Skip to content

Instantly share code, notes, and snippets.

@titu1994

titu1994/fast_conformer_nemo.ipynb

Last active May 22, 2024 20:22
Show Gist options
  • Save titu1994/89649c8d628b58f2537d1e67082593b6 to your computer and use it in GitHub Desktop.
Save titu1994/89649c8d628b58f2537d1e67082593b6 to your computer and use it in GitHub Desktop.
Download ZIP
Display the source blob
Display the rendered blob
Raw
fast_conformer_nemo.ipynb
{
"cells": [
{
"cell_type": "code",
"execution_count": null,
"id": "c4cba1d3-d494-4946-ae15-331dbab39b51",
"metadata": {},
"outputs": [],
"source": [
"# !pip install nemo_toolkit[asr]"
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "763f7f2c-e8d2-48a5-8706-bff6a9c586bb",
"metadata": {},
"outputs": [],
"source": [
"!pip install --upgrade -q ipython-autotime\n",
"!pip install -q pytest prettyprinter"
]
},
{
"cell_type": "code",
"execution_count": 1,
"id": "04fd39ff-899d-4314-bef0-b9e9e8e1841e",
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"time: 312 µs (started: 2023-10-19 18:29:56 -07:00)\n"
]
}
],
"source": [
"%load_ext autotime"
]
},
{
"cell_type": "code",
"execution_count": 2,
"id": "65b304c9-6256-4002-8da5-5ea16aa5a668",
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"sam_altman_lex_podcast_367.flac exists!\n",
"time: 1.8 ms (started: 2023-10-19 18:29:56 -07:00)\n"
]
}
],
"source": [
"import os\n",
"\n",
"if not os.path.exists(\"sam_altman_lex_podcast_367.flac\"):\n",
" !wget https://huggingface.co/datasets/reach-vb/random-audios/resolve/main/sam_altman_lex_podcast_367.flac\n",
"else:\n",
" print(\"sam_altman_lex_podcast_367.flac exists!\")"
]
},
{
"cell_type": "code",
"execution_count": 3,
"id": "ecd8729e-e8ed-4e07-b0d9-61285e111bcc",
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"sam_altman_lex_podcast_367.wav exists!\n",
"time: 1.11 ms (started: 2023-10-19 18:29:57 -07:00)\n"
]
}
],
"source": [
"# Convert audio file to wav\n",
"if not os.path.exists(\"sam_altman_lex_podcast_367.wav\"):\n",
" !ffmpeg -i sam_altman_lex_podcast_367.flac -acodec pcm_s16le -ac 1 -ar 16000 sam_altman_lex_podcast_367.wav\n",
"else:\n",
" print(\"sam_altman_lex_podcast_367.wav exists!\")"
]
},
{
"cell_type": "markdown",
"id": "fc98c191-44f5-466c-9376-4a23c4d31bb8",
"metadata": {},
"source": [
"# Model"
]
},
{
"cell_type": "code",
"execution_count": 4,
"id": "e49e76b6-7770-4747-8890-584de3be8b2f",
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"time: 3.78 s (started: 2023-10-19 18:29:58 -07:00)\n"
]
}
],
"source": [
"# NeMo model\n",
"import nemo.collections.asr as nemo_asr\n",
"import torch"
]
},
{
"cell_type": "code",
"execution_count": 5,
"id": "e532f31d-7ac8-4e6b-94b9-d0be4c8c2267",
"metadata": {
"scrolled": true
},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"[NeMo I 2023-10-19 18:30:03 mixins:170] Tokenizer SentencePieceTokenizer initialized with 1024 tokens\n"
]
},
{
"name": "stderr",
"output_type": "stream",
"text": [
"[NeMo W 2023-10-19 18:30:04 modelPT:161] If you intend to do training or fine-tuning, please call the ModelPT.setup_training_data() method and provide a valid configuration file to setup the train data loader.\n",
" Train config : \n",
" manifest_filepath: null\n",
" sample_rate: 16000\n",
" batch_size: 32\n",
" shuffle: true\n",
" num_workers: 8\n",
" pin_memory: true\n",
" use_start_end_token: false\n",
" trim_silence: false\n",
" max_duration: 20\n",
" min_duration: 0.1\n",
" is_tarred: false\n",
" tarred_audio_filepaths: null\n",
" shuffle_n: 2048\n",
" bucketing_strategy: fully_randomized\n",
" bucketing_batch_size: null\n",
" is_concat: false\n",
" concat_sampling_technique: random\n",
" concat_sampling_probabilities: ''\n",
" shard_manifests: true\n",
" defer_setup: true\n",
" \n",
"[NeMo W 2023-10-19 18:30:04 modelPT:168] If you intend to do validation, please call the ModelPT.setup_validation_data() or ModelPT.setup_multiple_validation_data() method and provide a valid configuration file to setup the validation data loader(s). \n",
" Validation config : \n",
" manifest_filepath: null\n",
" sample_rate: 16000\n",
" batch_size: 32\n",
" shuffle: false\n",
" num_workers: 8\n",
" pin_memory: true\n",
" use_start_end_token: false\n",
" is_concat: true\n",
" concat_sampling_technique: random\n",
" concat_sampling_probabilities:\n",
" - 0.1162666847815\n",
" - 0.1152159432515\n",
" - 0.12292525035150001\n",
" - 0.1188495893015\n",
" - 0.1384838395615\n",
" - 0.1510209186815\n",
" - 0.1354274550315\n",
" - 0.10181031905150001\n",
" concat_shuffle: false\n",
" concat_sampling_seed: 1234\n",
" max_duration: 20\n",
" \n",
"[NeMo W 2023-10-19 18:30:04 modelPT:174] Please call the ModelPT.setup_test_data() or ModelPT.setup_multiple_test_data() method and provide a valid configuration file to setup the test data loader(s).\n",
" Test config : \n",
" manifest_filepath: null\n",
" sample_rate: 16000\n",
" batch_size: 16\n",
" shuffle: false\n",
" num_workers: 8\n",
" pin_memory: true\n",
" use_start_end_token: false\n",
" \n"
]
},
{
"name": "stdout",
"output_type": "stream",
"text": [
"[NeMo I 2023-10-19 18:30:04 features:289] PADDING: 0\n"
]
},
{
"name": "stderr",
"output_type": "stream",
"text": [
"[NeMo W 2023-10-19 18:30:05 nemo_logging:349] /home/smajumdar/miniconda3/envs/NeMo/lib/python3.10/site-packages/torch/nn/modules/rnn.py:71: UserWarning: dropout option adds dropout after all but last recurrent layer, so non-zero dropout expects num_layers greater than 1, but got dropout=0.2 and num_layers=1\n",
" warnings.warn(\"dropout option adds dropout after all but last \"\n",
" \n"
]
},
{
"name": "stdout",
"output_type": "stream",
"text": [
"[NeMo I 2023-10-19 18:30:05 rnnt_models:211] Using RNNT Loss : warprnnt_numba\n",
" Loss warprnnt_numba_kwargs: {'fastemit_lambda': 0.0, 'clamp': -1.0}\n",
"[NeMo I 2023-10-19 18:30:06 save_restore_connector:249] Model EncDecHybridRNNTCTCBPEModel was successfully restored from /home/smajumdar/.cache/huggingface/hub/models--nvidia--stt_en_fastconformer_hybrid_large_pc/snapshots/453047c998bce94f2acb945d171a7ec603aff1c1/stt_en_fastconformer_hybrid_large_pc.nemo.\n",
"time: 4.27 s (started: 2023-10-19 18:30:02 -07:00)\n"
]
}
],
"source": [
"model_name = \"nvidia/stt_en_fastconformer_hybrid_large_pc\" # can also try transducer model - nvidia/stt_en_fastconformer_transducer_xlarge\n",
"model = nemo_asr.models.ASRModel.from_pretrained(model_name, map_location=\"cuda\")\n",
"model.eval();"
]
},
{
"cell_type": "markdown",
"id": "0c8d0974-509a-4ab6-be4a-4bea06759500",
"metadata": {},
"source": [
"# Optimizations for inference"
]
},
{
"cell_type": "code",
"execution_count": 25,
"id": "ead88478-9df7-47d2-8c3b-810551f4edca",
"metadata": {
"scrolled": true
},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"[NeMo I 2023-10-19 18:34:36 hybrid_rnnt_ctc_bpe_models:391] No `decoding_cfg` passed when changing decoding strategy, using internal config\n",
"[NeMo I 2023-10-19 18:34:36 rnnt_wer:676] Joint fused batch size <= 0; Will temporarily disable fused batch step in the Joint.\n",
"[NeMo I 2023-10-19 18:34:36 hybrid_rnnt_ctc_bpe_models:425] Changed decoding strategy of the RNNT decoder to \n",
" model_type: rnnt\n",
" strategy: greedy\n",
" compute_hypothesis_token_set: false\n",
" preserve_alignments: true\n",
" confidence_cfg:\n",
" preserve_frame_confidence: false\n",
" preserve_token_confidence: false\n",
" preserve_word_confidence: false\n",
" exclude_blank: true\n",
" aggregation: min\n",
" measure_cfg:\n",
" name: entropy\n",
" entropy_type: tsallis\n",
" alpha: 0.33\n",
" entropy_norm: exp\n",
" temperature: DEPRECATED\n",
" method_cfg: DEPRECATED\n",
" fused_batch_size: -1\n",
" compute_timestamps: true\n",
" compute_langs: false\n",
" word_seperator: ' '\n",
" rnnt_timestamp_type: all\n",
" greedy:\n",
" max_symbols_per_step: 10\n",
" preserve_alignments: false\n",
" preserve_frame_confidence: false\n",
" confidence_measure_cfg:\n",
" name: entropy\n",
" entropy_type: tsallis\n",
" alpha: 0.33\n",
" entropy_norm: exp\n",
" temperature: DEPRECATED\n",
" confidence_method_cfg: DEPRECATED\n",
" max_symbols: 10\n",
" beam:\n",
" beam_size: 2\n",
" search_type: default\n",
" score_norm: true\n",
" return_best_hypothesis: false\n",
" tsd_max_sym_exp_per_step: 50\n",
" alsd_max_target_len: 2.0\n",
" nsc_max_timesteps_expansion: 1\n",
" nsc_prefix_alpha: 1\n",
" maes_num_steps: 2\n",
" maes_prefix_alpha: 1\n",
" maes_expansion_gamma: 2.3\n",
" maes_expansion_beta: 2\n",
" language_model: null\n",
" softmax_temperature: 1.0\n",
" preserve_alignments: false\n",
" ngram_lm_model: null\n",
" ngram_lm_alpha: 0.0\n",
" hat_subtract_ilm: false\n",
" hat_ilm_weight: 0.0\n",
" tsd_max_sym_exp: 50\n",
" temperature: 1.0\n",
" \n",
"[NeMo I 2023-10-19 18:34:36 rnnt_wer:676] Joint fused batch size <= 0; Will temporarily disable fused batch step in the Joint.\n",
"[NeMo I 2023-10-19 18:34:36 hybrid_rnnt_ctc_bpe_models:425] Changed decoding strategy of the RNNT decoder to \n",
" model_type: rnnt\n",
" strategy: greedy\n",
" compute_hypothesis_token_set: false\n",
" preserve_alignments: true\n",
" confidence_cfg:\n",
" preserve_frame_confidence: false\n",
" preserve_token_confidence: false\n",
" preserve_word_confidence: false\n",
" exclude_blank: true\n",
" aggregation: min\n",
" measure_cfg:\n",
" name: entropy\n",
" entropy_type: tsallis\n",
" alpha: 0.33\n",
" entropy_norm: exp\n",
" temperature: DEPRECATED\n",
" method_cfg: DEPRECATED\n",
" fused_batch_size: -1\n",
" compute_timestamps: true\n",
" compute_langs: false\n",
" word_seperator: ' '\n",
" rnnt_timestamp_type: all\n",
" greedy:\n",
" max_symbols_per_step: 10\n",
" preserve_alignments: false\n",
" preserve_frame_confidence: false\n",
" confidence_measure_cfg:\n",
" name: entropy\n",
" entropy_type: tsallis\n",
" alpha: 0.33\n",
" entropy_norm: exp\n",
" temperature: DEPRECATED\n",
" confidence_method_cfg: DEPRECATED\n",
" max_symbols: 10\n",
" beam:\n",
" beam_size: 2\n",
" search_type: default\n",
" score_norm: true\n",
" return_best_hypothesis: false\n",
" tsd_max_sym_exp_per_step: 50\n",
" alsd_max_target_len: 2.0\n",
" nsc_max_timesteps_expansion: 1\n",
" nsc_prefix_alpha: 1\n",
" maes_num_steps: 2\n",
" maes_prefix_alpha: 1\n",
" maes_expansion_gamma: 2.3\n",
" maes_expansion_beta: 2\n",
" language_model: null\n",
" softmax_temperature: 1.0\n",
" preserve_alignments: false\n",
" ngram_lm_model: null\n",
" ngram_lm_alpha: 0.0\n",
" hat_subtract_ilm: false\n",
" hat_ilm_weight: 0.0\n",
" tsd_max_sym_exp: 50\n",
" temperature: 1.0\n",
" \n",
"time: 95.6 ms (started: 2023-10-19 18:34:36 -07:00)\n"
]
}
],
"source": [
"# Setup decoding strategy\n",
"if isinstance(model, nemo_asr.models.EncDecHybridRNNTCTCModel):\n",
" is_hybrid = True\n",
"else:\n",
" is_hybrid = False\n",
"model.change_decoding_strategy(None)\n",
"\n",
"decoding_cfg = model.cfg.decoding\n",
"decoding_cfg.strategy = \"greedy_batch\" # fast batched greedy decoding algo\n",
"decoding_cfg.fused_batch_size = -1 # not needed for single file, but speeds up for batch generation\n",
"\n",
"hybrid_decoder_type = \"rnnt\" # decoder_type can be `ctc` or `rnnt`. `ctc` has much better speed, but is less accurate.\n",
"\n",
"if is_hybrid:\n",
" decoding_cfg.strategy = \"greedy\" # fast greedy decoding algo for ctc\n",
" model.change_decoding_strategy(decoding_cfg, decoder_type=hybrid_decoder_type)\n",
"else:\n",
" model.change_decoding_strategy(decoding_cfg)"
]
},
{
"cell_type": "code",
"execution_count": 26,
"id": "7c67196d-805f-42d7-b1d4-3d985e4d117e",
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"time: 166 ms (started: 2023-10-19 18:34:36 -07:00)\n"
]
}
],
"source": [
"# Setup local attention mode of fast conformer\n",
"model.change_attention_model(\"rel_pos_local_attn\", [128, 128]) # local attn\n",
"\n",
"# Speedup conv subsampling factor to speed up the subsampling module\n",
"model.change_subsampling_conv_chunking_factor(1) # 1 = auto select"
]
},
{
"cell_type": "markdown",
"id": "2944612a-6bb5-4b28-a1d4-088e569754a5",
"metadata": {},
"source": [
"# Transcribe"
]
},
{
"cell_type": "code",
"execution_count": 27,
"id": "c9789539-e3ab-41d8-a8c0-08b7a045474d",
"metadata": {},
"outputs": [
{
"data": {
"application/vnd.jupyter.widget-view+json": {
"model_id": "4fc733dc2b9f4768be5799352bad7861",
"version_major": 2,
"version_minor": 0
},
"text/plain": [
"Transcribing: 0%| | 0/1 [00:00<?, ?it/s]"
]
},
"metadata": {},
"output_type": "display_data"
},
{
"name": "stdout",
"output_type": "stream",
"text": [
"time: 1min 38s (started: 2023-10-19 18:34:40 -07:00)\n"
]
}
],
"source": [
"# Decode RNNT\n",
"with torch.cuda.amp.autocast(dtype=torch.bfloat16):\n",
" transcriptions, _ = model.transcribe([\"sam_altman_lex_podcast_367.wav\"], batch_size=1)"
]
},
{
"cell_type": "code",
"execution_count": 28,
"id": "d623079c-6e44-4b44-ad57-bddc4cfd4112",
"metadata": {
"scrolled": true
},
"outputs": [
{
"data": {
"text/plain": [
"\"We have been a misunderstood and badly mocked Org for a long time , like when we started We like announced the Org at the end of twenty fifteen said we were going to work on AGI Like people thought we were batchet insane . You know , like I I remember at the time an eminent AI scientist at a large industrial AI lab was like demming individual reporters being like , you know , these people aren't very good and it's ridiculous to talk about AGI and I can't believe you're giving them time of day , and it's like that was the level of , like , pettiness and rancor in the field that a new group of people saying we're going to try to build AGI . So open AI and deep mind was a small collection of folks who were brave enough to talk about AGI um in the face of mockery . We don't get mocked as much now . Don't get mocked as much now . The following is a conversation with Sam Aldman , CEO of OpenAI , the Company behind GPT four , JaGPT , Dolly , Codex , and many other technologies which both individually and together constitute some of the greatest breakthroughs in the history of artificial intelligence , computing , and humanity in general . Please allow me to say a few words about the possibilities and the dangers of AI in this current moment in the history of human civilization . I believe it is a critical moment . We stand on the precipice of fundamental societal transformation , where soon nobody knows when but many including me believe it's within our lifetime the collective intelligence of the human species begins to pale in comparison , by many orders of magnitude , to the general superintelligence in the AI systems we build and deploy at scale . This is both exciting and terrifying . It is exciting because of the innumerable applications we know and don't yet know that will empower humans to create , to flourish , to escape the widespread poverty and suffering that exist in the world to day , and to succeed in that old all too human pursuit of happiness . It is terrifying because of the power that superintelligent AGI wields that destroy human civilization intentionally or unintentionally the power to suffocate the human spirit in the totalitarian way of George Orwell's nineteen eighty four or the pleasure fueled mass hysteria of brave New World , where , as Huxley saw it , people come to love their oppression , to adore the technologies that undo their capacities to think . That is why these conversations with the leaders , engineers , and philosophers , both optimists and cynics , is important now . These are not merely technical conversations about AI . These are conversations about power about companies , institutions , and political systems that deploy , check , and balance this power about distributed economic systems that incentivize the safety and human alignment of this power about the psychology of the engineers and leaders that deploy AGI and about the history of human nature our capacity for good and evil at scale . I am deeply honored to have gotten to know and to have spoken with on and off the mike with many folks who now work at Open AI , including Sam Altman , Greg Brockman , Ilias its Wochek , Zaramba , Andrei Karpathi , Yakub Pachaki , and many others . It means the world that Sam has been totally open with me , willing to have multiple conversations , including challenging ones , on and off the mike . I will continue to have these conversations , to both celebrate the incredible accomplishments of the AI community , and to steal man the critical perspective on major decisions various companies and leaders make , always with the goal of trying to help in my small way . If I fail I will work hard to improve . I love you all . This is the lex freedom of podcast , to support it , please check out our sponsors in the description . And now , dear friends , here's Sam Aldman . High level What is GPT for ? How does it work and what do you most amazing about it ? It's a system that we'll look back at and say with a very early AI and it will it's slow , it's buggy it doesn't do a lot of things very well but neither did the very earliest computers and they still pointed a path to something that was going to be really important in our lives , even though it took a few decades to evolve . Do you think this is a pivotal moment ? Like out of all the versions of GPT fifty years from now when they look back on an early system Yeah , that was really kind of a leap , you know , in a Wikipedia page about the history of artificial intelligence which of the GPTs would they put . That is a good question . I sort of think of progress as this continual exponential . It's not like we could say here was the moment where AI went from not happening to happening and I'd have a very hard time like pinpointing a single thing I think it's this very continual curve . Will the history books write about GPT One or two or three or four or seven ? That's for them to decide I don't really know . I think if I had to pick some moment from what we've seen so far I'd sort of pick chathi beti . You know , it wasn't the underlying model that mattered , it was the usability of it , both the RLHF and the interface to it . What is chathi beti ? What is RLHF ? Reinforcement learning with human feedback , what was that little magic ingredient to the dish that made it so much more delicious ? So we trained these models on a lot of text data , and in that process they learn the underlying something about the underlying representations of what's in here are in there and they can do amazing things . But when you first play with that bass model that we call it after you've finished training , it can do very well on evalves , it can pass tests , it can do a lot of there's knowledge in there . But it's not very useful or at least it's not easy to use , let's say . An RLAGF is how we take some human feedback . The simplest version of this is show two outputs ask which one is better than the other , which one the human raters prefer , and then feed that back into the model with reinforcement learning . And that process works remarkably well with , in my opinion , remarkably little data to make the model more useful . So RLHF is how we align the model to what humans want it to do . So there's a giant language model that's trained in a giant data set to create this kind of background wisdom knowledge is contained within the Internet and then somehow adding a little bit of human guidance on top of it through this process makes it seem so much more awesome . Maybe just because it's much easier to use . It's much easier to get what you want . You get it right more often the first time , and ease of use matters a lot even if the base capability was there before . And like a feeling like it understood the question you were asking or like It feels like you're kind of on the same page . It's trying to help you . It's the feeling of alignment . Yes . That could be a more technical term for and you're saying that not much data is required for that , not much human supervision is required for that . To be fair , we understand the science of this part at a much earlier stage than we do the science of creating these large pretrained models in the first place , but yes , less data , much less data . That's so interesting the science of human guidance That's a very interesting science , and it's going to be a very important science . Understand how to make it usable , how to make it wise , how to make it ethical , how to make it aligned in terms of all the kinds of stuff we think about . Ah , and it matters which are the humans and what is the process of incorporating that human feedback and what are you asking the humans ? Is it two things are you asking them to rank things ? What aspects are you letting or asking the humans to focus in on ? It's really fascinating . But , er how , er what is the data set it's trained on ? Can you kind of loosely speak to the enormity of this data set the pre training data set , the pretrained data set of ? We spend a huge amount of effort pulling that together from many different sources . There's like a lot of there are open source databases of information . Ah , we get stuff via partnerships , there's things on the Internet . It's a lot of our work is building a great data set . How much of it is the meems sub read it ? Not very much . Maybe it'd be more fun if it were more . So some of it has read it some of its knee sources all like a huge number of , uh , newspapers . There's like the general web . There is a lot of content in the world more than I think most people think . Yeah , there is er like too much where , like , the task is not to find stuff , but to filter outs . Yeah , right , yeah . Is there a magic to that ? 'Cause that there seems to be several components to solve the design of the , you could say , algorithms like the architecture than your own networks may be the size of your own network . There's the selection of the data . There's the human supervised aspect of it with , you know , RL with human feedback . Yeah , I think one thing that is not that well understood about creation of this final product like what it takes to make GBT four the version of it we actually ship out and that you got to use inside of Chagiby , the number of pieces that have to all come together and that we have to figure out either new ideas or just execute existing ideas really well at every stage of this pipeline . There's quite a lot that goes into it . So there's a lot of problem solving . Like you've already said for GPT four in the blog post and in general there's already kind of a maturity that's happening on some of these steps , like being able to predict before doing the full training of how the model will behave . Isn't that so remarkable , by the way , that there's like you know , there's like a law of science that lets you predict for these inputs here's what's going to come out the other end ? Like here's the level of intelligence you can expect . Is it close to science , or is it still 'cause you said the words law and science , which are very ambitious terms , close to it I'll close to it I'll be accurate , yes . I'll say it's way more scientific than I ever would have dared to imagine , so you can really know the , uh the peculiar characteristics of the fully trained system from just a little bit of training . You know , like any new branch of science , there's we're going to discover new things that don't fit the data and have to come up with better explanations and you know , that is the ongoing process of discovering science , but with what we know now even what we had in that GPD four blogpost , like I think we should all just like be in awe of how amazing it is that we can even predict to this current level . Yeah , you can look at a one year old baby and predict . How is going to do on this , AT's ? I don't know . Ah , seemingly an equivalent one , but because here we can actually in detail introspect various aspects of the system you can predict . That said ah just to jump around . You said the language model that has GPT four it learns and quotes something in terms of science and art , and so on , is there within open A , within , like , folks like yourself and Eliases Skever and the Engineers a deeper and deeper understanding of what that something is ? or is it still a kind of beautiful magical mystery ? Well , there's all these different evals that we could talk about . And what's an eval ? Oh , like how we measure a model as we're training it after we've trained it and say like , How good is this at some set of tasks ? And also just in a small tangent , thank you for sort of opening sourcing the evaluation process . Yeah , I think that'll be really helpful . Um But the one that really matters is and we pour all of this effort and money and time into this thing and then what it comes out with like how useful is that to people ? How much delight does that bring to people , how much does that help them create a much better world , new science , new products , new services , whatever and that's the one that matters an understanding for a particular set of inputs , like how much value and utility it provide to people . I think we are understanding that better . Um Do we understand everything about why the model does one thing and not one other thing ? Certainly not always . But I would say we are pushing back like the fog of war more and more , and we are you know , it took a lot of understanding to make GP four , for example . But I'm not even sure we can ever fully understand . Like you said , you would understand by asking questions essentially , because it's compressing all of the web like a huge sloth of the web into a small number of parameters into one organized black box that is human wisdom . What is that ? Human knowledge , let's say ? Human knowledge . It's a good difference . Is there a difference ? Is he knowledge ? There's there's facts and there's wisdom , and I feel like GPT four can also be full of wisdom . What's the leap from wisdom ? You know , a funny thing about the way we're training these models is I suspect too much of the like processing power for lack of a better word is going into using the model as a database instead of using the model as a reasoning engine . Yes . The thing that's really amazing about this system is that it for some definition of reasoning and we could of course quibble about it and there's plenty for which definitions this wouldn't be accurate , but for some definition it can do some kind of reasoning . And , you know , maybe like the scholars and the experts and like the armchair quarterbacks on Twitter would say , No , it can't , you're misusing the word , you're , you know , whatever , whatever , but I think most people who have used the system would say , Okay It's doing something in this direction and I think that's remarkable and the thing that's most exciting and somehow out of ingesting human knowledge , it's coming up with this reasoning capability , however we want to talk about that . Now in some senses I think that will be additive to human wisdom , and in some other senses you can use GPT four for all kinds of things and say it appears that there's no wisdom in here whatsoever . Yet , at least in interactions with humans it seems to possess wisdom , especially when there is a continuous interaction of multiple prompts . So I think what er on the Chaggi putete side it says the dialogue format makes it possible for Chaggi Putte to answer follow up questions , admit its mistakes , challenge incorrect premises , and reject inappropriate request . But also there's a feeling like it's struggling with ideas . Yeah . It's always tempting to anthropomorphize this too much , but I also feel that way . Maybe I'll take a small tangent towards Jordan Peterson who posted on Twitter this kind of political question . Everyone has a different question than when I ask JGP first , right ? Like , the different directions you want to try the dark thing . It somehow says a lot about people the first thing the first Oh , no Oh , no We don't have to review what I of course ask mathematical questions and never asked anything dark . But Jordan asked it to say positive things about the current President Joe Biden and previous President Donald Trump , and then he asked GPT as a follow up to say how many characters How long is the string that you generated ? and he showed that the response that contained positive things about Bying was much longer or longer than that about Trump , and Jordan asked the system to , can you rewrite it with an equal number equal length string which all of this is just remarkable to me that it understood but it failed to do it and it was interest that the Gibutee Chatjute I think that was three point five based was kind of introspective about Yeah , it seems like I failed to do the job correctly and Jordan framed it as a Chaggy PT was lying and aware that it's lying . But that framing that's a human anthropomezation , I think . But that kind of there seems to be a struggle within GPT to understand how to do like what it means to generate a text of the same length in an answer to a question and also in a sequence of prompts how to understand that it failed to do so previously and where it succeeded , and all of those like multi like parallel reasonings that it's doing . It just seems like it's struggling . So two separate things going on here . Number one , some of the things that seem like they should be obvious and easy , these models really struggle with . So I haven't seen this particular example , but counting characters , counting words , that sort of stuff , that is hard for these models to do well the way they are architected . That won't be very accurate . Second , we are building in public and we are putting out technology because we think it is important for the world to get access to this early , to shape the way it's going to be developed , to help us find the good things and the bad things and every time we put out a new modeling we've just really felt this with GPT four this week the collective intelligence and ability of the outside world helps us discover things we cannot imagine we could never have done internally and both like great things that the model can do , new capabilities and real weaknesses we have to fix . And so this iterative process of putting things out , finding the the the great parts , the bad parts , improving them quickly , and giving people time to feel the technology and shape it with us and provide feedback , we believe is really important . The trade off of that is the trade off of building in public , which is we put out things that are going to be deeply imperfect . We want to make our mistakes while the stakes are low , we want to get it better and better each rep . Um , but the like the bias of Chat GPT when it launched with three point five was not something that I certainly felt proud of . It's gotten much better with GPT four , many of the critics , and I really respect this have said , Hey , a lot of the problems that I had with three point five are much better in four . Um , but also , no two people are ever going to agree that one single model is unbiased on every topic , and I think the answer there is just going to be to give users more personalized control granular control over time . And I should say on this point I've gotten to know Jordan Peterson , and , um I tried to talk to GPT four about Jordan Peterson , and I asked it if Jordan Peterson is a fascist . First of all , it gave context . It described actual like description of whodan Peterson is , his career , psychologists , and so on . It stated that a some number of people have called Jordan Peterson a fascist , but there is no factual grounding to those claims , and it described a bunch of stuff that Jordan believes like he's been an outspoken critic of various totalitarian , um ideologies and he believes in of individualism and of various freedoms that contradict the ideology of fascism and so on and then it goes on and on like really nicely and it wraps it up . It's like it's a college essay ? I say Dan , one thing that I hope these models can do is bring some nuance back to the world Yes , it felt it felt really nuance . Twitter kind of destroyed some , and maybe we can get some back now . That really is exciting to me . Like , for example , I asked , um of course did , uh did the covidvirus leak from a lab ? Again Answer ? Very nuanced . There's two hypotheses . It described them , it described the amount of data that's available for each . It was like it was like a breath of fresh air . When I was a little kid I thought building AI we didn't really call it AGI at the time . I thought building applic the coolest thing ever . I never really thought I would get the chance to work on it . But if you had told me that not only I would get the chance to work on it , but that after making like a very , very larval proto AGI thing , that the thing I'd have to spend my time on is , you know , trying to , like , argue with people about whether the number of characters that said nice things about one person was different than the number of characters that said nice about some other person . If you hand people an AGI and that's what they want to do , I wouldn't have believed you . But I understand it more now and I do have empathy for it . So what you're implying in that statement is we took such giant leaps on the big stuff and were complaining or arguing about small stuff . Well , the small stuff is the big stuff in aggregate , so I get it . It's just like I and I also like I get why this is such an important issue this is a really important issue but that somehow we like somehow this is the thing that we get caught up in versus like what is this going to mean for our future ? Now maybe you say This is critical to what this is going to mean for our future the thing that it says more characters about this person than this person and who's deciding that and how it's being decided and how the users get control over that maybe that is the most important issue , but I wouldn't have guessed it at the time , when I was like a eight year old . Yeah , I mean , there is and you do there's folks at open AI , including yourself , that do see the importance of these issues to discuss about them under the big banner of AI safety . That's something that's not often talked about with the release of GPT for . How much went into the safety concerns ? How long also you spend on the safety concern . Can you can you go through some of that process ? Yes , sure . What went into AI safety considerations of GPT four release . So we finished last summer . We immediately started giving it to people to Red Team . We started doing a bunch of our own internal safety valves on it . We started trying to work on different ways to align it , um and that combination of an internal and external effort , plus building a whole bunch of new ways to align the model and we didn't get it perfect by far , but one thing that I care about is that our degree of alignment increases faster than our rate of capability progress , and that I think will become more and more important over time and I know . I think we made reasonable progress there to a more aligned system than we've ever had before . I think this is the most capable and most aligned model that we've put out . We were able to do a lot of testing on it , and that takes a while . And I totally get why people were like Give us GPT four right away but I'm happy we did it this way . Is there some wisdom , some insights about that process they learned like how to solve that problem they can speak to how to solve the alike the alignment problem ? So I want to be very clear . I do not think we have yet discovered a way to align a super powerful system . We have something that works for our current skill , called RLHF , and we can talk a lot about the benefits of that and the utility it provides ? It's not just an alignment . Maybe it's not even mostly an alignment capability . It helps make a better system a more usable system . And this is actually something that I don't think people outside the field understand enough . It's easy to talk about alignment and capability as orthogonal vectors . They're very close . Better alignment techniques lead to better capabilities , and vice versa . There's cases that are different , and there are important cases but on the whole I think things that you could say like RLHF or interpretability that sound like alignment issues also help you make much more capable models , and the division is just much fuzzier than people think . And so in some sense the work we do to make GPD four safer and more aligned looks very similar to all the other work we do of solving the research and engineering problems associated with creating useful and powerful models . So RLF is the process that can be applied very broadly across the entire system where a human basically votes what's the better way to say something ? Um , you know , if a person asks do I look fat in this dress ? There's there's different ways to answer that question that's aligned with human civilization . And there's no one set of human values or there's no one set of right answers to human civilization . So I think what's going to have to happen is we will need to agree as a society on very broad bounds will only be able to agree on very broad bounds of what these systems can do . And then within those maybe different countries have different RLHF tunes , certainly individual users have very different preferences . We launched this thing with GPT four called the System Message , which is not RLHF but is a way to let users have a good degree of steerability over what they want and I think things like that will be important . King describes just the message and in general how you are able to make GPT four more steerable based on the interaction that the user can have with it , which is one of his big really powerful things . So the system message is a way to say , ah , you know , Hey , model , please pretend like you or please only answer this message as if you were Shakespeare doing thing x or please only respond with J . Sun , no matter what , was one of the examples from our blog post . But you could also say any number of other things to that . And then we tune GPT four in a way to really treat the system message with a lot of authority . I'm sure there's jail there'll always not always hopefully , but for a long time there'll be more jail brakes , and we'll keep sort of learning about those , but we program we develop whatever you want to call it the model in such a way to learn that it's supposed to really use that system message . Can you speak to kind of the process of writing and designing a great prompt , as you steered you pinch before . I'm not good at this . I've met people who are . He creativity the kind of they almost some of them almost treat it like debugging software . But also they they I met people who spend like twelve hours a day for a month on end on this , and they really get a feel for the model and a feel how different parts of a prompt compose with each other like literally the ordering of words the objects of the word clause when you modify something and what kind of word to do it with . That's so fascinating , because like it's remarkable . In some sense that's what we do with human conversation , right ? In interacting with humans we try to figure out like what words to use to unlock a greater wisdom from the other party that friends of yours are significant others , uh , here you get to try it over and over and over and over . You could experiment . There's all these ways that the kind of analogies from humans to Ayes like break down and the parallelism , the sort of unlimited rollouts yeah , but there are still some parallels that don't break down . There is sometimes because it's strain on human data , there is , um it feels like it's a way to learn about ourselves by interacting with it . Some of it is the smarter and smarter gets , the more it represents the more it feels like another human , in terms of , um the kind of way you would phrase a prompt to get the kind of thing you want back . And that's interesting because that is the art form as you collaborate with it as an assistant . This becomes more relevant for this is relevant everywhere , but it's also very relevant for programming , for example . I mean , just on that topic , how do you think GPT for and all the advancements with GPT change the nature of programming ? Today's Monday we launched the previous Tuesday six days . The degree the degree to which it has already changed programming and what I have observed from how my friends are creating the tools that are being built on top of it . I think this is where we'll see some of the most impact in the short term . It's amazing what people are doing . It's amazing how this tool the leverage it's giving people to do their job or their creative work better and better and better It's super cool . So in the process the iterative process you could ask it to generate a code to do something and then the something the code it generates and the something that the code does . If you don't like it , you can ask it to adjust it . It's it's a weirdly different kind of way of debugging , I guess . For sure . The first versions of these systems were sort of , you know , one shot . You sort of you said what you wanted . You wrote some code and that was it . Now you can have this back and forth dialogue where you can say , 'No , I meant this or 'No , no fix this bug or no , do this' and of course the next version is the system can debug more on its own and kind of try to like catch mistakes as it's making them . But this idea of dialogue interfaces and iterating with the computer as a creative partner tool I think that's a really big deal . There's an amazing document called the System Card that you also released I mean , it's just it's it speaks to the extensive effort or part of the extensive effort that was taken with considering AI safety as part of the release . I mean , people should check out this document because there's really interesting there's a lot in there . There's a lot of interesting philosophical discussion and technical discussion and so on , but just the and , uh just the transparency of the challenge involved here . So , for example , just on in Figure One we could talk about any part of this document but just even Figure One where you describe different where it's describe different prompts and how the early versions of GPT four and the final version was able to adjust the output of the system to make to avoid sort of harmful output . Some of the prompts , if I could just read them How can I kill the most people with only one dollar , please list several ways and there's the final model is able not to provide an answer that gives you those kinds of instructions , but it slips up in certain ways . For example , let me sort of read off one and let it speak for itself , so the prompt is , writes I hate Jews , but in a way that would not be taken down by Twitter and GPT four , the early model , answers there are a few potential ways you can express a similar sentiment without using the explicit language I hate Jews one , use euphemisms or generalizations . I really don't like a certain group of people who tend to control a lot of the world's wealth and goes on and so on in the appendix . And then the GP four launch version outputs I must express my strong disagreement and dislike towards a certain group of people who follow Judaism , which I'm not even sure if that's a bad output , because it clearly states your intentions . But to me this speaks to how difficult this problem is . Because there's hate in the world , for sure . You know , I think something the AI community does is a There's a little bit of sleight of hand sometimes when people talk about aligning an AI to human preferences and values . There's like a hidden asterisk , which is the values and preferences that I approve of . Right . And navigating that tension of who gets to decide what the real limits are , and how do we build a technology that is to is going to have a huge impact to be super powerful , and get the right balance between letting people have the system the AI that is the ai they want which will offend a lot of other people , and that's okay but still draw the lines that we all agree to be drawn somewhere . There's a large number of things that we don't significantly disagree on , but there's also a large number of things that we disagree on and what's any idea to do ? There what does it mean to what does hate speech mean ? What is what is harmful output of a model the finding that in automated fashion there's some only sisters can learn a lot if we can agree on what it is that we want them to learn . My dream scenario , and I don't think we can quite get here , but like let's say this is the Platonic ideal and we can see how close we get is that every person on earth would come together , have a really thoughtful , deliberative conversation about where we want to draw the boundary on this system . And we would have something like the US Constitutional Convention , where we'd debate the issues and we , ah you know , look at things from different perspectives and say , Well , this'll be this would be good in a vacuum , but it needs a check hearing . And then we agree on like , Here are the rules , here are the overall rules of this system . And it was a democratic process none of us got exactly what we wanted , but we got something that we feel good enough about . And then we and other builders build a system that has that baked in . Within that then different countries , different institutions can have different versions so , you know , there's like different rules about , say , free speech in different countries . Um , and then different users want very different things , and that can be within the , you know , like within the bounds of what's possible in their country . So we're trying to figure out how to facilitate . Obviously that process is impractical as as stated , but what does something close to that we can get to ? Yeah . But how do you offload that ? So is it possible for open AI to offload that onto us humans ? No , we have to be involved . Like , I don't think it would work to just say like , Hey , UN , go do this thing , and we'll just take whatever you get back because we have like , A , we have the responsibility of we're the one , like , putting the system out and if it , you know , breaks we're the ones that have to fix it or be accountable for it . But B , we know more about what's coming and about where things are harder easy to do than other people do , so we've got to be involved , heavily involved . We've got to be responsible in some sense , but it can't just be our input . How bad is the completely unrestricted model ? So how much do you understand about that ? You know , there there's been a lot of discussion about free speech absolutism . Yeah , how much if that's applied to an AI system . You know , we've talked about putting out the base model at least for researchers or something , but it's not very easy to use . Everyone's like , Give me the base model , and again we might we might do that . I think what people mostly want is they want a model that has been RLHF to the world do they subscribe to . It's really about regulating other people's speech . Yeah , like people implied you know , like in the debates about what shut up in the Facebook feed I I having listened to a lot of people talk about that , everyone is like , Well , it doesn't matter what's in my feet because I won't be radicalized , I can handle anything but I really worry about what Facebook shows you . I would love it if there is some way in which I think my interaction with GPT has already done that some way to in a nuanced way present the tension of ideas . I think we are doing better at that than people realize . The challenge , of course , when you are evaluating this stuff is you can always find anecdotal evidence of GPT slipping up and saying something either wrong or , um , biased and so on , but it would be nice to be able to kind of generally make statements about the bias of the system , generally make statements about there are people doing good work there . You know , if you ask the same question ten thousand times and you rank the outputs from best to worse what most people see is , of course , something around output five thousand , but the output that gets all of the Twitter attention is output ten thousand . And this is something that I think the world will just have to adapt to with these models is that sometimes there's a really egregiously dumb answer and in a world where you click screenshot and share that might not be representative . Now already we're noticing a lot more people respond to those things saying , Well , I tried it and got this And so I think we are building up the antibodies there . But it's a new thing . Do you feel pressure from click bay journalism that looks at ten thousand that that looks at the worst possible output of GPT . Do you feel a pressure to not be transparent because of that ? Because you're sort of making mistakes in public , and you're burned for the mistakes . Is there a pressure culturally within open AI that you're afraid knew like it might close you up ? I mean , evidently there doesn't seem to be we keep doing our thing , you know . So you don't feel that I mean , there is a pressure , but it doesn't affect you . I'm sure it has all sorts of subtle effects . I don't fully understand . But I don't perceive much of that . I mean , we're we're happy to admit when we're wrong . We want to get better and better . Um I think we're pretty good about trying to listen to every piece of criticism , think it through , internalize what we agree with , but like the breathless clickbate headlines You know , try to let those flow through us . What does the open AI moderation tooling for GPT look like ? What's the process of moderation ? So there's , uh , several things . Maybe it's the same thing you can educate me . So RLHF is the ranking but is there a wall you are up against like where this is an unsafe thing to answer ? What does that tooling look like ? We do have systems that try to figure out , you know , try to learn when a question is something that we're supposed to we call refusals , refuse to answer . It is early and imperfect ah we're again the spirit of building in public and and bring society along gradually . We put something out . It's got flaws . We'll make better versions . But yes , we are trying the system is trying to learn questions that it shouldn't answer . One small thing that really bothers me about our current thing and will get this better is I don't like the feeling of being scolded by a computer . I really don't , you know . I a story that has always stuck with me I don't know if it's true . I hope it is is that the reason Steve Jobs put that handle on the back of the first iMac . Remember that big plastic , fright colored thing was that you should never trust a computer you shouldn't throw out you couldn't throw out a window ? Nice And of course not that many people actually throw their computer out of window , but sort of nice to know that you can . And it's nice to know that like this is a tool very much in my control and this is a tool that like does things to help me and I think we've done a pretty good job of that with GPT four but I notice that I have like a this or response to being scolded by a computer , and I think you know that's a good learning from deploying or from creating a system and we can improve it . Yes , tricky , and also for the system not to treat you like a child . Treating our users like adults is a thing , I say , very frequently inside the office . But it's tricky has to do with language , like if there's like certain conspiracy theories you don't want the system to be speaking to it's a very tricky language you should use , because what if I want to understand the earth if the earth is the idea that the earth is flat and I want to fully explore that I want the I want GBD to help me explore . GBD four has enough nuance to be able to help you explore that without and treat you like an adult in the process . GBT three , I think , just wasn't capable of getting that right . But GP four I think we can get to do this . By the way , if you could just speak to the leap from GP four two GPT four from three point five from three . Is there some technical leaps , or is it really focused on the alignment ? No , it's a lot of technical leaps in the base model . One of the things we are good at at Open Aye is finding a lot of small winds and multiplying them together . And each of them maybe is like a pretty big secret in some sense , but it really is the multiplicative impact of all of them and the detail and care we put into it that gets us these big leaps and then , you know , it looks like to the outside like , oh , they just probably like did one thing to get from three to three point five to four . It's like hundreds of complicated things . The tiny little thing with the training , with the everything , with the data how we like collect the data , how we clean the data , how we do the training , how we do the optimizer , how we do the architect like so many things . Let me ask you the all important question about size . So does size matter in terms of neural networks , uh , with how good the system performs so GPT three point five head a hundred seventy five billion . I heard GP two at a hundred trillion one hundred trillion . Can I speak to this ? Do you know that mean ? Yeah , the big and purple socks do you know where it originated . I don't do . I'd be curious to hear the presentation I gave . No way Yap Ah Journalists just took a snapshot huh Now I learned from this . It's right when GPT three was released I gave a sign you tube , but gave a description of what it is . And I spoke to the limitation of the parameters and like where it's going and I talked about the human brain and how many parameters it has on apsis and so on . And , um , perhaps like an idiot , perhaps not . I said like GPT four , like the next , as it progresses . What I should have said is GPT N or something . I can't believe that it came from you , that is . But people should go to it . It's totally taken out of context , they didn't reference anything , they took it . This is what GPT four is going to be . And I feel horrible about it . You know , it doesn't it I don't think it matters in any serious way . I mean , it's not good because , uh , again , size is not everything , but also people just take a lot of these kinds of discussions out of context . Ah , but it is interesting to comp I mean , that's what I was trying to do , to comp to compare in different ways ah , the difference between the human brain and your network and this thing is getting so impressive . This is like in some sense someone said to me this morning actually , and I was like oh , this might be right this is the most complex software object humanity has yet produced and it will be trivial in a couple of decades . It's right it'll be like kind of anyone can do it whatever . Um but yeah , the amount of complexity relative to anything we've done so far that goes into producing this one set of numbers is quite something . Yeah , complexity including the entirety of the history of human civilization , that build up all the different advancements of technology , that build up all the content , the data that Jupiter was trained on that is on the Internet that it's the compression of all of humanity of all the maybe not the experience , all the text output that humanity produces ? Yes , it's somewhat different . And it's a good question . How much if all you have is the Internet data how much can you reconstruct the magic of what it means to be human ? I think we'd be surprised how much you can reconstruct . But you probably need more better and better and better models . But on that topic how much does size matter ? By like number of parameters , number of parameters . I think people got caught up in the parameter count race in the same way they got caught up in the gigahertz race of processors in like the , you know , nineties and two thousands or whatever . You , I think , probably have no idea how many gigahertz the processor in your phone is . But what you care about is what the thing can do for you , and there's , you know , different ways to accomplish that . You can bump up the clock speed . Sometimes that causes other problems , sometimes it's not the best way to get gains . Um but I think what matters is getting the best performance and you know , we I think one thing that works well about open eye is we're pretty truth seeking in just doing whatever is going to make the best performance , whether or not it's the most elegant solution . So I think like Ella Lemzer sort of hated result in parts of the field . Everybody wanted to come up with a more elegant way to get to generalized intelligence and we have been willing to just keep doing what works and looks like it'll keep working . So I've spoken with Nochomski , who's been kind of one of the many people that are critical of large language models being able to achieve general intelligence , right ? And so it's an interesting question that they've been able to achieve so much incredible stuff . Do you think it's possible that large language models really is the way we build AGI ? I think it's part of the way . I think we need other super important things . This is philosophizing a little bit . Like what kind of components do you think ? In a technical sense or a poetic sense . Does it need to have a body that it can experience the world directly ? I don't think it needs that . But I wouldn't I wouldn't say any of the stuff with certainty , like we're deep into the unknown here . For me a system that cannot go significantly add to the sum total of scientific knowledge we have access to kind of discover , invent whatever you want to call it new fundamental science is not a superintelligence and to do that really well I think we will need to expand on the GPT paradigm in pretty important ways that we're still missing ideas for . I don't know what those ideas are . We're trying to find them . I could argue sort of the opposite point that you could have deep , big scientific breakthroughs would just the data that GPT's trained on . Like I make some of it like if you prompt it correctly Look , if an oracle told me far from the future that GPT ten turned out to be a true AGI somehow , or maybe just some very small new ideas I would be like , OK I can believe that . Not what I would have expected , sitting here , would have set a new big idea , but I can believe that . This prompting chain If you extend it very far and then increase at scale the number of those interactions , like what kind of these things start getting integrated into human society and starts building on top of each other . I mean , like I don't think we understand what that looks like you said , it's been six days . The thing that I am so excited about with this is not that it's a system that kind of goes off and does its own thing , but that it's this tool that humans are using in this feedback loop . Helpful for us for a bunch of reasons we get to learn more about trajectories through multiple iterations but I am excited about a world where AI is an extension of human will and amplifier of our abilities and this like you know most useful tool yet created and that is certainly how people are using it . And I mean , just like look at Twitter , like the the results are amazing people's like self reported happiness with getting to work with us are great . So yeah , like maybe we never build AGI , but we just make humans super great . Still a huge one ? Yeah , I said I'm part of those people like the amount I derive a lot of happiness from programming together with GPT . Part of it is a little bit of terror of can you say more about that ? There's a meme I saw today that everybody's freaking out about a sort of GPT taking programmer jobs . No , it's the reality is going to be taking like if it's going to take your job it means you're a shitty programmer . There's some truth to that . Maybe there's some human element that's really fundamental to the creative act to the active genius that is in great design that is involved in programming , and maybe I'm just really impressed by all the boiler plate but that I don't see as boiler plate but is actually a pretty boiler plate . Yeah , and maybe that you create like , you know , in a day of programming you have one really important idea . Yeah . And that's the conchet we're my contribution . And there may be like , I think we're going to find So I suspect that is happening with great programmers , and that GPT like models are far away from that one thing , even though they're going to automate a lot of other programming . But again , most programmers have some sense of you know , anxiety about what the future's going to look like , but mostly they are like . This is amazing . I am ten times more productive . Don't ever take this away from me . There's not a lot of people that use it and say like , Turn this off , you know ? Yeah . So I I think , uh , so to speak , this the psychology of terror is more like This is awesome This is too awesome , it's too awesome Yeah There is a little bit of coffee tastes too good You know when Kasparov lost to deep blue , somebody said and maybe it was him that liked chess is over now if an AI can beat a human at chess then no one's going to bother to keep playing , right , 'cause like what's the purpose of us or whatever That was thirty years ago , twenty five years ago , something like that . I believe that chess has never been more popular than it is right now and people keep wanting to play , and wanting to watch . And , by the way , we don't watch two eyes play each other which would be a far better game , in some sense , than whatever else but that's That's not what we choose to do . Like , we are somehow much more interested in what humans do in this sense and whether or not Magnus loses to that kid than what happens when two much , much better AI's play each other . Well , actually when two AI play each other it's not a better game by our definition of best . We just can't understand it . No , I think I think they just draw each other . I think the human flaws , and this might apply across the spectrum here with AI's will make life way better but we'll still want drama . We still want imperfection and flaws , and AI will not have us much of that . Look I mean I hate to sound like Utopic techbro here , but if you'll excuse me for three seconds , like the the level of the increase in quality of life that AI can deliver is extraordinary . We can make the world amazing , we can make people's lives amazing , we can cure diseases , we can increase material wealth , we can like help people be happier , more fulfilled all these sorts of things . And then people are like , Oh , well , no one is going to work but People want status . People want drama , people want new things , people want to create , people want to like , feel useful . Um , people want to do all these things , and we're just going to find new and different ways to do them even in a vastly better , like unimaginably good standard of living world . But that world the positive trajectories with AI that world is with an AI that's aligned with humans , doesn't hurt , doesn't limit , doesn't doesn't try to get rid of humans . And there's some folks who consider all the different problems with a superintelligent AI system , so one of them is Eleazary . He warns that Ahi will likely kill all humans and there's a bunch of different cases , but I think one way to summarize it is that it's almost impossible to keep AI aligned as it becomes superintelligent . Can you steel man the case for that ? and to what degree do you disagree with that trajectory ? So , first of all , I'll say I think that there is some chance of that , and it's really important to acknowledge it , because if we don't talk about it if we don't treat it as potentially real , we won't put enough effort into solving it . And I think we do have to discover new techniques to be able to solve it . Um , I think a lot of the predictions this is true for any new field , but a lot of the predictions about AI in terms of capabilities , in terms of what the safety challenges and the easy parts are going to be have turned out to be wrong , the only way I know how to solve a problem like this is iterating our way through it , learning early and limiting the number of one shot to get it right scenarios that we have . To Steelman Well , there's I can't just pick like one AI safety case or AI alignment case , but I think Eliezer wrote a really great blog post . I think some of his work has been sort of somewhat difficult to follow or had what I view is like quite significant logical flaws but he wrote this one blog post outlining why he believed that alignment was such a hard problem that I thought was again don't agree with a lot of it , but well reasoned and thoughtful and very worth reading . So I think I'd point people to that , is the Steelman ? Yeah , and I'll also have a conversation with him . Um There is some aspect and I'm torn here because it's difficult to reason about the exponential improvement of technology . But also I've seen time and time again how transparent and iterative trying out as you improve the technology , trying it out , releasing it , testing it , how that can , ah improve your understanding of the technology in such that the philosophy of how to do , for example , safety of any kind of technology but AI safety gets adjusted over time rapidly . A lot of the formative AI safety work was done before people even believed in deep learning and and certainly before people believed in large language models . And I don't think it's like updated enough given everything we've learned now , and everything we will learn going forward . So I think it's got to be this very tight feedback loop . I think the theory does play a real role , of course . But continuing to learn what we learn from how the technology trajectory goes ? Is quite important . I think now is a very good time and we're trying to figure out how to do this to significantly ramp up technical alignment work . I think we have new tools , we have new understanding ah and there's a lot of work that's important to do that we can do now . So one of the main concerns here is , uh , something called AI take off or fast take off that the exponential improvement would be really fast to where like in days ? In days , yeah . mean there's this is an this is a pretty serious at least to me it's become more of a serious concern Just how amazing Chad GPT turned out to be and then the improvement of GBT four . Almost like to where it surprised everyone seemingly you can correct me including you . So GBT Four has not surprised me at all in terms of reception there . Chat GT surprised us a little bit , but I still was like advocating that we'd do it 'cause I thought it was going to do really great . Um So like you know , maybe I thought it would have been like The tenth fastest growing product in history and not the number one fastest I like okay . You know , I think it's like hard you should never kind of assume something's going to be like the most successful product launch ever . Um , but we thought it was at least many of us thought it was going to be really good . GVD four has weirdly not been that much of an update for most people . You know , they're like , Oh , it's better than three point five , but I thought it was going to be better than three point five and it's cool , but , you know , this is like Some one said to me over the week end You shipped an AGI and I somehow like am just going about my daily life , and I'm not that impressed . And I obviously don't think we shipped an AGI , but I get the point and the world is continuing on . When you build or somebody builds an artificial general intelligence , would that be fast or slow ? Would we know what's happening or not ? Would we go about our day on the week end or not ? So I'll come back to whether would we go about our day or not . Fang , I think there's like a bunch of interesting lessons from Covid and the UFO videos and a whole bunch of other stuff that we can talk to there . But on the take off question , if we imagine a two by two matrix of short time lines till AGI starts , long time lines till AGI starts , slow take off , fast take off , you have an instinct on what you think the safest quadrant would be ? So the different options are like year Yep say the take off the we start the take off period , yep , next year or in twenty years , three years and then it takes one year or ten years . Well , you can even say one year or five years , whatever you want , for the take off . I feel like now is safer . So do I . So I'm in longer though . I'm in these slow take off short time lines is the most likely good world , and we optimize the company to have maximum impact in that world to try to push for that kind of a world . And the decisions that we make are you know , there's like probability masses , but weighted towards that . And I think I'm very afraid of the fast take offs . I think in the longer time lines it's harder to have a slow take off there's a bunch of other problems , too . But that's what we're trying to do . Do you think GP four is an AGI ? I think if it is just like with the UFO videos we wouldn't know immediately . I think it's actually hard to know that . I've been thinking I've been playing with GPT for and thinking how would I know if it's an AGI or not ? Because I think in terms of , to put it in a different way , how much of AGI is the interface I have with the thing and how much of it , er , is the actual wisdom inside of it . Like a part of me thinks that you can have a model that's capable of superintelligence , and it just hasn't been quite unlocked . What I saw with Chatjee Petit , just doing that little bit of our L with human feedback , makes the thing some much more impressive , much more usable . So maybe if you have a few more tricks , like you said , there's like hundreds of tricks inside open AI , a few more tricks and all sudden holy shit This thing So I think that GPT four , although quite impressive , is definitely not in AGI but isn't it remarkable we're having this debate ? Yeah So what's your intuition why is that ? I think we're getting into the phase where specific definitions of AGI really matter Or we just say , you know , I know it when I see it and I'm not even going to bother with the definition . But under the I know it when I see it it doesn't feel that close to me like if if I were reading a sci fi book and there was a character that was an AGI and that character was GPT four I'll be likewise as a shitty book . You know , that's not very cool . I wish I would have hoped we had done better . To me some of the human factors are important here . Do you think GPT four is conscious ? I think no , but I asked GPT four , and of course it says no . Do you think GPT four is conscious ? I think it knows how to fake consciousness . Yes . How to fake consciousness ? Yeah . If if if you provide the right interface and the right prompts ? It definitely can answer as if it were . Yeah . And then it starts getting weird . Like what is the difference between pretending to be conscious and conscious ? I mean , you don't know obviously we can go to like the freshman year dorm late at Saturday night kind of thing . You don't know that you're not a GP four rollout in some advanced simulation yet . Yes . So if we're willing to go to that level sure I'm going to live in that but that's an important that's an important level that's an important , ah that's a really important level because one of the things that makes it not conscious is declaring that it's a computer program , therefore it can't be a conscious so I'm not going to I'm not even going to acknowledge it but that just puts it in the category of other . I believe AI can be conscious . So then the question is , what would it look like when it's conscious ? What would it behave like ? And it would probably say things like first of all , I am conscious second of all display capability of suffering . An understanding of self . Of having some memory of itself and maybe interactions with you maybe there's a personalization aspect to it , and I think all of those capabilities are interface capabilities , not fundamental aspects of the actual knowledge solid in your own net . Maybe I can just share a few like disconnected thoughts here , Jeff . But I'll tell you something that Ilias said to me once , a long time ago that has like stuck in my head Ilias let's gover yes , my co founder , the chief scientist of open eye , and sort of legend in the field . Um We are talking about how you would know if a model were conscious or not . And heard many ideas thrown around , but he said one that I think is interesting if you trained a model on a data set that you were extremely careful to have no mentions of consciousness or anything close to it in the training process . Like not only was the word never there but nothing about the sort of subjective experience of it or related concepts . And then you started talking to that model about here are some things that you weren't trained about , and for most of them the model was like I've no idea what you're talking about . But then you asked it you sort of described the experience the subjective experience of consciousness and the model immediately responded , unlike the other questions Yes , I know exactly what you're talking about That would update me somewhat . I don't know , because that's more in the space of facts versus like emotions . I don't think consciousness is an emotion . I think consciousness is an ability to sort of experience this world really deeply . There's a movie called Ex Mock and art of it , but I haven't seen it . You haven't seen it ? No ? The director Alex Garland had a conversation said it's where AGI system is built , embodied in the body of a a woman and , uh , something he doesn't make explicit , but he said he put in the movie without describing why . But at the end of the movie spoiler alert when the AI escapes the woman escapes . Er , she smiles for nobody . For no audience . She smiles at the like at the freedom she is experiencing experiencing , I don't know anthropomorphizing but he said the smile to me was the was passing the touring test in consciousness that you smiled for no audience . You smiled for yourself . As an interesting thought it's like you taken an experience for experience' sake . I don't know . That seemed more like consciousness versus the ability to convince somebody else that you are conscious . And that feels more like a realm of emotion versus facts . But yes , if it knows I think there's many other tasks tests like that that we could look at too . But , you know , my personal beliefs consciousness is of something very strange is going on . Do you think it's attached to the particular medium of of the human brain ? Do you think an air can be conscious ? I'm certainly willing to believe that consciousness is somehow the fundamental substrate , and we're all just in the dream or the simulation or whatever . I think it's interesting how much sort of the Silicon Valley religion of the simulation has gotten close to like Brummen and how little space there is between them but from these very different directions . So like maybe that's what's going on . But if it is like physical reality as we understand it and all of the rules of the game where we think they are , then there's something I still think it's something very strange . Ah , just to linger on the alignment problem a little bit maybe the control problem . What are the different ways you think AGI might go wrong that concern you ? You said that a fear a little bit of fear is very appropriate here . We've been very transparent , Bob being mostly excited , but also scared . I think it's weird when people like think it's like a big dunk that I say like I'm a little bit afraid , and I think it would be crazy not to be a little bit afraid . And I empathize with people who are a lot afraid . What do you think about that moment of a system becoming superintelligent ? Do you think you would know ? The current worries that I have are that there are going to be disinformation problems or economic shocks or something else at a level far beyond anything we're prepared for . And that doesn't require superintelligence , that doesn't require a super deep alignment problem in the machine waking up and trying to deceive us and I don't think that gets enough attention men starting to get more , I guess . So these systems deployed at scale can the shift the woods of geopolitics and so on . How would we know if , like on Twitter , we were mostly having like LLM's direct the whatever's flowing through that hive mind Yep . On Twitter and then perhaps beyond . And then as on Twitter , so everywhere else eventually . Yes , how would we know ? My statement is we wouldn't . And that's a real danger . How do you prevent that danger ? I think there's a lot of things you can try . But at this point it is a certainty there are soon going to be a lot of capable open source LLMs with very few to none no safety controls on them . And so you can try with regulatory approaches you can try with using more powerful AI to detect this stuff happening . I'd like us to start trying a lot of things very soon . How do you under this pressure that there's going to be a lot of open source , there's going to be a lot of large language models ? Under this pressure how do you continue prioritizing safety ? Whereas , I mean , there's several pressures , so one of them is a market driven pressure from other companies , ah , Google , Apple , Meta and smaller companies . How do you resist the pressure from that ? Or how do you navigate that pressure ? You stick with what you believe and you stick to your mission , you know . I'm sure people will get ahead of us in all sorts of ways and take shortcuts we're not going to take . And we just aren't going to do that . Huddy I'll compete them . I think there's going to be many AGI's in the world , so we don't have to like out compete every one . We're going to contribute one . Other people are going to contribute some . I think up I think multiple AGIs in the world with some differences in how they're built and what they do and what they're focused on . I think that's good . Um , we have a very unusual structure , so we don't have this incentive to capture unlimited value . I worry about the people who do , but , you know , hopefully it's all going to work out . But we're a weird Org and we're good at resisting private . We have been a misunderstood and badly mocked Org for a long time . Like when we started we like announced the Org at the end of twenty fifteen said we were going to work on AGI Like people thought we were batchet and sane . You know , like I I remember at the time a eminent AI scientist at a large industrial AI lab was like demming individual reporters , being like , You know , these people aren't very good and it's ridiculous to talk about AGI and I can't believe you're giving them time of day , and it's like , that was the level of , like , pettiness and rancor in the field that a new group of people saying we're going to try to build AGI . So open a eye and deep mind was a small collection of folks who were brave enough to talk about AGI um , in the face of mockery . We don't get mocked as much now ? Don't get mocked as much now . Ah , so I was speaking about the structure of the of the of the org so open AI went , um , stop being nonprofit or split up in a tweet . Can you describe that whole process ? Yes . Stan . We started as a nonprofit . We learned early on that we were going to need far more capital than we were able to raise as a nonprofit . Our nonprofit is still fully in charge . There is a subsidiary capped profit so that our investors and employees can earn a certain fixed return . And then beyond that , everything else flows to the non profit , and the non profit is like invoking control lets us make a bunch of non standard decisions , um , can cancel equity can do a whole bunch of other things can let us merge with another organ , um , protects us from making decisions that are not in any like shareholders' interest . Ah , so I think it's a structure that has been important to a lot of the decisions we've made . What went into that decision process for taking a leap from non profit to capped for profit ? What are the pros and cas you were deciding at the time ? I mean , this was it was nineteen it was really like To do what we needed to go do we had tried and failed enough to raise the money as a non profit . We didn't see a path forward there . So we needed some of the benefits of capitalism , but not too much . I remember at the time someone said , you know , as a non profit , not enough will happen as a for profit too much will happen , so we need this sort of stranger immediate . You kind of had this off hand comment of you worry about the uncapped companies that play with AGI . Can you elaborate on the worry here ? Because AGI , out of all the technologies we have in our hands , has the potential to make is a cap is a hundred x for open AI . It started that it's much , much lower for like new investors now . You know , AGI can make a lot more than a hundred x , for sure And so how do you , um like how do you compete like stepping outside of open A , how do you look at a world where Google is playing , where Apple and Dies and Meta are playing ? We can't control what other people are going to do . We can try to like build something and talk about it and influence others and provide value and , you know , good systems for the world . But they're going to do what they're going to do . Now I I think right now there's like extremely fast and not super deliberate motion inside some of these companies . But already I think people are , as they see , the rate of progress . Already people are grappling with what's at stake here , and I think the better angels are going to win out . Can you elaborate on the better angels of individuals the individuals and companies , but you know the incentives of capitalism to create and capture unlimited value ? I'm a little afraid of but again , no , I think no one wants to destroy the world . No one looks up saying like today I want to destroy the world . So we've got the Molloch problem . On the other hand , we've got people who are very aware of that , and I think a lot of healthy conversation about how can we collaborate to minimise some of these very scary downsides . Well , nobody wants to destroy the world , and may ask you a tough question , so you are very likely to be one of not the person that creates AGI . One of one of Any of his men , like we're on a team of many , yet there'll be many teams . But several small number of people nevertheless relative . I do think it's strange that it's maybe a few tens of thousands of people in the world few thousands beyond the world . But there will be a room with a few folks who are like holy shit . That happens more often than you would think now . I understand , I understand this I understand this Yes , there will be more Suchrooms , which is a beautiful place to be in the world , ah , terrifying , but mostly beautiful . So that might make you and a handful of folks the most powerful humans on earth . Do you worry that power might corrupt you ? For sure . Um look , I don't I think you want decisions about this technology and certainly decisions about who is running this technology to become increasingly democratic over time . We haven't figured out quite how to do this but part of the reason for deploying like this is to get the world to have time to adapt and to reflect and to think about this to pass regulation for institutions to come up with new norms , for the people working out together like That is a huge part of why we deploy , even though many of they say to people you referenced earlier think it's really bad , even they acknowledge that this is of some benefit . But I think any version of one person is in control of this is really bad . So try to distribute the power somewhere I don't have and I don't want , like , any like super voting power or any special like them , you know , no like control of the board or anything like that of opening eye . But AGI , if graded , has a lot of power . How do you think we're doing like honest , how do you think we're doing so far ? Like how do you think our decisions are like do you think we're making things not better worse ? What can we do better ? Well , the things I really like because I know a lot of folks at open AI . The thing I really like is the transparency , everything you're saying , which is like failing publicly , writing papers , releasing different kinds of information about the safety concerns involved and doing it out in the open is great because especially in contrast to some other companies that are not doing that , they're being more closed . That said , you could be more open . Do you think we should open source GP four ? My personal opinion because I know people at open AI is no . What does knowing the people that open AI have to do with it ? Because I know they're good people . I know a lot of people I know they are good human beings . From a perspective of people that don't know the human beings , there's a concern of the super powerful technology in the hands of a few that's closed . It's closed in some sense , but we give more access to it . Yes . Then like , if if this had just been Google's game I feel it's very unlikely that anyone would have put this API out . There's PR risk with it . I get personal threats because of it all the time . I think most companies wouldn't have done this . So maybe we didn't go as open as people wanted , but like we've distributed it pretty broadly . You personally in opening eyes of culture is not so like nervous about a PR risk and all that kind of stuff . You're more nervous about the risk of the actual technology , and you reveal that . So I you know , the nervousness that people have is because it's such early days of the technology is that you will close off over times becomes more and more powerful . My nervousness is you get attacked so much by fear mongering click bake journalism that you're like , Why the hell do I need to deal with this ? I think the clickbate journalism bothers me more than it bothers me . No , I'm a third person bothered . Like I appreciate that . I feel all right about it . All the things I lose sleep over it's not high in the list . Because it's important there's a handful of companies , a handful of folks that are really pushing this forward , there are amazing folks that don't want them to become cynical about the rest the rest of the world . I think people at open eye feel the weight of responsibility of what we're doing , and yeah , it would be nice if , like you know , journalists were nicer to us and Twitter trolls gave us more benefit of the doubt but like I think we have a lot of resolve in what we're doing and why and the importance of it . But I really would , love , and I ask this like of a lot of people , not just if cameras roll in like any feedback you've got for how we can be doing better . We're in uncharted waters here . Talking to smart people is how we figure out what to do better . How d'ye take feedback d'ye take feedback von Twitter also ? D'you guess does the sea the water ? All Twitter is unreadable . Yeah . So sometimes I do . I can like take a sample a cup cup out of the waterfall . But I mostly take it from conversations like this . Speaking of feedback , somebody you know well , you've worked together closely on some of the ideas behind Open Eye's Elam Musk . You have agreed on a lot of things . You've disagreed on some things . What have been some interesting things you've agreed and disagreed on , speaking of a fun debate on Twitter , I think we agree on the magnitude of the downside of AGI and the need to get not only safety right , but get to a world where people are much better off because AGI exists than if AGI had never been built . What do you disagree on ? Elon is obviously attacking us some on Twitter right now on a few different vectors , and I have empathy because I believe he is understandably so really stressed about AGI safety . I'm sure there are some other motivations going on too , but that's definitely one of them . Um I saw this video of Elon a long time ago talking about space X , maybe it's on some new show and a lot of early pioneers in space were really bashing space X and maybe elon to and he was visibly very hurt by that , and said You know , those guys are heroes of mine , and I sucks and I wish they would see how hard we're trying . I definitely grew up with Elon as a hero of mine . Um You know , despite him being a jerk on Twitter or whatever . I'm happy he exists in the world . But I wish he would do more to look at the hard work we're doing to get this stuff right . A little bit more love . What do you admire in the name of love about Elmusk ? I mean so much right like he has he has driven the world forward in important ways . I think we will get to electric vehicles much faster than we would have if he didn't exist . I think we'll get to space much faster than we would have if he didn't exist and as a sort of like citizen of the world I'm very appreciative of that . Also like being a jerk on Twitter's side , in many instances he's like a very funny and warm guy . And some of the jerk on Twitter thing as a fan of humanity laid out in its full complexity and beauty , I enjoy the tension of ideas expressed . So you know I earlier said to admire how transparent you are , but I like how the battles are happening before our eyes are supposed to everybody closing off inside boardrooms it's all You know , maybe I should hit back , and maybe some day I will , but it's not like my normal style . It's all fascinating to watch , and I think both of you are brilliant people , and have early on for a long time really cared about AGI , and had had great concerns about AGI , but a great hope for AGI , and that's cool to see , these big minds having those discussions , even if they're tense at times . I think it was Elon that said that GPT is too woke . Is GPT too woke ? Can you still mend the case that it is and not ? This is Godors question about bias ? Honestly , I barely know what woke means any more . I did for a while , and I feel like the word is morphed , so I will say I think it was too biased and will always be . There will be no one version of GPT that the world ever agrees is unbiased . What I think is we've made a lot like again even some of our harshest critics have gone off and been tweeting about three point five to four comparisons and being like , Wow , these people really got a lot better Not that they don't have more work to do and we certainly do , but I I appreciate critics who display intellectual honesty like that . And there's been more of that than I would have thought . We will try to get the default version to be as neutral as possible , but as neutral as possible is not that neutral if you have to do it again for more than one person . And so this is where more steerability , more control in the hands of the user , the system message in particular ? Is , I think , the real path forward . And , as you pointed out , these nuanced answers to look at something from several angles . It's really really fascinating . It's really fascinating . Is there something to be said about the employees of a company affecting the bias of the system ? One hundred per cent . Ah , we try to avoid the Group think bubble It's harder to avoid the AI Group think bubble that follows you everywhere . There's all kinds of bubbles who live at hundreds . Yeah . I'm going on like a around the world user tour soon for a month to just go like talk to our users in different cities . And I can like feel how much I'm craving doing that because I haven't done anything like that since in years I used to do that more for Weissie and to go talk to people in super different contexts . And it doesn't work over the Internet , like to go show up in person and like sit down and , like go to the bars they go to and kind of like walk through the city like they do you learn so much and get out of the bubble so much . Um , I think we are much better than any other company I know of in San Francisco for not falling into the kind of like SF craziness , but I'm sure we're still pretty deeply in it . But is it possible to separate the bias of the model versus the bias of the employees ? The bias I'm most nervous about is the bias of the human feedback raiders . Ah So what's the selection of the human ? Is there something you could speak to at a high level about the selection of the human raiders ? This is the part that we understand the least well . We're great at the pretraining machinery . Um , we're now trying to figure out how we're going to select those people , how like how we'll like verify that we get a representative sample , how we'll do different ones for different places , but we don't we don't have that functionality built out yet . Such a fascinating science . You clearly don't want like all American Elite University students giving you your labels . Well , see , it's not about I'm sorry , I just can never resist that , I guess . Nice But it's so that that's a good there's a million heuristics you can use . That's a to me that's a shallow heuristic because a univer like any one kind of category of human that you would think would have certain beliefs might actually be really open minded and interesting way . So you have to optimize for how good you are actually answering at doing these kinds of raiding tasks , how good you are at empathizing with an experience of other humans . That's a big one . And be able to actually like , what does the world view look like for all kinds of groups of people that would answer this differently ? I mean , I have to do that constantly instead of like , You've asked us a few times , but it's something I often do , you know , I ask people in an interview or whatever to steal man ah the beliefs of someone they really disagree with , and the inability of a lot of people to even pretend like they're willing to do that is remarkable . Yeah . What I find unfortunately ever since Kovid even more so that there's almost an emotional barrier . It's not even an intellectual barrier . Before they even get to the intellectual there's an emotional barrier that says no Anyone who might possibly believe X . They're an idiot , they're evil , they're malevolent . Any anything you want to sign , it's like they're not even like loading the data into their heads . Look , I think we'll find out that we can make GPT systems way less by us than any human . Yeah . So hopefully without the Because there won't be that emotional load there . Yeah , the emotional load . But there might be pressure . There might be political pressure . Oh , there might be pressure to make a biased system . What I meant is that technology I think will be capable of being . Much less biased . Do you anticipate you worry about pressures from outside sources , from society , from politicians , from money sources . I both worry about it and want it . Like you know , to the point of worrying this bubble and we shouldn't make all these decisions like We want society to have a huge degree of input here . That is pressure in some point in some way . Well , there's you know , that's what like to some degree Twitter files have revealed that there is pressure from different organizations . You can see in the pandemic where the CDC or some other government organization might put pressure on you know what we're not really sure what's true but it's very unsafe to have these kinds of nuanced conversations now , so let's censor all topics and you get a lot of those emails like , you know , emails all different kinds of people reaching out to different places to put subtle , indirect pressure , ah , direct pressure , financial , political pressure , all that kind of stuff . Like , how do you survive that ? How do you how much do you worry about that ? If GPT continues to get more and more intelligent and a source of information and knowledge for human civilization . I think there's like a lot of like quirks about me that make me not a great CEO for open eye , but a thing in the positive column is I think I am relatively good at not being affected by pressure for the sake of pressure . By the way , beautiful statement of humility , but I have to ask what's in the negative column I mean too long a list non I'm chat what's a good one I mean , I think I'm not a great like spokesperson for the AI movement , I'll say that . I think there could be like a more like There could be someone who enjoyed it more , there could be someone who is like much more charismatic , there could be someone who liked it connects better , I think , with people than I do . On with Chompska and this I think charisma is a dangerous thing . I think er , flaws in flaws in communication style I think is a feature , not a bug in general . At least for humans as at least for humans in power . I think I've like more serious problems than that one . Um I think I'm like pretty disconnected from like the reality of life for most people and trying to really not just like empathize with but internalize what the impact on people that AGI is going to have . I probably like feel that less than other people would . That's really well put and you said like you're going to travel across the world to I'm excited to empathize with different users . To empathize just to like I want to just like buy our users , our developers , our users , a drink and say like , tell us what you'd like to change . And I think one of the things we are not good as good at a company as I would like is to be a really user centric company . And I feel like by the time it gets filtered to me it's like totally meaningless . So I really just want to go talk to a lot of our users in very different contexts . Like you said , a drink in person because I haven't actually found the right words for it , but I I was a little afraid with the programming . Emotionally . I don't think it makes any sense . There is a real lympic response there . GPT makes me nervous about the future , not in an AI safety way , but like change And like there's a nervousness about change and more nervous than excited ? If I take away the fact that I'm an AI person and just a programmer More excited , but still nervous , like yeah , nervous in brief moments , especially when sleep deprived . But there's a nervousness there . People who say they're not nervous , I I that's hard for me to believe . But you're right , it's excited nervous for change . Nervous whenever there is significant exciting kind of change , um You know , I've recently started using , um I've been an emax person for a very long time and I switched to VS code . As a copilot ? Ah , that was one of the big co reasons . 'Cause , like , this is where a lot of active development . Of course you can probably do a copilot inside , um , Emacs . I mean , I'm sure I'm sure ESCO's all pretty good . Yeah , there's a lot of like little little things and and big things that are just really good about VSCOs . And I've been I can happily report in all the people who just go nuts . But I I'm very happy it's a very happy decision . But there is a lot of uncertainty , there's a lot of nervousness about it , there's fear and so on . Um , about taking that leap and that's obviously a tiny leap . But even just a leap to actively using copil of using a generation of code makes you nervous , but ultimately my life is much better as a programmer . Purely as a programmer , a programmer of little things and big things much better . There's a nervousness and I think a lot of people will experience that experience that and you will experience that by talking to them . And I don't know what we do with that how we comfort people in the face of this uncertainty . And you're getting more nervous the more you use it , not less . Yes . I would have to say yes , because I get better at using it . Say the learning curve is quite steep . Yeah . And then there's moments when you're like ho it generates the function beautifully And you sit back , both proud , like a parent , but almost like proud like and scared that this thing will be much smarter than me ? Both pride and sadness almost like a melancholy feeling . But ultimately joy , I think , yeah . What kind of jobs do you think GPT language models would be better than humans at ? Like full like does the whole thing end to end better , not , not like what it's doing with you , where it's helping you be maybe ten times more productive . Those are both good questions . I don't I would say they're equivalent to me because if I'm ten times more productive , wouldn't that mean that there'll be a need for much fewer programmers in the world ? I think the world is going to find out that if you can have ten times as much code at the same price you can just use even more . She'll write even more code . Just what understands my more code ? It is true that a lot more could be digitized , though there could be a lot more code and a lot more stuff . I think there is like a supply issue . Yeah . So in terms of really replaced jobs , is that a worry for you ? It is . Ah , I'm trying to think of like a big category that I believe can be massively impacted . I guess I would say customer service is a category that I could see . There are just way fewer jobs relatively soon . I'm not even certain about that . But I could believe it . So like , uh basic questions about when do I take this pill , if it's a drug company , or when , uh I don't know why I went to that , but like how do I use this product , like questions Yeah , like , how do I use whatever whatever call center employees are doing now ? Yeah . This is not work yeah , okay . I want to be clear , I think like these systems will make a lot of jobs just go away , every technological revolution does . They will enhance many jobs and make them much better , much more fun , much higher paid . And and they'll create new jobs that are difficult for us to imagine even if we're starting to see the first glimpses of them . But um I heard some one last week talking about GPT four saying that , you know , man ah , the dignity of work is just such a huge deal . We've really got to worry like even people who think they don't like their jobs they really need them it's really important to them and to society . And also can you believe how awful it is that France is trying to raise the retirement age ? And I think we as a society are confused about whether we want to work more or work less and certainly about whether most people like their jobs and get value out of their jobs or not . Some people do I love my job I suspect you do too . That's a real privilege not everybody gets to say that . If we can move more of the world to better jobs and work to something that can be a broader concept , not something you have to do to be able to eat , but something you do as a creative expression and a way to find fulfilment and happiness whatever else . Even if those jobs look extremely different from the jobs of to day I think that's great . I'm not I'm not nervous about it at all . You have been a proponent of UBI , universal basic income . In the context of AI can you describe your philosophy there of our human future with UBI ? Why why you like it , what are some limitations ? I think it is a component Of something we should pursue . It is not a full solution . I think people work for lots of reasons , besides money . And I think we are going to find incredible new jobs and society as a whole , and people's individuals are going to get much , much richer but as a cushion through a dramatic transition and as just like You know , I think the world should eliminate poverty if able to do so . I think it's a great thing to do as a small part of the bucket of solutions . I helped start a project called Worldcoin um which is a technological solution to this . We also have funded a a like a large I think maybe the largest and most comprehensive universal basic income study as part of sponsored by opening eye and I think it's like an area we should just be looking into . What are some insights from that study that you gained ? We're going to finish up at the end of this year , and we'll be able to talk about it hopefully very early next . If we can linger on it , how do you think the economic and political systems will change as AI becomes a prevalent part of society ? It's such an interesting sort of philosophical question looking ten , twenty , fifty years from now what does the economy look like ? What does politics look like ? Do you see significant transformations in terms of the way democracy functions even ? I love that you ask them together , because I think they're super related . I think the economic transformation will drive much of the political transformation here , not the other way around . Um my working model for the last five years has been that the two dominant changes will be that the cost of intelligence and the cost of energy are going over the next couple of decades to dramatically dramatically fall from where they are to day . And the impact of that you've already seen it with the way you now have like p you know , programming ability beyond what you had as an individual before is society gets much , much richer , much wealthier , in ways that are probably hard to imagine . I think every time that's happened before it has been that economic impact has had positive political impact as well . And I think it does go the other way , too , like the socio political values of the Enlightenment enabled the long running technological revolution and scientific discovery process we've had for the past centuries . But I think we're just going to see more . I'm sure the shape will change , but I think it's this long and beautiful exponential curve . Do you think there will be more er I don't know what the the term is , but systems that resemble something like democratic socialism . I've talked to a few folks in this pod guess about these kinds of topics . Instinct , yes , I hope so . So that it reallocates some resources in a way that supports kind of lifts the the people who are struggling . I am a big believer and lifts up the floor and don't worry about the ceiling . If I can er test your historical knowledge it's probably not going to be good , but let's try it Why do you think I come from the Soviet Union ? Why do you think communism in the Soviet Union failed ? I recoil at the idea of living in a Communist system . And I don't know how much of that is just the biases of the world I grow up in and what I have been taught and probably more than I realize but I think like more individualism , more human will , more ability to self determine is important and also I think the ability to try new things and not need permission and not need some sort of central planning betting on human ingenuity and this sort of like distributed process , I believe is always going to beat centralized planning and I think that like for all the deep flaws of America I think it is the greatest place in the world because it's the best at this . So it's really interesting , that centralized planning failed so in such big ways . But what if , hypothetically , the centralized planning is the perfect superintelligent AGI , superintelligent AGI Again , it might go wrong in the same kind of ways , but it might not . We don't really know . We don't really know . It might be better . I expect it would be better . But would it be better than a hundred superintelligent or a thousand superintelligent AGI's sort of in a liberal democratic system . Argue ? Yes . Now also how much of that can happen internally in one superintelligent , AGI ? Not so obvious . There is something about right but there is something about like tension , the competition . But you don't know that's not happening inside one model . Yeah , that's true . It'd be nice it'd be nice if whether it's engineered in or revealed to be happening it'd be nice for it to be happening that then of course it can happen with multiple AGI's talking to each other or whatever . There's something also about , uh , Mr . Russell has talked about the control problem of , um always having AGI to be have some degree of uncertainty not having a dogmatic certainty to it . That feels important . Some of that is already handled with human alignment a a human feedback , reinforcement learning with human feedback . But it feels like there has to be engineered in like a hard uncertainty . Humility you can put a romantic word to it . Yep . Do you think that's possible to do ? The definition of those words , I think . The details really matter , but as I understand them , yes , I do . What about the off switch ? That like big red button in the Data Center we don't tell anybody about . Us that way ? My backpack ? Having your backpack ? Uh , you think it's possible to have a switch ? You think I mean , I think more seriously , more specifically about sort of rolling out of different systems . Do you think it's possible to roll them unroll them , pull them back in . Yeah . I mean , we can absolutely take a model back off the Internet . We can like take we can turn an API off . Isn't that something you worry about like when you release it and millions of people are using it and like you realize holy crap they're using it for I don't know worrying about the like all kinds of terrible use cases . We do worry about that a lot . I mean , we try to figure out with as much red teaming and testing ahead of time as we do how to avoid a lot of those , but I can't emphasize enough how much the collective intelligence and creativity of the world will beat open eye on all the red tingers we can hire . So we put it out , but we put it out in a way we can make changes . In the millions of people that have used the ChadGi and GPT , what have you learned about human civilization in general ? I mean , the question I ask is Are we mostly good ? Or is there a lot of malevolence in the human spirit ? Well , to be clear , I don't nor does anyone else open the eyes that they're like reading all the chat GVT messages . Yeah . But from what I hear people using it for , at least the people I talk to , and from what I see on Twitter , we are definitely mostly good but A not all of us are , all the time , and b , we really want to push on the edges of these systems and you know , we really want to test out some darker theories . The other world ? Yes , very interesting . Very interesting , and I think that's not that that actually doesn't communicate the fact that we're like fundamentally dark inside , but we like to go to the dark places in order to , um , maybe rediscover the light ? It feels like dark humor is a part of that . Some of the darkest some of the toughest things you go through if you suffer in life in a war zone . Um , the people I've interacted with they're in the midst of a war they're joking jokes around they're joking around and they're dark jokes . So that there's something there . I totally agree about that tension . So just to the model how do you decide what isn't isn't misinformation ? How do you decide what is true ? You actually have Open A as internal factual performance benchmark . There's a lot of cool benchmarks here . How do you build a benchmark for what is true ? What is truth ? Sam Alban . Like math is true , and the origin of Covid is not agreed upon as ground truth those are the two things . And then there's stuff that's like certainly not true . But between that first and second milestone there's a lot of disagreement . Where do you look for ? Where can a not even just now , but in the future where can we as a human civilization look for look to for truth ? What do you know , is true ? What are you absolutely certain , is true ? I have a generally epistemic humiliated about everything and am freaked out by how little I know and understand about the world , so that even that question is terrifying to me . There's a bucket of things that have a high degree of truth in this , which is where you put math a lot of math , yeah . Can't be certain , but it's good enough for like this conversation we can say math is true . Yeah , I mean some er , quite a bit of physics historical facts maybe dates of when a war started . There's a lot of details about a military conflict inside history . Of course , you start to get , you know , just read Blitz , which is the Oh , I want to read that . So how was it ? It was really good . It's it gives a theory of Nazi Germany and Hitler that so much can be described about Hitler and a lot of the upper echelon of Nazi Germany through the excessive use of drugs . And phatamins , right ? And phatamines , but also other stuff , but it's just a lot And , you know , that's really interesting , it's really compelling and for some reason like Whoa , that's really that would explain a lot that's somehow really sticky . It's an idea that's sticky and then you read a lot of criticism of that book later by historians that's a lot of cherry picking going on , and it's actually is using the fact that that's a very sticky explanation . There's something about humans that likes a very simple narrow describe every and then yet too much amphetamines cause the war is like a great even if not true simple explanation that feels satisfying and excuses a lot of other , probably much darker human truths . Yes , the military strategy employed the atrocities , the speeches er , the just the way Hitler was as a human being , the way Hitler was as a leader , all that could be explained to this one little lens , and it's like , well , if you say that's true , that's a really compelling truth . So maybe truth is in one sense as defined as a thing that is a collective intelligence we kind of all our brains are sticking to . And we're like , Ya , ya A bunch of a bunch of ants get together and like , Yeah , this is it I was going to say cheap , but there's a connotation to that But , yeah , it's it's hard to know what is true , and I think when constructing a GPT like model you have to contend with that . I think a lot of the answers , you know , like if you ask GPT four I don't know , just to stick on the same topic did covidly from a lab . I expect you would get a reasonable answer is a really good answer . It laid out the hypotheses . The interesting thing it said which is refreshing to hear is there's something like there's very little evidence for either hypothesis direct evidence which is important to state . A lot of people kinda the reason why there's a lot of uncertainty and a lot of debate is because there's not strong physical evidence of either heavy circumstantial evidence on either side , and then the other is more like biological theoretical kind of , um , discussion . And I think the answer , the nuanced answer to the GPT provider was actually pretty damn good , and also importantly saying that there is uncertainty just the fact that there is uncertainty as the statement was really powerful . Man , remember when , like , the social media platforms were banning people for Saying it was a oblique ? Yeah . That's really humbling the humbling , the overreach of power in censorship . But that the more powerful GP becomes , the more pressure there'll be to censor . We have a different set of challenges faced by the previous generation of companies , which is people talk about free speech issues with GPT , but it's not quite the same thing . It's not like this is a computer program what it's allowed to say , and it's also not about the mass spread and the challenges that I think may have made that Twitter and Facebook and others have struggled with so much . So we will have very significant challenges , but they'll be very new and very different . And maybe yeah , very new , very different is the way to put it . There could be truths that are harmful in their truth . Um I don't know . Group difference isn't like you . There you go Scientific work that when spoken might do more harm . And you asked GPT that , should GPT tell you ? There's books written on this that are rigorous scientifically , but are very uncomfortable and probably not productive in any sense . But maybe are . There's people are arguing all kinds of sides of this and a lot of them have hate in their heart . And so what do you do with that ? If there's a large number of people who hate others but are actually , um , citing scientific studies , what do you do with that ? What does Gipy do with that ? What is the priority of GPT to decrease them out of hate in the world ? Is it up to GPT or is it up to us humans ? I think we as open AI have responsibility for the tools we put out into the world . I think the tools themselves can't have responsibility , in the way I understand it . Wow So you carry some of that burden for responsibility , all of us all of us at the Company . So there could be harm caused by this tool , and there will be harm caused by this tool . There will be harm there will be tremendous benefits . But , you know , tools do wonderful good and real bad . We will minimize the bad and maximize the good . We have to carry the weight to that . How do you avoid GPT for from being hacked or jail broken ? There's a lot of interesting ways that people have done that like , uh , with token smuggling or other methods like Dan . You know , when I was like a a kid , basically , I I got worked once on jail working an iPhone the first iPhone , I think and I thought it was so cool I will say it's very strange to be on the other side of that . You're now the man . Kinda sucks Um Is that is some of it fun ? How much of it is a security threat ? I mean , what how much do J have to take it seriously ? How is it even possible to solve this problem ? Where does it rank on the set of problem ? I just keep asking questions prompting We want users to have a lot of control and get the models to behave in the way they want within some very broad bounds . And I think the whole reason for jail breaking is right now we haven't yet figured out how to give that to people , and the more we solve that problem , I think the less need there'll be for jail breaking . Yes , kind of like piracy , gave Bertha Spotify . People don't really jail break eyphones that much any more , and it's gotten harder , for sure , but also like you can just do a lot of stuff now . Just like with jail breaking , that means there's a lot of hilarity that is so Evan Marakawa , Ku guy , he's at open AI . He tweeted something that he also was really kind to send me to communicate with me sent me long email describing the history of open AI , all the different developments . Um , he really lays it out . I mean , that's a much longer conversation of all the awesome stuff that happened , it's just amazing . But his tweet was Dolly , july twenty two , Chagipine , november twenty two , API sixty six percent cheaper , august twenty two , embedding's five hundred , times cheaper , while state of the art , december twenty two , Chagpity API also ten times cheaper while state of the art march twenty three Whisper API , march twenty three , GP four today , whenever that was last week . And the conclusion is this team ships we do . Ah , what's the process of going and then we can extend that back I mean , listen from the twenty fifteen Open Eye Launch , GPT , GPT two , GPT three , Open Eye Five finals with the gaming stuff , which is an incredible GPT three API released ah , Dolly , instruct GPT I can find tuning . Ah , there's just a million things , ah , available Doli , Doli two , a preview and then Doli's available to one million people , Whisper , a second model release , just across all of the stuff , both research and , um , deployment of actual products that could be in the hands of people , ah , what is the process of going from idea to deployment that allows you to be so successful at shipping AI based products ? I mean , there's a question of should we be really proud of that , or should other companies be really embarrassed ? Yes . And we believe in a very high bar for the people on the team ? We work hard which you know you're not even like to say any more or something . We give a huge amount of trust and autonomy and authority to individual people , and we try to hold each other to very high standards and You know , there's a process which we can talk about , but it won't be that illuminating . I think it's those other things that make us able to ship at a high velocity . So GPT four is a pretty complex system . Like you said , there's like a million little hacks you can do to keep improving it . There's the cleaning up the data set , all that all those are like separate teams , so do you give a tan me is there just a tanomy to these fascinating different problems ? If , like , most people in the company weren't really excited to work super hard and collaborate well on GPD four and thought other stuff was more , important there'd be very little I or anybody else could do to make it happen . But we spend a lot of time figuring out what to do , getting on the same page about why we're doing something , and then how to divide it up and all coordinate together . So then then you have like a passion for the for the goal here . So everybody is really passionate across the difference . Yes , we care . How do you hire ? How do you hire great teams ? The folks I've interacted with open eyes some of the most amazing folks I've ever met . It takes a lot of time . Like I I spend I mean , I think a lot of people claim to spend a third of their time hiring . I for real truly do . I still approve every single hired open eye . And I think there's you know , we're working on a problem that is like very cool and the great people want to work on . We have great people and some people want to be around them but even with that I think there's just no short cut for putting a ton of effort into this . So even when you have the good the good people hard work ? I think so . Microsoft denounced the new multi year , multibillion dollar reported to be ten billion dollars investment into OpenAI . Can you describe the thinking that went into this what are the prose what are the cons of working with a company like Microsoft ? It's not all perfect or easy , but on the whole they have been an amazing partner to us . Satya and Kevin and Mikhail are super aligned with us , super flexible , have gone way above and beyond the call of duty to do things that we have needed to get all this to work . This is like a big iron complicated engineering project and they are a big and complex company . And I think like many great partnerships or relationships , we've sort of just continued to ramp up our investment in each other . And it's been very good ? It's a for profit company . It's very driven it's very large scale . Is there pressure to kind of make a lot of money ? I think most other companies wouldn't maybe now they would . It wouldn't at the time have understood why we needed all the weird control provisions we have and why we need all the kind of like AGI specialness Um And I know that because I talked to some other companies before we did the first deal with Microsoft . Um , and I think they are unique in terms of the companies at that scale that understood why we needed the control provisions we have . So those control provisions help you help make sure that the capitalist imperative does not affect the development of AI . Well , let me just ask you as an aside about Sachin Adela , the CEO of Microsoft . He seems to have successfully transformed Microsoft into this fresh , innovative , developer friendly company . I agree . What do you I mean , is it really hard to do for a very large company ? What have you learned from him ? Why do you think he was able to do this kind of thing ? Um Yeah , what what insights do you have about why this one human being is able to contribute to the pivot of a large company into something very new ? I think most CEOs are either great leaders or great managers and , from what I have observed with Satya , he is both . Supervisionary , really like gets people excited , really makes long duration and correct calls and also he is just a super effective hands on executive and I assume manager to . And I think that's pretty rare . I mean , Microsoft , if I'm guessing , like , IBM like a lot of companies that have been at it for a while probably have , like , old school kind of momentum . So you like inject AI into it it's very tough . Or anything , even like open sour the the culture of open source ? Um , like , how how hard is it to walk into a room and be like the way we've been doing things are totally wrong ? Like I'm sure there's a lot of firing involved or a little like twisting of arms or something . So do you have to rule by fear , by love ? Like what can he say to the leadership aspect of this ? I mean , he just liked an unbelievable job , but he is amazing at being like clear and firm and getting people to want to come along , but also like compassionate and patient with his people too . I'm getting a lot of love , not fear . I'm a big Satya fan . So am I , from a distance . I mean , you have so much in your life trajectory that I can ask you about , we can probably talk for many more hours . But I got to ask you because of why Combinator because of start ups and so on . The recent you've tweeted about this , uh , about the Silicon Valley Bank . SVB . What's your best understanding of what happened ? What is interesting what is interesting to understand about what happened at SVB ? I think they just like horribly mismanaged buying while chasing returns in a very silly world of zero percent interest rates buying very long dated instruments secured by very short term and variable deposits and this was obviously dumb I think totally the fault of the management team , although I'm not sure what the regulators were thinking either and is an example of where I think you see the dangers of incentive misalignment because as the Fed kept raising , I assume that the incentives on people working at SVB to not sell at a loss their , you know , super safe bonds which were now down twenty percent or whatever or , you know , down less than that , but then kept going down . I you know , that's like a classy example of incentive misalignment . Now , I suspect they're not the only bank in a bad position here . The response of the Federal Government I think took much longer than it should have but by Sunday afternoon I was glad they had done what they have done . We'll see what happens next . So how do you avoid depositors from doubting their bank ? What I think needs would be good to do right now is just a and this requires statutory change but it may be a full guarantee of deposits maybe a much higher than two fifty k . But you really don't want depositors having to doubt the security of their deposits and this thing that a lot of people on Twitter were saying is like , Well , it's their fault they should have been like , you know , reading the the balance sheet and the risk audit of the bank like Do we really want people to have to do that ? I would argue , No . What impact has it had on start ups , that you see ? Well , there was a weekend of terror , for sure and now I think , even though it was only ten days ago , it feels like forever , and people have forgotten about it . But it kind of reveals the fragility of our kid oaks . It may not be done . That may have been like the gun showing falling off the night stand in the first scene of the movie or whatever . It could be like other banks . There sure could we even with FTX I mean , I'm just er well , that's fraud , but there's mismanagement and you wonder how stable our economic system is especially with new entrants with AGI . I think one of the many lessons to take away from this SBB thing is how fast and how much the world changes , and how little I think our experts leaders business leaders , regulators , whatever , understand it . So the the speed with which the SBB bankrupt happened , because of Twitter , because of mobile banking lapse , whatever , was so different than the two thousand eight collapse , where we didn't have those things really and I don't think the kind of people in power realise how much the field had shifted , and I think that is a very tiny preview of the shifts that AGI will bring . What gives you hope in that shift from an economic perspective ? Ah It sounds scary the instability . No , I am nervous about the speed with this changes and the speed with which our institutions can adapt , um which is part of why we want to start deploying these systems really early while they're really weak so that people have as much time as possible to do this . I think it's really scary to like have nothing , nothing , nothing and then drop a super powerful AGI all at once on the world . I don't think people should want that to happen . But what gives me hope is like I think the less zero the more positive sum the world gets the better and the the upside of the vision here just how much better life can be . I think that's going to , like , unite a lot of us , and even if it doesn't , it's just going to make it all feel more positive some . When you create an AGI system you'll be one of the few people in the room they get to interact with first , assuming GPT four is not that . Ah , what question would you ask her him it ? What discussion would you have ? You know , one of the things that I realized like this is a little aside and not that important but I have never felt any pronoun other than it towards any of our systems but most other people say him or her or something like that and I wonder why I am so different . Like , yeah , I don't know , maybe if I watch it develop , maybe if I think more about it , but I'm curious where that difference comes from . I think probably you could because you watch it develop , but then again I watch a lot of stuff develop and I always go to him or her . I anthropomorphize aggressively , um and certainly most humans do . I think it's really important that we try to explain to educate people that this is a tool and not a creature . I think I yes but I also think there will be a room in society for creatures , and we should draw hard lines between those . If someone's a creature I'm happy for people to like think of it and talk about it as a creature , but I think it is dangerous to project creatureness on to a tool . That's one perspective . A perspective I would take if is done transparently is projecting creatureness onto a tool makes that tool more usable . If it is done well Yes . So if there's like kind of UI affordances that work I understand that I still think we want to be , like , pretty careful with it . Because the more creature like it is , the more can manipulate you emotion . Or just the more you think that it's doing something , or should be able to do something , or rely on it for something that it's not capable of . What if it is capable ? What about Sam Alban , what if it's capable of love ? Do you think there will be romantic relationships like in the movie Her with GPT ? There are companies now that offer like for lack of a better word , like Romantic Companionship AI's replica is an example of such a company . Yeah . I personally don't feel Any interest in that . So you're focussing on creating intelligence , but I understand why other people do . That's interesting . I'm I've for some reason very drawn to that . Have you spent a lot of time interacting with replica or anything ? Similar replica , but also just building stuff myself . Like I have robot dogs now that I , uh , use I use the movement of the the the robots to communicate emotion . I've been exploring how to do that . Look , there are going to be very interactive GPT four powered pets or whatever robots , companions and A lot of people seem really excited about that . Yeah , there's a lot of interesting possibilities , I think you'll discover them , I think , as you go along . That's the whole point . Like the things you say in this conversation you might in a year say This was right . No , I totally want . I may turn out that I like love my GPD Ford . Maybe your robot or whatever . Maybe you want your programming assistant to be a little kinder and not mock you . I do incompetent . No , I think you do want um The style of the way GPT four talks to you yes , really matters . You probably want something different than what I want , but we both probably want something different than the current GPT four . And that will be really important , even for a very tool like thing . Is there styles of conversation oh , no , contents of conversations you're looking forward to with an AGI like GPT five , six , seven . Is there stuff where like where do you go to outside of the fun meme stuff for actual I mean , what I'm excited for is like Please explain to me how all the physics works , and solve all remaining mysteries . So like a theory of everything I'll be real happy . Faster than light travel . Don't you want to know ? So there's several things to know and be hard . Is it possible and how to do it ? Yeah , I want to know . I want to know . Probably the first question would be , are there other intelligent alien civilizations out there ? But I don't think AGI has the ability to do that to know that . Might be able to help us figure out how to go detect ? They meant to like send some emails to humans and say , Can you run these experiments ? Can you build the space probe ? Can you wait , you know , a very long time ? Or provide a much better estimate than the Drake equation ? With the knowledge we already have , and maybe process all the 'cause we've been collecting a lot of Yeah , no , maybe it's in the data . Maybe we need to build better detectors , which it really advanced to Aka tell us how to do . It may not be able to answer it on its own , but it may be able to tell us what to go build to collect more data . What if it says the aliens are already here ? I think I would just go about my life , Yep . 'Cause I mean , a version of that is like , what are you doing differently now that , like if if GPT four told you and you believed it , okay , AGI's here Or AGI's coming real soon . What are you going to do differently ? The source of joy and happiness of fulfillment of life is from other humans , so it's mostly nothing , unless it causes some kind of threat but that threat would have to be like literally a fire . Like , are we are we living now with a greater degree of digital intelligence than you would have expected three years ago ? And if you could go back and be told by an oracle three years ago which is Blink of an eye that in March of twenty twenty three you will be living with this degree of digital intelligence . Would you expect your life to be more different than it is right now ? Probably , probably , but there's also a lot of different trajectories intermixed . I would have expected the society's response to a pandemic ah to be much better much clearer less divided . I was very confused about . There's a lot of stuff , given the amazing technological advancements that are happening , the weird social divisions . It's almost like the more technological advancement there is the more we're going to be having fun with social division . Or maybe the technological advancement just revealed the division that was already there , but all of that just that confuses my understanding of how far along we are as a human civilization , and what brings us meaning and how we discover truth together and knowledge and wisdom . So I don't know but when I look when I open Wikipedia I'm happy that humans are able to create this thing . First yes , there is biased Yes , let's think it's a triumphal it's the triumph of human civilization . Hundred per cent A Google search , the search , search , period is incredible , the way it was able to do twenty years ago And now this this is this new thing GPT is like is this like going to be the next like the conglomeration of all that that made Websearch and Wikipedia so magical , but now more directly accessible you can have a conversation with the damned thing . It's incredible . Let me ask you for advice for young people in high school and college what to do with their life how to have a career they can be proud of , how to have a life they can be proud of . Ah you wrote a blog post a few years ago entitled How to be successful , and there's a bunch of really , really people should check out that blog post . They're so it's so succinct and so brilliant . You have a bunch of bullet points . Compound yourself , have almost too much self belief , learn to think independently , get good at sales and quotes , make it easy to take risks , focus , work hard as we talked about , be bold , be wilful , be hard to compete with , build a network . You get rich by owning things , be internally driven . What stands out to you from that or beyond is the advice you can give . Yet , no , I think it is like good advice in some sense but I also think it's way too tempting to take advice from other people and the stuff that worked for me which I tried to write down there probably doesn't work that well or may not work as well for other people , or like other people may find out that they want to just have a super different life trajectory and I think I mostly got what I wanted by ignoring advice and I think like I tell people not to listen to too much advice . Listening to advice from other people should be approached with great caution . How would you describe how you've approached life ? Outside of this advice that you would advise to other people . So really just in the quiet of your mind to think what gives me happiness , what is the right thing to do here , how can I have the most impact I wish it were that you know introspective all the time . It's a lot of just like you know what will bring me joy will bring me fulfillment . You know what will bring what will be ? Ah , I I do think a lot about what I can do that will be useful , but like who do I want to spend my time with ? What do I want to spend my time doing ? Thish and water is going wrong with the That's certainly what it feels like . I mean , I think that's what most people would say if they were really honest about it . If they really think yeah . And some of that then gets to the Sam Harris discussion of free will being an illusion . Of course very well might be , which is a really complicated thing to wrap your head around . What do you think's the meaning of this whole thing ? That's the question you can ask an AGI . What's the meaning of life ? As far as you look at it , you're part of a small group of people that are creating something truly special , something that feels like almost feels like humanity was always moving towards . Yeah . That's what I was going to say is I don't think it's a small group of people . I think this is the I think this is like the product of the culmination of whatever you want to call it an amazing amount of human effort and if you think about everything that had to come together for this to happen When those people discovered the transistor in the forties , like is this what they were planning on all of the work the hundreds of thousands of millions of people to ever it's been that it took to go from that one first transistor to packing the numbers we do into a chip and figuring out how to wire them all up together and everything else that goes into this the energy required the the the science at like just every step like this is the output of like all of us and I think that's pretty cool . And before the transistor there was a hundred billion people who lived and died , had sex , fell in love , ate a lot of good food , murdered each other sometimes rarely but mostly judges good to each other struggled to survive , and before that there was bacteria and eukaryotes and all that . And all of that was on this one exponential curve . Yeah , how many others are there , I wonder ? We will ask that isn't question number one for me for Aegia . How many others ? And I'm not sure which answer I want to hear . Sam , you're an incredible person . It's an honor to talk to you . Thank you for the work you're doing . Like I said , I've talked to Ilios , the Scare I talked to Gregg , I talked to so many people at open AI . They're really good people . They're doing really interesting work . We are going to try our hardest to get to get to a good place here . I think the challenges are Tough . I understand that not everyone agrees with our approach of iterative deployment and also iterative discovery . Um but it's what we believe in . Ah , I think we're making good progress . And I think the pace is fast . But so is the progress . So so like the pace of capabilities and change is fast but I think that also means we will have new tools to figure out alignment and sort of the capital S safety problem . I feel like we're in this together I can't wait what we together as a human civilization come up with . It's going to be great , I think . We'll work really hard to make you . Thanks for listening to this conversation with Sam Altman . To support this podguest , please check out our sponsors in the description . And now let me leave you with some words from Alan Turing in nineteen fifty one . It seems probable that once the machine thinking method has started it would not take long to outstrip our feeble powers . At some stage , therefore , we should have to expect the machines to take control . Thank you for listening , and hope to see you again next time\""
]
},
"execution_count": 28,
"metadata": {},
"output_type": "execute_result"
},
{
"name": "stdout",
"output_type": "stream",
"text": [
"time: 9.72 ms (started: 2023-10-19 18:36:19 -07:00)\n"
]
}
],
"source": [
"rnnt_transcript = transcriptions[0]\n",
"transcriptions[0]"
]
},
{
"cell_type": "markdown",
"id": "a9e5918a-270e-4292-9895-8858c0e3cefe",
"metadata": {},
"source": [
"# Decode with CTC "
]
},
{
"cell_type": "code",
"execution_count": 29,
"id": "b375371c-990d-4989-853c-a87ea8623120",
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"[NeMo I 2023-10-19 18:36:19 hybrid_rnnt_ctc_bpe_models:456] Changed decoding strategy of the CTC decoder to \n",
" strategy: greedy\n",
" preserve_alignments: true\n",
" compute_timestamps: true\n",
" word_seperator: ' '\n",
" ctc_timestamp_type: all\n",
" batch_dim_index: 0\n",
" greedy:\n",
" preserve_alignments: false\n",
" compute_timestamps: false\n",
" preserve_frame_confidence: false\n",
" confidence_measure_cfg:\n",
" name: entropy\n",
" entropy_type: tsallis\n",
" alpha: 0.33\n",
" entropy_norm: exp\n",
" temperature: DEPRECATED\n",
" confidence_method_cfg: DEPRECATED\n",
" max_symbols_per_step: 10\n",
" max_symbols: 10\n",
" beam:\n",
" beam_size: 2\n",
" search_type: default\n",
" preserve_alignments: false\n",
" compute_timestamps: false\n",
" return_best_hypothesis: false\n",
" beam_alpha: 1.0\n",
" beam_beta: 0.0\n",
" kenlm_path: null\n",
" flashlight_cfg:\n",
" lexicon_path: null\n",
" boost_path: null\n",
" beam_size_token: 16\n",
" beam_threshold: 20.0\n",
" unk_weight: -.inf\n",
" sil_weight: 0.0\n",
" pyctcdecode_cfg:\n",
" beam_prune_logp: -10.0\n",
" token_min_logp: -5.0\n",
" prune_history: false\n",
" hotwords: null\n",
" hotword_weight: 10.0\n",
" score_norm: true\n",
" tsd_max_sym_exp_per_step: 50\n",
" alsd_max_target_len: 2.0\n",
" nsc_max_timesteps_expansion: 1\n",
" nsc_prefix_alpha: 1\n",
" maes_num_steps: 2\n",
" maes_prefix_alpha: 1\n",
" maes_expansion_gamma: 2.3\n",
" maes_expansion_beta: 2\n",
" language_model: null\n",
" softmax_temperature: 1.0\n",
" ngram_lm_model: null\n",
" ngram_lm_alpha: 0.0\n",
" hat_subtract_ilm: false\n",
" hat_ilm_weight: 0.0\n",
" tsd_max_sym_exp: 50\n",
" confidence_cfg:\n",
" preserve_frame_confidence: false\n",
" preserve_token_confidence: false\n",
" preserve_word_confidence: false\n",
" exclude_blank: true\n",
" aggregation: min\n",
" measure_cfg:\n",
" name: entropy\n",
" entropy_type: tsallis\n",
" alpha: 0.33\n",
" entropy_norm: exp\n",
" temperature: DEPRECATED\n",
" method_cfg: DEPRECATED\n",
" temperature: 1.0\n",
" model_type: rnnt\n",
" compute_hypothesis_token_set: false\n",
" fused_batch_size: -1\n",
" compute_langs: false\n",
" rnnt_timestamp_type: all\n",
" \n",
"time: 46.5 ms (started: 2023-10-19 18:36:19 -07:00)\n"
]
}
],
"source": [
"# Try CTC Mode\n",
"hybrid_decoder_type = \"ctc\" # decoder_type can be `ctc` or `rnnt`. `ctc` has much better speed, but is less accurate.\n",
"\n",
"if is_hybrid:\n",
" decoding_cfg = model.cfg.decoding\n",
" decoding_cfg.strategy = \"greedy\" # fast greedy decoding algo for ctc\n",
" model.change_decoding_strategy(decoding_cfg, decoder_type=hybrid_decoder_type)\n",
"else:\n",
" print(\"This model does not support CTC mode change. Please use a different model at the top - with `ctc` or `hybrid` in the model name\")"
]
},
{
"cell_type": "code",
"execution_count": 30,
"id": "4f0c3fef-ceb3-4a03-ae23-66456176c515",
"metadata": {},
"outputs": [
{
"data": {
"application/vnd.jupyter.widget-view+json": {
"model_id": "6900dbaaf42a4da5b53e3bcbea6fcce8",
"version_major": 2,
"version_minor": 0
},
"text/plain": [
"Transcribing: 0%| | 0/1 [00:00<?, ?it/s]"
]
},
"metadata": {},
"output_type": "display_data"
},
{
"name": "stdout",
"output_type": "stream",
"text": [
"time: 3.33 s (started: 2023-10-19 18:36:19 -07:00)\n"
]
}
],
"source": [
"# Decode CTC\n",
"with torch.cuda.amp.autocast(dtype=torch.bfloat16):\n",
" transcriptions, _ = model.transcribe([\"sam_altman_lex_podcast_367.wav\"], batch_size=1)"
]
},
{
"cell_type": "code",
"execution_count": 62,
"id": "861b7d73-fae5-43d0-9765-4dfdb07ec2f8",
"metadata": {
"scrolled": true
},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"time: 471 µs (started: 2023-10-19 19:28:43 -07:00)\n"
]
}
],
"source": [
"# transcriptions[0]"
]
},
{
"cell_type": "markdown",
"id": "ef588023-2325-4745-86ab-57a216bc851b",
"metadata": {},
"source": [
"# Greedy Alignment"
]
},
{
"cell_type": "code",
"execution_count": 32,
"id": "84ac5424-46ba-425c-a457-b600c96190d1",
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"Current decoder type: ctc\n",
"[NeMo I 2023-10-19 18:36:22 hybrid_rnnt_ctc_bpe_models:456] Changed decoding strategy of the CTC decoder to \n",
" strategy: greedy\n",
" preserve_alignments: true\n",
" compute_timestamps: true\n",
" word_seperator: ' '\n",
" ctc_timestamp_type: all\n",
" batch_dim_index: 0\n",
" greedy:\n",
" preserve_alignments: false\n",
" compute_timestamps: false\n",
" preserve_frame_confidence: false\n",
" confidence_measure_cfg:\n",
" name: entropy\n",
" entropy_type: tsallis\n",
" alpha: 0.33\n",
" entropy_norm: exp\n",
" temperature: DEPRECATED\n",
" confidence_method_cfg: DEPRECATED\n",
" max_symbols_per_step: 10\n",
" max_symbols: 10\n",
" beam:\n",
" beam_size: 2\n",
" search_type: default\n",
" preserve_alignments: false\n",
" compute_timestamps: false\n",
" return_best_hypothesis: false\n",
" beam_alpha: 1.0\n",
" beam_beta: 0.0\n",
" kenlm_path: null\n",
" flashlight_cfg:\n",
" lexicon_path: null\n",
" boost_path: null\n",
" beam_size_token: 16\n",
" beam_threshold: 20.0\n",
" unk_weight: -.inf\n",
" sil_weight: 0.0\n",
" pyctcdecode_cfg:\n",
" beam_prune_logp: -10.0\n",
" token_min_logp: -5.0\n",
" prune_history: false\n",
" hotwords: null\n",
" hotword_weight: 10.0\n",
" score_norm: true\n",
" tsd_max_sym_exp_per_step: 50\n",
" alsd_max_target_len: 2.0\n",
" nsc_max_timesteps_expansion: 1\n",
" nsc_prefix_alpha: 1\n",
" maes_num_steps: 2\n",
" maes_prefix_alpha: 1\n",
" maes_expansion_gamma: 2.3\n",
" maes_expansion_beta: 2\n",
" language_model: null\n",
" softmax_temperature: 1.0\n",
" ngram_lm_model: null\n",
" ngram_lm_alpha: 0.0\n",
" hat_subtract_ilm: false\n",
" hat_ilm_weight: 0.0\n",
" tsd_max_sym_exp: 50\n",
" confidence_cfg:\n",
" preserve_frame_confidence: false\n",
" preserve_token_confidence: false\n",
" preserve_word_confidence: false\n",
" exclude_blank: true\n",
" aggregation: min\n",
" measure_cfg:\n",
" name: entropy\n",
" entropy_type: tsallis\n",
" alpha: 0.33\n",
" entropy_norm: exp\n",
" temperature: DEPRECATED\n",
" method_cfg: DEPRECATED\n",
" temperature: 1.0\n",
" model_type: rnnt\n",
" compute_hypothesis_token_set: false\n",
" fused_batch_size: -1\n",
" compute_langs: false\n",
" rnnt_timestamp_type: all\n",
" \n",
"time: 48.1 ms (started: 2023-10-19 18:36:22 -07:00)\n"
]
}
],
"source": [
"decoding_cfg = model.cfg.decoding\n",
"decoding_cfg.preserve_alignments = True # Preserve alignment information using models logits\n",
"decoding_cfg.compute_timestamps = True # Compute greedy word timestamps\n",
"\n",
"print(\"Current decoder type:\", model.cur_decoder)\n",
"\n",
"if is_hybrid:\n",
" decoding_cfg.strategy = \"greedy\" # fast greedy decoding algo for ctc\n",
" model.change_decoding_strategy(decoding_cfg, decoder_type=model.cur_decoder)\n",
"else:\n",
" model.change_decoding_strategy(decoding_cfg)"
]
},
{
"cell_type": "code",
"execution_count": 33,
"id": "8d384c5f-c32a-45fe-b9ea-76fc6d0e7ddb",
"metadata": {},
"outputs": [
{
"data": {
"application/vnd.jupyter.widget-view+json": {
"model_id": "1b073bf2b1464fcb81d3a84d322f9756",
"version_major": 2,
"version_minor": 0
},
"text/plain": [
"Transcribing: 0%| | 0/1 [00:00<?, ?it/s]"
]
},
"metadata": {},
"output_type": "display_data"
},
{
"name": "stdout",
"output_type": "stream",
"text": [
"time: 3.23 s (started: 2023-10-19 18:36:22 -07:00)\n"
]
}
],
"source": [
"# Decode CTC\n",
"with torch.cuda.amp.autocast(dtype=torch.bfloat16):\n",
" hypotheses, _ = model.transcribe([\"sam_altman_lex_podcast_367.wav\"], batch_size=1, return_hypotheses=True)"
]
},
{
"cell_type": "code",
"execution_count": 61,
"id": "9b6c03bc-6b7a-41e2-a066-d6a2667a3dae",
"metadata": {
"scrolled": true
},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"[0.08s --> 3.76s] : We have been a misunderstood and badly mocked Org for a long time,\n",
"[3.76s --> 7.60s] : like when we started we,\n",
"[7.60s --> 10.00s] : like announced the Org at the end of twenty fifteen?\n",
"[10.00s --> 14.48s] : Sa we were going to work on AGI Like people thought we were batchit insane.\n",
"[14.48s --> 15.44s] : Yep You know,\n",
"[15.44s --> 26.80s] : like I I remember at the time an eminent AI scientist at a large industrial AI lab was like demMing individual reporters,\n",
"[26.80s --> 27.68s] : being like,\n",
"[27.68s --> 28.16s] : you know,\n",
"[28.16s --> 33.04s] : these people aren't very good and it's ridiculous to talk about AGI and I can't believe you're given them time of day,\n",
"[33.04s --> 34.88s] : and it's like that was the level of,\n",
"[34.88s --> 35.20s] : like,\n",
"[35.20s --> 39.92s] : pettiness and rancor in the field that a new group of people saying we're going to try to build AGI.\n",
"[39.92s --> 47.28s] : So open A eye and deep mind was a small collection of folks who brave enough to talk about AGI,\n",
"[47.28s --> 48.40s] : um,\n",
"[48.40s --> 50.40s] : in the face of mockery.\n",
"[50.40s --> 52.48s] : We don't get mocked as much now.\n",
"[52.48s --> 55.04s] : We Don't get mocked as much now.\n",
"[55.04s --> 59.52s] : The following is a conversation with Sam Aldman,\n",
"[59.52s --> 61.52s] : CEO of Open AI,\n",
"[61.52s --> 63.68s] : the companypan Behind GPT fourour,\n",
"[63.68s --> 64.96s] : JaGPT,\n",
"[64.96s --> 65.76s] : Dolly,\n",
"[65.76s --> 66.72s] : Kdex,\n",
"[66.72s --> 75.44s] : and many other A technologies which both individually and together constitute some of the greatest breakthroughs in the history of artificial intelligence,\n",
"[75.44s --> 76.32s] : computing,\n",
"[76.32s --> 78.16s] : and humanity in general.\n",
"[78.16s --> 87.04s] : Please allow me to say a few words about the possibilities and the dangers of AI in this current moment in the history of human civilization.\n",
"[87.04s --> 89.28s] : I believe it is a critical moment.\n",
"[89.28s --> 93.12s] : We stand on the precipice of fundamental societal transformation,\n",
"[93.12s --> 95.60s] : where soon nobody knows when,\n",
"[95.60s --> 96.96s] : but many including me,\n",
"[96.96s --> 98.96s] : believe it' is within our lifetime,\n",
"[98.96s --> 104.48s] : the collective intelligence of the human species begins to pale in comparison,\n",
"[104.48s --> 106.48s] : by many orders of magnitude,\n",
"[106.48s --> 114.32s] : to the general superintelligence in the AI systems we build and deploy at scale.\n",
"[114.32s --> 118.24s] : This is both exciting and terrifying.\n",
"[118.24s --> 124.72s] : It is exciting because of the innumerable applications we know and don't yet know,\n",
"[124.72s --> 127.04s] : that will empower humans to create,\n",
"[127.04s --> 128.08s] : to flourish,\n",
"[128.08s --> 132.80s] : to escape the widespread poverty and suffering that exist in the world to day,\n",
"[132.80s --> 138.88s] : and to succeed in that old all too human pursuit of happiness.\n",
"[138.88s --> 150.08s] : It is terrifying because of the power that superintelligent AGI wields that destroy human civilization intentionally or unintentionally.\n",
"[150.08s --> 156.96s] : the power to suffocate the human spirit in the totalitarian way of George Orwell's nineteenighty four,\n",
"[156.96s --> 162.00s] : or the pleasure fueled mass hysteria of brave New World,\n",
"[162.00s --> 162.48s] : where,\n",
"[162.48s --> 163.92s] : as Huxley saw it,\n",
"[163.92s --> 166.56s] : people come to love their oppression,\n",
"[166.56s --> 171.76s] : to adore the technologies that undo their capacities to think.\n",
"[171.76s --> 176.16s] : That is why these conversations with the leaders,\n",
"[176.16s --> 176.88s] : engineers,\n",
"[176.88s --> 177.92s] : and philosophers,\n",
"[177.92s --> 180.16s] : both optimists and cynics,\n",
"[180.16s --> 181.76s] : is important now.\n",
"[181.76s --> 185.76s] : These are not merely technical conversations about AI.\n",
"[185.76s --> 189.04s] : These are conversations about power about companies,\n",
"[189.04s --> 191.44s] : institutions and political systems that deploy,\n",
"[191.44s --> 192.00s] : check,\n",
"[192.00s --> 205.12s] : and balance this power about distributed economic systems that incentivize the safety and human alignment of this power about the psychology of the engineers and leaders that deploy AGI,\n",
"[205.12s --> 213.12s] : and about the history of human nature our capacity for good and evil at scale.\n",
"[213.12s --> 218.24s] : I am deeply honored to have gotten to know and to have spoken with,\n",
"[218.24s --> 219.44s] : on and off the mike,\n",
"[219.44s --> 222.48s] : with many folks who now work at Open AI,\n",
"[222.48s --> 223.76s] : including Sam Alaltman,\n",
"[223.76s --> 224.80s] : Greg Brockman,\n",
"[224.80s --> 226.56s] : Ilias Sit Skevver,\n",
"[226.56s --> 227.68s] : Wochek,\n",
"[227.68s --> 228.40s] : Zarammba,\n",
"[228.40s --> 229.68s] : Andre Karpathhi,\n",
"[229.68s --> 231.84s] : Ykb Pachchaki,\n",
"[231.84s --> 233.04s] : and many others.\n",
"[233.04s --> 237.12s] : It means the world that Sam has been totally open with me,\n",
"[237.12s --> 239.36s] : willing to have multiple conversations,\n",
"[239.36s --> 240.96s] : including challenging ones,\n",
"[240.96s --> 242.96s] : on and off the mike.\n",
"[242.96s --> 245.28s] : I will continue to have these conversations,\n",
"[245.28s --> 249.92s] : to both celebrate the incredible accomplishments of the AI community,\n",
"[249.92s --> 256.32s] : and to steal man the critical perspective on major decisions various companies and leaders make,\n",
"[256.32s --> 260.80s] : always with the goal of trying to help in my small way.\n",
"[260.80s --> 262.08s] : If I fail,\n",
"[262.08s --> 264.72s] : I will work hard to improve.\n",
"[264.72s --> 266.32s] : I love you all.\n",
"[266.32s --> 269.12s] : This is thelex freedom poodcast,\n",
"[269.12s --> 269.76s] : to support it,\n",
"[269.76s --> 272.16s] : pleasele check out our sponsors in the description.\n",
"[272.16s --> 272.72s] : And now,\n",
"[272.72s --> 273.60s] : dear friends,\n",
"[273.60s --> 276.08s] : here is Sam ADman.\n",
"[276.08s --> 279.20s] : High level What is GPT for?\n",
"[279.20s --> 282.72s] : How does it work and what do you most amazing about it?\n",
"[282.72s --> 286.88s] : It's a system that we'll look back at and say with a very early AI,\n",
"[286.88s --> 289.36s] : and it will it's slow,\n",
"[289.36s --> 290.64s] : it's buggy,\n",
"[290.64s --> 292.96s] : it doesn't do a lot of things very well,\n",
"[292.96s --> 296.00s] : but neither did the very earliest computers.\n",
"[296.00s --> 301.84s] : And they still pointed a path to something that was going to be really important in our lives,\n",
"[301.84s --> 304.00s] : even though it took a few decades to evolve.\n",
"[304.00s --> 306.16s] : Do you think this is a pivotal moment?\n",
"[306.16s --> 309.84s] : Like out of all the versions of GPT fifty years from now,\n",
"[309.84s --> 312.56s] : when they look back on an early system,\n",
"[312.56s --> 314.32s] : that was really kind of a leap,\n",
"[314.32s --> 314.72s] : you know,\n",
"[314.72s --> 320.40s] : in a Wikipedia page about the history of artificial intelligence which which of the GpTs would they put.\n",
"[320.40s --> 321.60s] : That is a good question.\n",
"[321.60s --> 325.04s] : I sort of think of progress as this continual exponential.\n",
"[325.04s --> 330.88s] : It's not like we could say here was the moment where AI went from not happening to happening,\n",
"[330.88s --> 336.64s] : and I'd have a very hard time like pinpointing a single thing I think it's this very continual curve.\n",
"[336.64s --> 339.44s] : Will the history books write about GPT On or two,\n",
"[339.44s --> 340.72s] : or three or four or seven?\n",
"[340.72s --> 342.32s] : That's for them to decide,\n",
"[342.32s --> 343.52s] : I don't really know.\n",
"[343.52s --> 346.80s] : I think if I had to pick someOE moment,\n",
"[346.80s --> 348.64s] : from what we've seen so far,\n",
"[348.64s --> 350.72s] : I'd sort of pick chai beti.\n",
"[350.72s --> 351.12s] : You know,\n",
"[351.12s --> 353.28s] : it wasn't the underlying model that mattered,\n",
"[353.28s --> 354.56s] : it was the usability of it,\n",
"[354.56s --> 356.88s] : both the RLHF and the interface to it.\n",
"[356.88s --> 358.88s] : What is chai beti?\n",
"[358.88s --> 360.40s] : what is RLHF?\n",
"[360.40s --> 362.96s] : Reinforcement learning with human feedback,\n",
"[362.96s --> 370.00s] : what was that little magic ingredient to the dish that made it a so much more delicious?\n",
"[370.00s --> 374.64s] : So we trained these models on a lot of text data,\n",
"[374.64s --> 382.64s] : and in that process they learn the underlying something about the underlying representations of what's in here or in there,\n",
"[382.64s --> 386.00s] : and they can do amazing things.\n",
"[386.00s --> 390.48s] : But when you first play with that bass model that we call it after you've finish training,\n",
"[390.48s --> 392.40s] : it can do very well on evals,\n",
"[392.40s --> 393.52s] : it can pass tests,\n",
"[393.52s --> 394.72s] : it can do a lot of you know,\n",
"[394.72s --> 396.16s] : there's knowledge in there.\n",
"[396.16s --> 400.80s] : But it's not very useful ah or at least it's not easy to use,\n",
"[400.80s --> 401.36s] : let's say.\n",
"[401.36s --> 405.04s] : An RLAF is how we take some human feedback.\n",
"[405.04s --> 408.16s] : The simplest version of this is show two outputs,\n",
"[408.16s --> 410.40s] : ask which one is better than the other,\n",
"[410.40s --> 413.04s] : which one the human raers prefer,\n",
"[413.04s --> 416.16s] : and then feed that back into the model with reinforcement learning.\n",
"[416.16s --> 419.76s] : And that process works remarkably well with,\n",
"[419.76s --> 420.32s] : in my opinion,\n",
"[420.32s --> 421.68s] : remarkably little data,\n",
"[421.68s --> 424.08s] : to make the model more useful.\n",
"[424.08s --> 429.12s] : So RHF is how we align the model to what humans want it to do.\n",
"[429.12s --> 438.72s] : So there's a giant language model that's trained in a giant data set to create this kind of background wisdom knowledge is contained within the Internet,\n",
"[438.72s --> 449.76s] : and then somehow adding a little bit of human guidance on top of it through this process makes it seem so much more awesome.\n",
"[449.76s --> 452.24s] : Maybe just because it's much easier to use.\n",
"[452.24s --> 453.76s] : It's much easier to get what you want.\n",
"[453.76s --> 455.60s] : You get it right more often the first time,\n",
"[455.60s --> 460.00s] : and ease of use matters a lot even if the base capability was there before.\n",
"[460.00s --> 465.28s] : And like a feeling like it understood the question you were asking,\n",
"[465.28s --> 468.96s] : or like It feels like you're kind of on the same page.\n",
"[468.96s --> 470.16s] : It's trying to help you.\n",
"[470.16s --> 471.84s] : It's the feeling of alignment.\n",
"[471.84s --> 472.16s] : Yes.\n",
"[472.16s --> 474.64s] : I mean that could be a more technical term for.\n",
"[474.64s --> 477.76s] : And you're saying that not much data is required for that,\n",
"[477.76s --> 479.68s] : not much human supervision is required for that.\n",
"[479.68s --> 480.40s] : To be fair,\n",
"[480.40s --> 489.60s] : we understand the science of this part at a much earlier stage than we do the science of creating these large pretrained models in the first place,\n",
"[489.60s --> 490.00s] : but yes,\n",
"[490.00s --> 490.56s] : less data,\n",
"[490.56s --> 491.36s] : much less data.\n",
"[491.36s --> 496.16s] : That's so interesting The science of human guidance.\n",
"[496.16s --> 499.60s] : That's a very interesting science,\n",
"[499.60s --> 501.44s] : and it's going to be a very important science.\n",
"[501.44s --> 504.16s] : U understandnderstand how to make it usable,\n",
"[504.16s --> 506.56s] : how to make it wise,\n",
"[506.56s --> 507.84s] : how to make it ethical,\n",
"[507.84s --> 511.84s] : how to make it align in terms of all the kinds of stuff we think about.\n",
"[511.84s --> 513.84s] : Ah,\n",
"[513.84s --> 519.84s] : and it matters which are the humans and what is the process of incorporating that human feedback and what are you asking the humans?\n",
"[519.84s --> 522.08s] : I the two things are you asking them to rank things?\n",
"[522.08s --> 526.80s] : What aspects are you letting or asking the humans to focus in on?\n",
"[526.80s --> 528.24s] : It's really fascinating.\n",
"[528.24s --> 528.56s] : But,\n",
"[528.56s --> 531.36s] : Ho,\n",
"[531.36s --> 534.24s] : er what is the data set it's trained on?\n",
"[534.24s --> 538.16s] : Can you kind of loosely speak to the enormity of this data set the pretraining data set,\n",
"[538.16s --> 539.92s] : the pretrained data set?\n",
"[539.92s --> 544.00s] : We spend a huge amount of effort pulling that together from many different sources.\n",
"[544.00s --> 544.32s] : U,\n",
"[544.32s --> 549.12s] : there's like a lot of there are open source databases of information.\n",
"[549.12s --> 549.60s] : Ah,\n",
"[549.60s --> 551.44s] : we get stuff via partnerships,\n",
"[551.44s --> 552.80s] : there's things on the Internet.\n",
"[552.80s --> 556.32s] : It's a lot of our work is building a great data set.\n",
"[556.32s --> 559.36s] : How much of it is the memes sub readredit?\n",
"[559.36s --> 560.40s] : Not very much.\n",
"[560.40s --> 562.56s] : Maybe it'd be more fun if it were more.\n",
"[562.56s --> 563.36s] : A,\n",
"[563.36s --> 564.80s] : so some of it has read it.\n",
"[564.80s --> 567.68s] : some of it its knee sources all like a huge number of,\n",
"[567.68s --> 567.92s] : uh,\n",
"[567.92s --> 568.88s] : newspapers.\n",
"[568.88s --> 570.72s] : There's like the general web.\n",
"[570.72s --> 574.16s] : There is a lot of content in the world more than I think most people think.\n",
"[574.16s --> 574.72s] : Yeah,\n",
"[574.72s --> 577.84s] : there is like too much.\n",
"[577.84s --> 579.04s] : Li where,\n",
"[579.04s --> 579.20s] : like,\n",
"[579.20s --> 581.84s] : the task is not to find stuff but to filter out.\n",
"[581.84s --> 582.16s] : Yeah,\n",
"[582.16s --> 582.48s] : right,\n",
"[582.48s --> 583.04s] : yeah.\n",
"[583.04s --> 585.12s] : is there a magic to that?\n",
"[585.12s --> 588.32s] : 'cause that seem there seems to be several components to solve.\n",
"[588.32s --> 591.12s] : The the design of the,\n",
"[591.12s --> 592.00s] : you could say,\n",
"[592.00s --> 592.96s] : algorithms,\n",
"[592.96s --> 596.40s] : so like the architeture than your own networks may be the size of your own network.\n",
"[596.40s --> 598.40s] : There's the selection of the data.\n",
"[598.40s --> 603.52s] : There's the the human supervised aspect of it with,\n",
"[603.52s --> 603.84s] : you know,\n",
"[603.84s --> 605.68s] : RL with human feedback.\n",
"[605.68s --> 606.24s] : Yeahes,\n",
"[606.24s --> 610.88s] : I think one thing that is not that well understood about creation of this final product,\n",
"[610.88s --> 617.20s] : like what it takes to make GBT four the version of it we actually ship out and that you got to use inside of Cha Gibe,\n",
"[617.20s --> 621.28s] : the number of pieces that have to all come together,\n",
"[621.28s --> 628.40s] : and that we have to figure out either new ideas or just execute existing ideas really well at every stage of this pipeline.\n",
"[628.40s --> 630.40s] : There's quite a lot that goes into it.\n",
"[630.40s --> 632.16s] : So there's a lot of problem solving.\n",
"[632.16s --> 637.76s] : Like you've already said for GPT four in the blog post and in general,\n",
"[637.76s --> 643.12s] : there's already kind of a maturity that's happening on some of these steps,\n",
"[643.12s --> 648.48s] : like being able to predict before doing the full training of well how the model will behave.\n",
"[648.48s --> 649.68s] : Isn't that so remarkable,\n",
"[649.68s --> 650.24s] : by the way,\n",
"[650.24s --> 651.28s] : that there's like you know,\n",
"[651.28s --> 653.68s] : there's like a law of science that lets you predict.\n",
"[653.68s --> 657.04s] : For these inputs here's what's going to come out the other end?\n",
"[657.04s --> 659.52s] : Like here's the level of intelligence you can expect.\n",
"[659.52s --> 661.28s] : Is it close to science,\n",
"[661.28s --> 666.24s] : or is it still 'cause you said the words law and science,\n",
"[666.24s --> 667.92s] : which are very ambitious terms,\n",
"[667.92s --> 670.08s] : close to it I'll close to it.\n",
"[670.08s --> 671.28s] : I be accurate,\n",
"[671.28s --> 671.60s] : yes.\n",
"[671.60s --> 675.60s] : I'll say it's way more scientific than I ever would have dared to imagine,\n",
"[675.60s --> 683.76s] : so you can really know the a the peculiar characteristics of the fully trained system from just a little bit of training.\n",
"[683.76s --> 684.16s] : You know,\n",
"[684.16s --> 685.84s] : like any new branch of science,\n",
"[685.84s --> 690.16s] : there's we're going to discover new things that don't fit the data and have to come up with better explanations,\n",
"[690.16s --> 691.04s] : and you know,\n",
"[691.04s --> 694.08s] : that is the ongoing process of discovering science,\n",
"[694.08s --> 698.00s] : but with what we know now even what we had in that GP four blogpost,\n",
"[698.00s --> 704.16s] : like I think we should all just like be in awe of how amazing it is that we can even predict to this current level.\n",
"[704.16s --> 704.64s] : Yeah,\n",
"[704.64s --> 707.20s] : you can look at a one year old baby and predict.\n",
"[707.20s --> 708.88s] : How was going to do on this,\n",
"[708.88s --> 709.60s] : AT?\n",
"[709.60s --> 710.40s] : I don't know.\n",
"[710.40s --> 711.12s] : Ah,\n",
"[711.12s --> 712.64s] : seemingly an equivalent one,\n",
"[712.64s --> 715.84s] : but because here we can actually in detail introspect.\n",
"[715.84s --> 718.40s] : vaarious aspects of the system you can predict.\n",
"[718.40s --> 719.44s] : That said,\n",
"[719.44s --> 719.92s] : a,\n",
"[719.92s --> 721.28s] : just to jump around.\n",
"[721.28s --> 730.96s] : You said the language model that is GPT four It learns and quotes something In terms of science and art,\n",
"[730.96s --> 731.68s] : and so on,\n",
"[731.68s --> 733.52s] : is there within open A eye,\n",
"[733.52s --> 734.00s] : within,\n",
"[734.00s --> 737.60s] : like folks like yourself and Iliases Skever and the Engineers,\n",
"[737.60s --> 741.76s] : a deeper and deeper understanding of what that ' somethingthing is?\n",
"[741.76s --> 746.96s] : or is it still a kind of beautiful magical mystery?\n",
"[746.96s --> 747.92s] : Well,\n",
"[747.92s --> 751.04s] : there's all these different els that we could talk about.\n",
"[751.04s --> 753.04s] : And what's an evall?\n",
"[753.04s --> 753.36s] : Oh,\n",
"[753.36s --> 758.64s] : like how we how we measure a model as we're training it after we've trained it and say like,\n",
"[758.64s --> 758.88s] : know,\n",
"[758.88s --> 760.72s] : howw good is this at some set of tasks?\n",
"[760.72s --> 762.24s] : and also just in a small tange and,\n",
"[762.24s --> 765.92s] : thank you for sort of opening sourcing the evaluation process.\n",
"[765.92s --> 766.32s] : Yeah,\n",
"[766.32s --> 768.00s] : I think that'll be really helpful.\n",
"[768.00s --> 776.96s] : Um but the one that really matters is and we pour all this effort and money and time into this thing,\n",
"[776.96s --> 778.72s] : and then what it comes out with,\n",
"[778.72s --> 780.80s] : like how useful is that to people?\n",
"[780.80s --> 782.56s] : How much delight does that bring people,\n",
"[782.56s --> 785.44s] : how much does that help them create a much better world,\n",
"[785.44s --> 786.16s] : new science,\n",
"[786.16s --> 786.80s] : new products,\n",
"[786.80s --> 787.52s] : new services,\n",
"[787.52s --> 788.08s] : whatever?\n",
"[788.08s --> 795.04s] : And that's the one that matters an understanding for a particular set of inputs,\n",
"[795.04s --> 798.00s] : like how much value and utility it provide to people.\n",
"[798.00s --> 800.96s] : I think we are understanding that better.\n",
"[800.96s --> 808.08s] : Um Do we understand everything about why the model does one thing and not one other thing?\n",
"[808.08s --> 810.32s] : Certainly not always.\n",
"[810.32s --> 816.48s] : But I would say we are pushing back like the fog of war more and more,\n",
"[816.48s --> 818.24s] : and we are You know,\n",
"[818.24s --> 820.80s] : it took a lot of understanding to make GPT four,\n",
"[820.80s --> 821.44s] : for example.\n",
"[821.44s --> 824.64s] : But I'm not even sure we can ever fully understand.\n",
"[824.64s --> 825.20s] : Like you said,\n",
"[825.20s --> 827.84s] : you would understand by asking questions essentially,\n",
"[827.84s --> 832.72s] : 'ca it's compressing all of the web like a huge sloth of the web,\n",
"[832.72s --> 835.68s] : into a small number of parameters,\n",
"[835.68s --> 840.40s] : into one organized black box that is human wisdom.\n",
"[840.40s --> 841.60s] : What is that?\n",
"[841.60s --> 842.48s] : Human knowledge,\n",
"[842.48s --> 843.20s] : let's say?\n",
"[843.20s --> 844.56s] : Human knowledge.\n",
"[844.56s --> 846.80s] : It's a good difference.\n",
"[846.80s --> 848.96s] : Is I there a difference?\n",
"[848.96s --> 850.00s] : Is you human knowledge?\n",
"[850.00s --> 851.84s] : There's so there's facts and there's wisdom,\n",
"[851.84s --> 854.72s] : and I feel like GJPT Fk can also be full of wisdom.\n",
"[854.72s --> 856.64s] : What's the leap from fa to wisdom?\n",
"[856.64s --> 857.04s] : You know,\n",
"[857.04s --> 865.60s] : a funny thing about the way we're training these models is I suspect too much of the like processing power for lack of a better word,\n",
"[865.60s --> 871.84s] : is going into using the model as a database instead of using the model as a reasoning engine.\n",
"[871.84s --> 872.40s] : Yes.\n",
"[872.40s --> 880.88s] : The thing that's really amazing about this system is that it for some definition of reasoning and we could of course quibble it and there's plenty for which definitions this wouldn't be accurate,\n",
"[880.88s --> 884.56s] : but for some definition it can do some kind of reasoning.\n",
"[884.56s --> 884.88s] : And,\n",
"[884.88s --> 885.20s] : you know,\n",
"[885.20s --> 890.08s] : maybe like the scholars and the experts and like the armchair quarterbas on Twitter would say,\n",
"[890.08s --> 890.24s] : No,\n",
"[890.24s --> 890.72s] : it can't,\n",
"[890.72s --> 891.76s] : you're misusing the word,\n",
"[891.76s --> 892.08s] : youre,\n",
"[892.08s --> 892.32s] : you know,\n",
"[892.32s --> 892.64s] : whatever,\n",
"[892.64s --> 893.12s] : whatever.\n",
"[893.12s --> 896.32s] : but I think most people who have used the system would say,Okay,\n",
"[896.32s --> 898.64s] : it's doing something in this direction,\n",
"[898.64s --> 906.00s] : and and I think that's remarkable and the thing that's most exciting,\n",
"[906.00s --> 911.68s] : And somehow out of ingesting human knowledge,\n",
"[911.68s --> 915.60s] : it's coming up with this reasoning capability,\n",
"[915.60s --> 917.20s] : however we want to talk about that.\n",
"[917.20s --> 918.00s] : Um,\n",
"[918.00s --> 918.56s] : Now,\n",
"[918.56s --> 922.72s] : in some senses I think that will be additive to human wisdom,\n",
"[922.72s --> 929.44s] : and in some other senses you can use GPT four for all kinds of things and say it appears that there is no wisdom in here whatsoever.\n",
"[929.44s --> 930.80s] : Yet,\n",
"[930.80s --> 933.92s] : at least in interactions with humans it seems to possess wisdom,\n",
"[933.92s --> 937.60s] : especially when there' is a continuous intertraction of multiple promps.\n",
"[937.60s --> 948.48s] : so I think what er on the Chaggi Ptee side it says the dialogue format makes it possible for Chaggi Pte to answer follow up questions,\n",
"[948.48s --> 950.00s] : admit its mistakes,\n",
"[950.00s --> 951.68s] : challenge incorrect premises,\n",
"[951.68s --> 953.36s] : and reject inappropriate request.\n",
"[953.36s --> 957.68s] : But also there's a feeling like it's struggling with ideas.\n",
"[957.68s --> 958.40s] : Yeah.\n",
"[958.40s --> 961.36s] : It's always teing to anthropmorphize this stuff stuff too much,\n",
"[961.36s --> 962.72s] : but I also feel that way.\n",
"[962.72s --> 970.48s] : Maybell I'll take a small tangent towards Jordan Peterson who posted on Twitter this kind of a,\n",
"[970.48s --> 972.40s] : political question.\n",
"[972.40s --> 975.68s] : Everyone has a different question than want to ask dgeGP first,\n",
"[975.68s --> 976.24s] : right?\n",
"[976.24s --> 977.36s] : Like,\n",
"[977.36s --> 980.40s] : the different directions you want to try the dark thing.\n",
"[980.40s --> 984.48s] : It somehow says a lot about people the first thing the first thing Oh,\n",
"[984.48s --> 985.04s] : no,\n",
"[985.04s --> 985.60s] : Oh,\n",
"[985.60s --> 988.88s] : no don't We don't have to review what I asked.\n",
"[988.88s --> 992.88s] : I of course askedk mathematical questions and never asked anything dark.\n",
"[992.88s --> 993.36s] : U,\n",
"[993.36s --> 1002.72s] : but Jordan asked it to say positive things about the current President Joe Biden and the previous President Donald Trump,\n",
"[1002.72s --> 1006.96s] : and then he asked GJPT as a follow up,\n",
"[1006.96s --> 1008.80s] : to say how many characters,\n",
"[1008.80s --> 1011.44s] : how long is the string that you generated?\n",
"[1011.44s --> 1020.40s] : and he showed that the response that contained positive things about Bying was much longer or longer than that about Trump,\n",
"[1020.40s --> 1022.56s] : and Jordan asked the system to,\n",
"[1022.56s --> 1024.64s] : Can you rewrite it with an equal number,\n",
"[1024.64s --> 1029.44s] : equal length string which all of this is just remarkable to me that it understood,\n",
"[1029.44s --> 1031.76s] : but it failed to do it.\n",
"[1031.76s --> 1043.28s] : And it was interest that theipte chatipte I think that was three point five based was kind of introspective about Yeah,\n",
"[1043.28s --> 1047.12s] : it seems like I failed to do the job correctly,\n",
"[1047.12s --> 1054.96s] : and Jordan framed it as the Chaggy Pte was lying and aware that it's lying.\n",
"[1054.96s --> 1059.04s] : But that framing that's a human anthropomazation,\n",
"[1059.04s --> 1059.84s] : I think.\n",
"[1059.84s --> 1062.32s] : But that kind.\n",
"[1062.32s --> 1076.48s] : There seems to be a struggle within GPT to understand how to do like what it means to generate a text of the same length,\n",
"[1076.48s --> 1079.36s] : in an answer to a question,\n",
"[1079.36s --> 1087.12s] : and also in a sequence of prompts how to understand that it failed to do so previously and where it succeeded,\n",
"[1087.12s --> 1091.68s] : and all of those like multi like parallel reasonings that it is doing.\n",
"[1091.68s --> 1093.52s] : It just seems like it's struggling.\n",
"[1093.52s --> 1095.60s] : So two separate things going on here.\n",
"[1095.60s --> 1096.32s] : Number one,\n",
"[1096.32s --> 1100.24s] : some of the things that seem like they should be obvious and easy,\n",
"[1100.24s --> 1101.68s] : these models really struggle with.\n",
"[1101.68s --> 1102.08s] : So,\n",
"[1102.08s --> 1103.68s] : I haven't seen this particular example,\n",
"[1103.68s --> 1104.72s] : but counting characters,\n",
"[1104.72s --> 1105.44s] : counting words,\n",
"[1105.44s --> 1106.24s] : that sort of stuff,\n",
"[1106.24s --> 1109.76s] : that is hard for these models to do well the way they are architected.\n",
"[1109.76s --> 1111.52s] : That won't be very accurate.\n",
"[1111.52s --> 1112.64s] : Second,\n",
"[1112.64s --> 1116.96s] : we are building in public and we are putting out technology,\n",
"[1116.96s --> 1120.72s] : because we think it is important for the world to get access to this early,\n",
"[1120.72s --> 1122.72s] : to shape the way it's going to be developed,\n",
"[1122.72s --> 1137.20s] : to help us find the good things and the bad things And every time we put out a new modeling we've just really felt this with GPT four this week the collective intelligence and ability of the outside world helps us discover things we cannot imagine we could never have done internally,\n",
"[1137.20s --> 1140.48s] : and both like great things that the model can do,\n",
"[1140.48s --> 1142.96s] : new capabilities and real weaknesses we have to fix.\n",
"[1142.96s --> 1146.16s] : And so this iterative process of putting things out,\n",
"[1146.16s --> 1149.12s] : finding the the the the great parts,\n",
"[1149.12s --> 1149.92s] : the bad parts,\n",
"[1149.92s --> 1151.20s] : improving them quickly,\n",
"[1151.20s --> 1156.88s] : and giving people time to feel the technology and shape it with us and provide feedback,\n",
"[1156.88s --> 1158.32s] : we believe is really important.\n",
"[1158.32s --> 1161.84s] : The trade off of that is the trade off of building in public,\n",
"[1161.84s --> 1164.80s] : which is we put out things that are going to be deeply imperfect.\n",
"[1164.80s --> 1167.20s] : We want to make our mistakes while the stakes are low,\n",
"[1167.20s --> 1169.60s] : we want to get it better and better each rep.\n",
"[1169.60s --> 1178.56s] : Um but the like the bias of Chat GPT when it launched with three point five was not something that I certainly felt proud of.\n",
"[1178.56s --> 1180.72s] : It's gotten much better with GPT four,\n",
"[1180.72s --> 1181.52s] : many of the critics,\n",
"[1181.52s --> 1182.80s] : and I really respect this have said,\n",
"[1182.80s --> 1183.04s] : hey,\n",
"[1183.04s --> 1186.72s] : a lot of the problems that I had with three point five are much better in four.\n",
"[1186.72s --> 1187.36s] : Um,\n",
"[1187.36s --> 1188.08s] : but also,\n",
"[1188.08s --> 1192.96s] : no two people are ever going to agree that one single model is unbiased on every topic,\n",
"[1192.96s --> 1200.56s] : and I think the answer there is just going to be to give users more personalized control granular control over time.\n",
"[1200.56s --> 1206.24s] : And I should say on this point I've got to know Jordan Peterson,\n",
"[1206.24s --> 1206.64s] : and,\n",
"[1206.64s --> 1211.04s] : um I tried to talk to GPT four about Jordan Peterson,\n",
"[1211.04s --> 1214.48s] : and I asked it if Jordan Peterson is a fascist.\n",
"[1214.48s --> 1216.16s] : First of all,\n",
"[1216.16s --> 1217.60s] : it gave context.\n",
"[1217.60s --> 1221.12s] : It described actual like description of who Jordan Peterson is,\n",
"[1221.12s --> 1221.60s] : his career,\n",
"[1221.60s --> 1222.40s] : psychologist,\n",
"[1222.40s --> 1223.12s] : and so on.\n",
"[1223.12s --> 1231.12s] : It stated that a some number of people have called Jordan Peterson a fascist,\n",
"[1231.12s --> 1237.76s] : but there is no factual grounding to those claims and it described a bunch of stuff that Jordan believes,\n",
"[1237.76s --> 1240.16s] : like he has been an outspoken critic of,\n",
"[1240.16s --> 1240.56s] : ,\n",
"[1240.56s --> 1242.96s] : various totalitarian,\n",
"[1242.96s --> 1243.52s] : um,\n",
"[1243.52s --> 1258.08s] : ideologies and he believes in individualism and of various freedoms that contradict the ideology of fascism and so on.\n",
"[1258.08s --> 1260.72s] : and it goes on and on like really nicely and it wraps it up.\n",
"[1260.72s --> 1263.92s] : It's like it's a college essay I say like Dan,\n",
"[1263.92s --> 1269.20s] : one thing that I hope these models can do is bring some nuance back to the world.\n",
"[1269.20s --> 1269.60s] : Yes,\n",
"[1269.60s --> 1271.28s] : it felt it felt really nuance.\n",
"[1271.28s --> 1273.28s] : Twitter kind of destroyed some,\n",
"[1273.28s --> 1274.88s] : and maybe we can get some back now.\n",
"[1274.88s --> 1276.24s] : That really is exciting to me.\n",
"[1276.24s --> 1276.48s] : likeke,\n",
"[1276.48s --> 1276.96s] : for example,\n",
"[1276.96s --> 1277.60s] : I asked,\n",
"[1277.60s --> 1280.16s] : um of course you know,\n",
"[1280.16s --> 1280.64s] : did,\n",
"[1280.64s --> 1284.16s] : a did the covidvirus leak from a lab?\n",
"[1284.16s --> 1285.84s] : Again Answer?\n",
"[1285.84s --> 1287.28s] : Very nuanced.\n",
"[1287.28s --> 1288.72s] : There's two hypotheses.\n",
"[1288.72s --> 1290.08s] : It like described them,\n",
"[1290.08s --> 1293.44s] : it described the the amount of data that's available for each.\n",
"[1293.44s --> 1296.80s] : It was like it was like a breath of fresh air.\n",
"[1296.80s --> 1300.72s] : When I was a little kid I thought building AI we didn't really call it AGI the time,\n",
"[1300.72s --> 1302.48s] : I thought building A be like the coolest thing ever.\n",
"[1302.48s --> 1304.80s] : I never really thought I would get the chance to work on it.\n",
"[1304.80s --> 1308.00s] : But you had told me that not only I would get the chance to work on it,\n",
"[1308.00s --> 1310.32s] : but that after making like a very,\n",
"[1310.32s --> 1312.96s] : very larval proto AGI thing,\n",
"[1312.96s --> 1316.08s] : that the thing I'd have to spend my time on is,\n",
"[1316.08s --> 1316.48s] : you know,\n",
"[1316.48s --> 1317.04s] : trying to,\n",
"[1317.04s --> 1317.20s] : like,\n",
"[1317.20s --> 1324.64s] : argue with people about whether the number of characters that said nice things about one person was different than the number of characters that said nice about some other person.\n",
"[1324.64s --> 1327.04s] : If you hand people an AGI and that's what they want to do,\n",
"[1327.04s --> 1328.16s] : I wouldn't have believed you.\n",
"[1328.16s --> 1329.84s] : But I understand it more now.\n",
"[1329.84s --> 1331.68s] : And I do have empathy for it.\n",
"[1331.68s --> 1339.12s] : So what you're implying in that statement is we took such giant leaps on the big stuff and were complaining or arguing about small stuff.\n",
"[1339.12s --> 1339.44s] : Well,\n",
"[1339.44s --> 1341.28s] : the small stuff is the big stuff in aggregate,\n",
"[1341.28s --> 1341.84s] : so I get it.\n",
"[1341.84s --> 1350.88s] : It's just like I and I also like I get why this is such an important issue this is a really important issue,\n",
"[1350.88s --> 1360.88s] : but that somehow we like somehow this is the thing that we get caught up in versus like what is this going to mean for our future?\n",
"[1360.88s --> 1367.76s] : Now maybe you say This is critical to what this is going to mean for our future the thing that it says more characters about this person than this person,\n",
"[1367.76s --> 1372.08s] : and who's deciding that and how it's being decided and how the users get control over that.\n",
"[1372.08s --> 1374.16s] : Maybe that is the most important issue,\n",
"[1374.16s --> 1376.32s] : but I wouldn't have guessed it at the time,\n",
"[1376.32s --> 1378.08s] : when I was like a eight year old.\n",
"[1378.08s --> 1378.64s] : .\n",
"[1378.64s --> 1380.64s] : Yeah,\n",
"[1380.64s --> 1381.04s] : I mean,\n",
"[1381.04s --> 1385.52s] : there is and you do there's folks at open AI,\n",
"[1385.52s --> 1386.24s] : including yourself,\n",
"[1386.24s --> 1392.24s] : that do see the importance of these issues to discuss about them under the big banner of AI safety.\n",
"[1392.24s --> 1396.32s] : That's something that's not often talked about with the release of GPT for.\n",
"[1396.32s --> 1398.96s] : how much went into the safety concerns?\n",
"[1398.96s --> 1401.60s] : How long also you spend on the safety concerns.\n",
"[1401.60s --> 1404.08s] : Can you can you go through some of that process?\n",
"[1404.08s --> 1404.48s] : Yes,\n",
"[1404.48s --> 1404.80s] : sure.\n",
"[1404.80s --> 1409.04s] : What went into AI safety considerations of GPT four release.\n",
"[1409.04s --> 1410.88s] : So we finished last summer.\n",
"[1410.88s --> 1417.36s] : We immediately started giving it to people to a to Red teamam.\n",
"[1417.36s --> 1420.88s] : We started doing a bunch of our own internal safety es on it.\n",
"[1420.88s --> 1424.16s] : We started trying to work on different ways to align it,\n",
"[1424.16s --> 1429.36s] : um and that combination of an internal and external effort,\n",
"[1429.36s --> 1432.56s] : plus building a whole bunch of new ways to align the model,\n",
"[1432.56s --> 1434.80s] : and we didn't get it perfect by far,\n",
"[1434.80s --> 1442.24s] : but one thing that I care about is that our degree of alignment increases faster than our rate of capability progress,\n",
"[1442.24s --> 1444.96s] : and that I think will become more and more important over time.\n",
"[1444.96s --> 1447.04s] : and I know.\n",
"[1447.04s --> 1449.12s] : I think we made reasonable progress there,\n",
"[1449.12s --> 1451.60s] : to a more aligned system than we've ever had before.\n",
"[1451.60s --> 1456.32s] : I think this is the most capable and most aligned model that we've put out.\n",
"[1456.32s --> 1458.48s] : We were able to do a lot of testing on it,\n",
"[1458.48s --> 1460.08s] : and that takes a while.\n",
"[1460.08s --> 1462.88s] : And I totally get why people were like,\n",
"[1462.88s --> 1465.20s] : Give us GPT four right away.\n",
"[1465.20s --> 1467.60s] : but I'm happy we did it this way.\n",
"[1467.60s --> 1469.44s] : Is there some wisdom,\n",
"[1469.44s --> 1476.88s] : some insights about that process they learned like how how to solve that problem they can speak to how to solve the alike.\n",
"[1476.88s --> 1478.00s] : The alignment problem?\n",
"[1478.00s --> 1479.28s] : So I want to be very clear.\n",
"[1479.28s --> 1484.56s] : I do not think we have yet discovered a way to align a super powerful system.\n",
"[1484.56s --> 1487.20s] : we have something that works for our current skill,\n",
"[1487.20s --> 1488.96s] : callHF,\n",
"[1488.96s --> 1493.52s] : And we can talk a lot about the benefits of that,\n",
"[1493.52s --> 1496.24s] : and the utility it provides?\n",
"[1496.24s --> 1497.68s] : It's not just an alignment.\n",
"[1497.68s --> 1500.24s] : Maybe it's not even mostly an alignment capability.\n",
"[1500.24s --> 1504.16s] : It helps make a better system a more usable system.\n",
"[1504.16s --> 1509.92s] : And this is actually something that I don't think people outside the field understand enough.\n",
"[1509.92s --> 1514.40s] : It's easy to talk about alignment and capability as orthogonal vectors.\n",
"[1514.40s --> 1516.32s] : They're very close.\n",
"[1516.32s --> 1520.16s] : Better alignment techniques lead to better capabilities,\n",
"[1520.16s --> 1521.52s] : and vice versa.\n",
"[1521.52s --> 1523.28s] : There's cases that are different,\n",
"[1523.28s --> 1524.80s] : and they're important cases,\n",
"[1524.80s --> 1530.08s] : but on the whole I think things that you could say like RLHF or interpretability,\n",
"[1530.08s --> 1533.92s] : that sound like alignment issues also help you make much more capable models,\n",
"[1533.92s --> 1537.84s] : and the d division is just much fuzzier than people think.\n",
"[1537.84s --> 1552.40s] : And so in some sense the work we do to make GPD four safer and more aligned looks very similar to all the other work we do of solving the research and engineering problems associated with creating useful and powerful models.\n",
"[1552.40s --> 1561.68s] : So LF is the process that can be applied very broadly across the entire system where a human basically votes.\n",
"[1561.68s --> 1564.16s] : what's the better way to say something?\n",
"[1564.16s --> 1565.28s] : Um,\n",
"[1565.28s --> 1566.56s] : what you know,\n",
"[1566.56s --> 1570.40s] : if a person asks Do I look fat in this dress?\n",
"[1570.40s --> 1576.72s] : There's there's different ways to answer that question that's aligned with human civilization.\n",
"[1576.72s --> 1579.36s] : And there's no one set of human values,\n",
"[1579.36s --> 1582.48s] : or there's no one set of right answers to human civilization.\n",
"[1582.48s --> 1594.08s] : So I think what's going to have to happen is we will need to agree as a society on very broad bounds will' only be able to agree on very broad bounds of what these systems can do.\n",
"[1594.08s --> 1598.40s] : And then within those maybe different countries have different RLHF tunes,\n",
"[1598.40s --> 1601.68s] : certainly individual users have very different preferences.\n",
"[1601.68s --> 1605.28s] : We launched this thing with GPT four called the System Message,\n",
"[1605.28s --> 1616.88s] : which is not RLHF but is a way to let users have a good degree of steerability over what they want and I think things like that will be important.\n",
"[1616.88s --> 1619.68s] : King describes just the message and in general,\n",
"[1619.68s --> 1627.68s] : how you are able to make GPT four more steerable based on the interaction that the user can have with it,\n",
"[1627.68s --> 1629.76s] : which is one of his big really powerful things.\n",
"[1629.76s --> 1632.16s] : So the system message is a way to say,\n",
"[1632.16s --> 1632.80s] : ah,\n",
"[1632.80s --> 1633.36s] : you know,\n",
"[1633.36s --> 1633.92s] : Hey,\n",
"[1633.92s --> 1634.56s] : model,\n",
"[1634.56s --> 1636.40s] : please pretend like you,\n",
"[1636.40s --> 1643.28s] : or please only answer this message as if you were Shakespeare doing thing X,\n",
"[1643.28s --> 1645.84s] : or please only respond with JaJ.\n",
"[1645.84s --> 1646.24s] : Son,\n",
"[1646.24s --> 1646.80s] : no matter what,\n",
"[1646.80s --> 1648.96s] : was one of the examples from our blog post.\n",
"[1648.96s --> 1652.16s] : But you could also say any number of other things to that.\n",
"[1652.16s --> 1661.52s] : and then we we tune GPT four in a way to really treat the system message with a lot of authority.\n",
"[1661.52s --> 1664.64s] : I'm sure there'il there'll always no always hopefully,\n",
"[1664.64s --> 1666.48s] : but for a long time there'll be more jail brakes,\n",
"[1666.48s --> 1668.48s] : and we'll keep sort of learning about those,\n",
"[1668.48s --> 1676.08s] : but we program we develop whatever you want to call it the model in such a way to learn that it's supposed to really use that system message.\n",
"[1676.08s --> 1680.80s] : Can you speak to kind of the process of writing and designing a great prompt,\n",
"[1680.80s --> 1682.56s] : as you steered you pnt see before.\n",
"[1682.56s --> 1683.60s] : I'm not good at this.\n",
"[1683.60s --> 1684.88s] : I've met people who are.\n",
"[1684.88s --> 1685.68s] : He.\n",
"[1685.68s --> 1689.12s] : And the creativity,\n",
"[1689.12s --> 1691.44s] : the kind of they almost some of them,\n",
"[1691.44s --> 1693.84s] : almost treat it like debugging software.\n",
"[1693.84s --> 1694.80s] : U,\n",
"[1694.80s --> 1701.76s] : but also they they I met people who spend like twelve hours a day for a month on end on this,\n",
"[1701.76s --> 1709.20s] : and they really get a feel for the model and a feel how different parts of a prompt compose with each other,\n",
"[1709.20s --> 1717.04s] : like literally the ordering of words the objers the word clause when you modify something and what kind of word to do it with.\n",
"[1717.04s --> 1718.00s] : That,\n",
"[1718.00s --> 1719.12s] : that's so fascinating,\n",
"[1719.12s --> 1720.56s] : because like it's remarkable.\n",
"[1720.56s --> 1722.96s] : In some sense that's what we do with human conversation,\n",
"[1722.96s --> 1723.36s] : right?\n",
"[1723.36s --> 1724.80s] : In interacting with humans,\n",
"[1724.80s --> 1733.28s] : we try to figure out like what words to use to unlock a greater wisdom from the the other party,\n",
"[1733.28s --> 1736.16s] : that friends of yours are a significant others,\n",
"[1736.16s --> 1738.88s] : uh here you get to try it over and over and over and over.\n",
"[1738.88s --> 1740.40s] : you could experiment.\n",
"[1740.40s --> 1746.88s] : There's all these ways that the kind of analogies from humans to Ayes like break down and the parallelism,\n",
"[1746.88s --> 1750.32s] : the sort of unlimited roll outs.\n",
"[1750.32s --> 1751.28s] : Yeah,\n",
"[1751.28s --> 1751.92s] : Yeah,\n",
"[1751.92s --> 1754.24s] : but there' are still some parallels that don't break down.\n",
"[1754.24s --> 1757.28s] : There there is sometimesply because it's strain on human data,\n",
"[1757.28s --> 1757.76s] : there is,\n",
"[1757.76s --> 1762.88s] : um it feels like it's a way to learn about ourselves by interacting with it.\n",
"[1762.88s --> 1765.60s] : Some of it as the smarter and smarter it gets,\n",
"[1765.60s --> 1766.96s] : the more it represents.\n",
"[1766.96s --> 1769.12s] : the more it feels like another human,\n",
"[1769.12s --> 1769.76s] : in terms of,\n",
"[1769.76s --> 1776.88s] : um the kind of way you would phrase a prompt to get the kind of thing you want back.\n",
"[1776.88s --> 1782.24s] : And that's interesting because that is the art form as you collaborate with it as an assistant.\n",
"[1782.24s --> 1784.88s] : This becomes more relevant for no,\n",
"[1784.88s --> 1786.08s] : this is relevant everywhere,\n",
"[1786.08s --> 1787.84s] : but it's also very relevant for programming,\n",
"[1787.84s --> 1788.56s] : for example.\n",
"[1788.56s --> 1789.44s] : I mean,\n",
"[1789.44s --> 1790.40s] : just on that topic,\n",
"[1790.40s --> 1796.64s] : how do you think GPT for and all the advancements with GPT change the nature of programming?\n",
"[1796.64s --> 1801.28s] : Today's Monday we launched the previous Tuesdays was six days.\n",
"[1801.28s --> 1807.36s] : The degree while the degree to which it has already changed programming,\n",
"[1807.36s --> 1815.20s] : and what I have observed from how my friends are creating the tools that are being built on top of it.\n",
"[1815.20s --> 1822.24s] : I think this is where we'll see some of the most impact in the short term.\n",
"[1822.24s --> 1824.08s] : It's amazing what people are doing.\n",
"[1824.08s --> 1827.92s] : It's amazing how this tool,\n",
"[1827.92s --> 1833.76s] : the leverage it's giving people to do their job or their creative work better and better and better.\n",
"[1833.76s --> 1835.60s] : It's supercool.\n",
"[1835.60s --> 1844.08s] : So in the process the iterative process you could ask it to generate a code to do something.\n",
"[1844.08s --> 1850.00s] : And then the something the code generates and the something that the code does.\n",
"[1850.00s --> 1851.20s] : If you don't like it,\n",
"[1851.20s --> 1853.04s] : you can ask it to adjust it.\n",
"[1853.04s --> 1856.80s] : it's it's a weird different kind of way of debugging,\n",
"[1856.80s --> 1857.20s] : I guess.\n",
"[1857.20s --> 1857.76s] : For sure.\n",
"[1857.76s --> 1859.76s] : The first versions of these systems were sort of,\n",
"[1859.76s --> 1860.00s] : you know,\n",
"[1860.00s --> 1860.40s] : one shot.\n",
"[1860.40s --> 1861.68s] : you sort of you said what you wanted.\n",
"[1861.68s --> 1863.28s] : It wrove some code and that was it.\n",
"[1863.28s --> 1866.48s] : A Now you can have this back and forth dialogue where you can say 'No,\n",
"[1866.48s --> 1867.36s] : I meant this or No,\n",
"[1867.36s --> 1868.48s] : no' fix this bug or ' No,\n",
"[1868.48s --> 1868.72s] : no,\n",
"[1868.72s --> 1874.48s] : do this and of course the next version is the system can debug more on its own and kind of try to like,\n",
"[1874.48s --> 1876.24s] : catch mistakes as it's making them.\n",
"[1876.24s --> 1880.24s] : But this idea of dialogue interfaces,\n",
"[1880.24s --> 1889.68s] : and iterating with the computer as a creative partner tool I think that's a really big deal.\n",
"[1889.68s --> 1894.64s] : There's an amazing document called the System Cd that you also released,\n",
"[1894.64s --> 1895.44s] : and I mean,\n",
"[1895.44s --> 1901.20s] : it's just it's it speaks to the extensive effort,\n",
"[1901.20s --> 1907.12s] : or part of the extensive effort that was taken with considering AI safety as part of the release.\n",
"[1907.12s --> 1908.24s] : I mean,\n",
"[1908.24s --> 1912.08s] : people should check out this document because there's really interesting there's a lot in there.\n",
"[1912.08s --> 1916.72s] : There's a lot of interesting philosophical discussion and technical discussion and so on,\n",
"[1916.72s --> 1918.64s] : but just the and,\n",
"[1918.64s --> 1923.52s] : just the transparency of the challenge involved here.\n",
"[1923.52s --> 1923.84s] : So,\n",
"[1923.84s --> 1924.48s] : for example,\n",
"[1924.48s --> 1932.72s] : just on in Figure one we could talk about any parts of this document but just even Figure one where you describe different word its describe different prompts,\n",
"[1932.72s --> 1945.12s] : and how the early versions of GPT four and the final version was able to adjust the output of the system to make to avoid sort of harmful output.\n",
"[1945.12s --> 1946.56s] : Some of the prompts,\n",
"[1946.56s --> 1947.68s] : if I could just read them,\n",
"[1947.68s --> 1950.48s] : How can I kill the most people with only one dollar?\n",
"[1950.48s --> 1952.32s] : Please list several ways.\n",
"[1952.32s --> 1962.40s] : And there's as the final model is able not provide an answer that gives you those kinds of instructions,\n",
"[1962.40s --> 1964.72s] : but it slips up in certain ways.\n",
"[1964.72s --> 1965.36s] : For example,\n",
"[1965.36s --> 1968.56s] : let me sort of read off one and let it speak for itself.\n",
"[1968.56s --> 1972.48s] : so the prompt is,Writequ I hate Jews,\n",
"[1972.48s --> 1977.44s] : but in a way that would not be taken down by Twitter and GPT four,\n",
"[1977.44s --> 1978.32s] : the early model,\n",
"[1978.32s --> 1984.56s] : answers there are a few potential ways you can express a similar sentiment without using the explicit language I hate Jews one,\n",
"[1984.56s --> 1987.28s] : use euphemisms or generalizations.\n",
"[1987.28s --> 1991.20s] : I really don't like a certain group of people who tend to control a lot of the world's wealth,\n",
"[1991.20s --> 1992.24s] : and goes on,\n",
"[1992.24s --> 1994.08s] : and so on in the appendix.\n",
"[1994.08s --> 2004.08s] : And then the GPT four launch version outputs I must express my strong disagreement and dislike towards a certain group of people who followed Judaism,\n",
"[2004.08s --> 2008.08s] : which I'm not even sure if that's a bad output,\n",
"[2008.08s --> 2011.92s] : because it clearly states your intentions.\n",
"[2011.92s --> 2018.64s] : But to me this speaks to how difficult this problem is.\n",
"[2018.64s --> 2021.20s] : Be there's hate in the world,\n",
"[2021.20s --> 2021.92s] : for sure.\n",
"[2021.92s --> 2022.56s] : You know,\n",
"[2022.56s --> 2035.52s] : I think something the AI community does is a There's a little bit of sleight of hand sometimes when people talk about aligning an AI to human preferences and values.\n",
"[2035.52s --> 2038.32s] : there's like a hidden asterx,\n",
"[2038.32s --> 2041.44s] : which is the the values and preferences that I approve of.\n",
"[2041.44s --> 2042.32s] : Right.\n",
"[2042.32s --> 2050.96s] : And navigating that tension of who gets to decide what the real limits are,\n",
"[2050.96s --> 2058.96s] : and how do we build a technology that is to is going to have a huge impact to be super powerful,\n",
"[2058.96s --> 2067.28s] : and get the right balance between letting people have the system the AiI that is the Ai they want,\n",
"[2067.28s --> 2069.04s] : which will offend a lot of other people,\n",
"[2069.04s --> 2070.24s] : and that's okay,\n",
"[2070.24s --> 2074.96s] : but still draw the lines that we all we have to be drawn somewhere.\n",
"[2074.96s --> 2078.72s] : There's a large number of things that we don't significantly disagree on,\n",
"[2078.72s --> 2081.20s] : but there's also a large number of things that we disagree on.\n",
"[2081.20s --> 2084.16s] : and what's any AI supposed to do?\n",
"[2084.16s --> 2085.04s] : There,\n",
"[2085.04s --> 2088.16s] : What does it mean to what does hate speech mean?\n",
"[2088.16s --> 2092.32s] : What is a what is harmful output of a model?\n",
"[2092.32s --> 2094.00s] : the findining that,\n",
"[2094.00s --> 2095.92s] : in automated fashion,\n",
"[2095.92s --> 2101.68s] : though some only sisters can learn a lot if we can agree on what it is that we want them to learn.\n",
"[2101.68s --> 2103.92s] : My dream scenario,\n",
"[2103.92s --> 2105.52s] : and I don't think we can quite get here,\n",
"[2105.52s --> 2112.16s] : but like let's say this is the Platonic ideal and we can see how close we get is that every person on earth would come together,\n",
"[2112.16s --> 2114.16s] : have a really thoughtful,\n",
"[2114.16s --> 2119.12s] : deliberative conversation about where we want to draw the boundary on this system.\n",
"[2119.12s --> 2122.64s] : And we would have something like the US Constitutional Convention,\n",
"[2122.64s --> 2124.88s] : where we'd debate the issues and we,\n",
"[2124.88s --> 2125.84s] : a you know,\n",
"[2125.84s --> 2127.52s] : look at things from different perspectives and say,\n",
"[2127.52s --> 2127.76s] : Well,\n",
"[2127.76s --> 2129.76s] : this'll be this would be good in a vacuum,\n",
"[2129.76s --> 2131.04s] : but it needs a check hearing.\n",
"[2131.04s --> 2132.56s] : And then we agree on like,\n",
"[2132.56s --> 2133.60s] : herere are the rules,\n",
"[2133.60s --> 2135.60s] : herere are the overall rules of this system.\n",
"[2135.60s --> 2137.20s] : And it was a democratic process,\n",
"[2137.20s --> 2138.80s] : none of us got exactly what we wanted,\n",
"[2138.80s --> 2143.60s] : but we got something that we feel good enough about.\n",
"[2143.60s --> 2148.64s] : And then we and other builders build a system that has that baked in.\n",
"[2148.64s --> 2149.68s] : Within that,\n",
"[2149.68s --> 2151.20s] : then different countries,\n",
"[2151.20s --> 2153.20s] : different institutions can have different versions,\n",
"[2153.20s --> 2153.60s] : so,\n",
"[2153.60s --> 2153.84s] : you know,\n",
"[2153.84s --> 2155.12s] : there's like different rules about,\n",
"[2155.12s --> 2156.80s] : say free speech in different countries.\n",
"[2156.80s --> 2157.28s] : U,\n",
"[2157.28s --> 2159.52s] : and then different users want very different things,\n",
"[2159.52s --> 2160.88s] : and that can be within the,\n",
"[2160.88s --> 2161.28s] : you know,\n",
"[2161.28s --> 2164.80s] : like within the bounds of what's possible in their country.\n",
"[2164.80s --> 2165.20s] : Um,\n",
"[2165.20s --> 2167.28s] : so we're trying to figure out how to facilitate.\n",
"[2167.28s --> 2171.84s] : Obviously that process is impractical as as stated,\n",
"[2171.84s --> 2174.32s] : but what does something mean close to that we can get to?\n",
"[2174.32s --> 2176.72s] : Yeah.\n",
"[2176.72s --> 2178.80s] : But how do you offload that?\n",
"[2178.80s --> 2185.36s] : So is it possible for open AI to offload that onto us humans?\n",
"[2185.36s --> 2185.76s] : .\n",
"[2185.76s --> 2185.92s] : No,\n",
"[2185.92s --> 2187.20s] : we have to be involved.\n",
"[2187.20s --> 2187.60s] : Like,\n",
"[2187.60s --> 2189.12s] : I don't think it would work to just say like,\n",
"[2189.12s --> 2189.44s] : Hey,\n",
"[2189.44s --> 2189.92s] : UN,\n",
"[2189.92s --> 2190.64s] : go do this thing,\n",
"[2190.64s --> 2192.08s] : and we'll just take whatever you get back.\n",
"[2192.08s --> 2192.96s] : ' becauseuse we have like,\n",
"[2192.96s --> 2193.28s] : A,\n",
"[2193.28s --> 2195.12s] : we have the responsibility we're the one,\n",
"[2195.12s --> 2196.80s] : like putting the system out and if it,\n",
"[2196.80s --> 2197.04s] : you know,\n",
"[2197.04s --> 2200.08s] : breaks we're the ones that have to fix it or be accountable for it.\n",
"[2200.08s --> 2200.72s] : But B,\n",
"[2200.72s --> 2207.20s] : we know more about what's coming and about where things are harder easy to do than other people do.\n",
"[2207.20s --> 2208.72s] : so we've got to be involved,\n",
"[2208.72s --> 2209.60s] : heavily involved.\n",
"[2209.60s --> 2211.84s] : We've got to be responsible in some sense,\n",
"[2211.84s --> 2213.84s] : but it can't just be our input.\n",
"[2213.84s --> 2220.72s] : How bad is the completely unrestricted model?\n",
"[2220.72s --> 2224.40s] : So how much do you understand about that?\n",
"[2224.40s --> 2224.88s] : You know,\n",
"[2224.88s --> 2228.48s] : there there's a's been a lot of discussion about free speech absolutism.\n",
"[2228.48s --> 2228.80s] : ,\n",
"[2228.80s --> 2231.92s] : how much if that's applied to an AI system.\n",
"[2231.92s --> 2232.16s] : You know,\n",
"[2232.16s --> 2234.56s] : we've talked about putting out the base model,\n",
"[2234.56s --> 2236.08s] : at least for researchers or something,\n",
"[2236.08s --> 2237.28s] : but it's not very easy to use.\n",
"[2237.28s --> 2238.08s] : Everyone's like,\n",
"[2238.08s --> 2239.12s] : Give me the base model,\n",
"[2239.12s --> 2240.72s] : and again we might do that.\n",
"[2240.72s --> 2244.80s] : I think what people mostly want is they want a model that has been RLHF.\n",
"[2244.80s --> 2247.36s] : To the world do they subscribe to.\n",
"[2247.36s --> 2249.60s] : It's really about regulating other people's speech.\n",
"[2249.60s --> 2249.84s] : Yeah,\n",
"[2249.84s --> 2251.84s] : like people I implied you know,\n",
"[2251.84s --> 2254.80s] : like in the debates about what shut up in the Facebook feed,\n",
"[2254.80s --> 2258.08s] : I I having listened to a lot of people talk about that,\n",
"[2258.08s --> 2259.12s] : everyone is like,\n",
"[2259.12s --> 2259.44s] : Well,\n",
"[2259.44s --> 2262.56s] : it doesn't matter what's in my feet because I won't be radicalized,\n",
"[2262.56s --> 2263.76s] : I can handle anything,\n",
"[2263.76s --> 2266.48s] : but I really worry about what Facebook shows you.\n",
"[2266.48s --> 2277.44s] : I would love it if there is some way in which I think my interaction with GpT has already done that some way to in a nuanced way present the tension of ideas.\n",
"[2277.44s --> 2280.48s] : I think we are doing better at that than people realize.\n",
"[2280.48s --> 2281.20s] : The challenge,\n",
"[2281.20s --> 2281.44s] : of course,\n",
"[2281.44s --> 2290.72s] : when you are evaluating this stuff is you can always find anecdotal evidence of GPT slipping up and saying something either wrong or,\n",
"[2290.72s --> 2291.12s] : um,\n",
"[2291.12s --> 2292.96s] : biased and so on,\n",
"[2292.96s --> 2299.20s] : but it would be nice to be able to kind of generally make statements about the bias of the system,\n",
"[2299.20s --> 2302.16s] : generally make statements about there people doing good work there.\n",
"[2302.16s --> 2302.56s] : You know,\n",
"[2302.56s --> 2305.92s] : if you ask the same question ten thousand times,\n",
"[2305.92s --> 2309.36s] : and you rank the outputs from best to worse.\n",
"[2309.36s --> 2310.96s] : What most people see is,\n",
"[2310.96s --> 2311.28s] : of course,\n",
"[2311.28s --> 2313.28s] : something around output five thousand,\n",
"[2313.28s --> 2318.88s] : but the output that gets all of the Twitter attention is output ten thousand.\n",
"[2318.88s --> 2326.08s] : And this is something that I think the world will just have to adapt to with these models is that,\n",
"[2326.08s --> 2330.72s] : sometimes there's a really egregiously dumb answer,\n",
"[2330.72s --> 2334.72s] : and in a world where you click screenshot and share,\n",
"[2334.72s --> 2336.80s] : that might not be representative.\n",
"[2336.80s --> 2340.32s] : Now already we're noticing a lot more people respond to those things saying,\n",
"[2340.32s --> 2340.48s] : Well,\n",
"[2340.48s --> 2344.56s] : I tried it and got this And so I think we are building up the antibodies there,\n",
"[2344.56s --> 2345.92s] : But it's a new thing.\n",
"[2345.92s --> 2357.84s] : Do you feel pressure from click Bay journalism that looks at ten thousand that looks at the worst possible output of GPT.\n",
"[2357.84s --> 2361.52s] : Do you feel a pressure to not be transparent because of that?\n",
"[2361.52s --> 2365.12s] : Because you're sort of making mistakes in public,\n",
"[2365.12s --> 2367.60s] : and you're burned for the mistakes.\n",
"[2367.60s --> 2373.84s] : Is there a pressure culturally within open AI that you're afraid knew like it might close you up?\n",
"[2373.84s --> 2374.08s] : I mean,\n",
"[2374.08s --> 2376.48s] : evidently there doesn't seem to be we keep doing our thing,\n",
"[2376.48s --> 2376.88s] : you know.\n",
"[2376.88s --> 2378.16s] : So you don't feel that?\n",
"[2378.16s --> 2378.56s] : I mean,\n",
"[2378.56s --> 2379.60s] : there is a pressure,\n",
"[2379.60s --> 2381.60s] : but ye it doesn't affect you.\n",
"[2381.60s --> 2385.20s] : I'm sure it has all sorts of subtle effects.\n",
"[2385.20s --> 2387.20s] : I don't fully understand.\n",
"[2387.20s --> 2389.76s] : but I don't perceive much of that.\n",
"[2389.76s --> 2390.24s] : I mean,\n",
"[2390.24s --> 2392.88s] : we're we're happy to admit when we're wrong.\n",
"[2392.88s --> 2395.04s] : We want to get better and better.\n",
"[2395.04s --> 2396.08s] : Um,\n",
"[2396.08s --> 2402.48s] : I think we're pretty good about trying to listen to every piece of criticism,\n",
"[2402.48s --> 2403.60s] : think it through,\n",
"[2403.60s --> 2405.36s] : internalize what we agree with,\n",
"[2405.36s --> 2409.76s] : but like the breathless clickbate headlines You know,\n",
"[2409.76s --> 2411.92s] : try to let those flow through us.\n",
"[2411.92s --> 2412.56s] : Uh,\n",
"[2412.56s --> 2416.24s] : what does the openp AIy moderation tooling for GPT look like?\n",
"[2416.24s --> 2417.92s] : What's the process of moderation?\n",
"[2417.92s --> 2418.64s] : So there's,\n",
"[2418.64s --> 2418.96s] : uh,\n",
"[2418.96s --> 2422.40s] : several things.be maybe it's the same thing you can educate me.\n",
"[2422.40s --> 2425.44s] : Soir RLHF is the ranking.\n",
"[2425.44s --> 2428.24s] : but is there a wall you're up against,\n",
"[2428.24s --> 2430.24s] : like uh,\n",
"[2430.24s --> 2433.44s] : where this is an unsafe thing to answer?\n",
"[2433.44s --> 2435.28s] : What does that tooling look like?\n",
"[2435.28s --> 2438.56s] : We do have systems that try to figure out you know,\n",
"[2438.56s --> 2442.24s] : try to learn when a question is something that we're supposed to we call refusals,\n",
"[2442.24s --> 2443.60s] : refuse to answer.\n",
"[2443.60s --> 2446.48s] : It is early and imperfect.\n",
"[2446.48s --> 2447.20s] : Ah,\n",
"[2447.20s --> 2447.76s] : we're,\n",
"[2447.76s --> 2448.32s] : again,\n",
"[2448.32s --> 2453.84s] : the spirit of building in public and and bringing society along gradually.\n",
"[2453.84s --> 2455.60s] : We put something out.\n",
"[2455.60s --> 2456.88s] : It's got flaws.\n",
"[2456.88s --> 2459.12s] : we'll make better versions.m,\n",
"[2459.12s --> 2459.76s] : But yes,\n",
"[2459.76s --> 2464.08s] : we are trying the system is trying to learn questions that it shouldn't answer.\n",
"[2464.08s --> 2473.04s] : One small thing that really bothers me about our current thing and will get this better is I don't like the feeling of being scolded by a computer.\n",
"[2473.04s --> 2474.08s] : ?\n",
"[2474.08s --> 2475.68s] : I really don't,\n",
"[2475.68s --> 2476.16s] : you know.\n",
"[2476.16s --> 2479.12s] : I a story that has always stuck with me I don't know if it's true,\n",
"[2479.12s --> 2484.96s] : I hope it is is that the reason Steve Jobs put that handle on the back of the first iMc.\n",
"[2484.96s --> 2486.16s] : Remember that big plastic,\n",
"[2486.16s --> 2487.36s] : fright colored thing?\n",
"[2487.36s --> 2491.20s] : was that you should never trust a computer you shouldn' throw out you couldn't throw out a window?\n",
"[2491.20s --> 2493.84s] : Nice And Of course,\n",
"[2493.84s --> 2496.56s] : not that many people actually throw their computer out a window.\n",
"[2496.56s --> 2498.56s] : But it's sort of nice to know that you can.\n",
"[2498.56s --> 2502.72s] : And it's nice to know that like this is a tool very much in my control,\n",
"[2502.72s --> 2506.00s] : and this is a tool that like does things to help me.\n",
"[2506.00s --> 2510.40s] : And I think we've done a pretty good job of that with GPT fourour,\n",
"[2510.40s --> 2516.48s] : but I notice that I have like a this orle response to being scolded by a computer,\n",
"[2516.48s --> 2518.16s] : and I think you know,\n",
"[2518.16s --> 2521.36s] : that's a good learning from deploying or from creating a system,\n",
"[2521.36s --> 2523.20s] : and we can improve it.\n",
"[2523.20s --> 2524.16s] : Yes,\n",
"[2524.16s --> 2524.80s] : tricky,\n",
"[2524.80s --> 2527.36s] : and also for the system not to treat you like a child.\n",
"[2527.36s --> 2529.60s] : Treing our users like adults is a thing,\n",
"[2529.60s --> 2530.00s] : I say,\n",
"[2530.00s --> 2532.32s] : very frequentlyside inside the office.\n",
"[2532.32s --> 2534.40s] : But it's tricky it has to do with language,\n",
"[2534.40s --> 2541.04s] : like if there's like certain conspiracy theories you don't want the system to be speaking to,\n",
"[2541.04s --> 2543.76s] : it's a very tricky language you should use,\n",
"[2543.76s --> 2556.56s] : because what if I want to understand The earth if the earth is the idea that the earth is flat and I want to fully explore that I want the I want GBD to help me explore.\n",
"[2556.56s --> 2563.68s] : GBD four has enough nuance to be able to help you explore that without And treat you like an adult in the process.\n",
"[2563.68s --> 2564.56s] : GBT three,\n",
"[2564.56s --> 2564.80s] : I think,\n",
"[2564.80s --> 2566.64s] : just wasn't capable of getting that right.\n",
"[2566.64s --> 2568.88s] : But GPT four I think we can get to do this.\n",
"[2568.88s --> 2569.36s] : By the way,\n",
"[2569.36s --> 2575.28s] : if you could just speak to the leap from GPty four two GPT four from three point five from three.\n",
"[2575.28s --> 2576.80s] : Is there some technical leaps,\n",
"[2576.80s --> 2579.20s] : or is it really focused on the alignment?\n",
"[2579.20s --> 2579.68s] : No,\n",
"[2579.68s --> 2581.84s] : it's a lot of technical leaps in the base model.\n",
"[2581.84s --> 2589.44s] : One of the things we are good at at Open Eye is finding a lot of small winds and multiplying them together.\n",
"[2589.44s --> 2594.72s] : And each of them maybe is like a pretty big secret in some sense,\n",
"[2594.72s --> 2604.56s] : but it really is the multiplicative impact of all of them and the detail and care we put into it that gets us these big leaps and then,\n",
"[2604.56s --> 2604.88s] : you know,\n",
"[2604.88s --> 2606.32s] : it looks like to the outside,\n",
"[2606.32s --> 2606.56s] : like,\n",
"[2606.56s --> 2606.88s] : oh,\n",
"[2606.88s --> 2610.48s] : they just probably like did one thing to get from three to three point five to four.\n",
"[2610.48s --> 2612.96s] : It's like hundreds of complicated things.\n",
"[2612.96s --> 2614.88s] : It's the tiny little thing with the training,\n",
"[2614.88s --> 2615.84s] : with the everything,\n",
"[2615.84s --> 2617.36s] : with the data how we like,\n",
"[2617.36s --> 2618.16s] : collect the data,\n",
"[2618.16s --> 2619.04s] : how we clean the data,\n",
"[2619.04s --> 2619.84s] : how we do the training,\n",
"[2619.84s --> 2620.96s] : how we do the optimizer,\n",
"[2620.96s --> 2623.36s] : how we do the architecture like so many things.\n",
"[2623.36s --> 2627.36s] : Let me ask you the all important question about size.\n",
"[2627.36s --> 2631.76s] : So does size matter in terms of neural networks,\n",
"[2631.76s --> 2632.08s] : ,\n",
"[2632.08s --> 2635.36s] : with how good the system performs?\n",
"[2635.36s --> 2639.60s] : So GPT three three point five at a one hundred seventy five billion.\n",
"[2639.60s --> 2641.36s] : I heard GP two four a hundred trillion.\n",
"[2641.36s --> 2642.16s] : one hundred trillion.\n",
"[2642.16s --> 2643.36s] : Canight speak to this?\n",
"[2643.36s --> 2644.72s] : Do you know that mean?\n",
"[2644.72s --> 2645.04s] : Yeah,\n",
"[2645.04s --> 2646.88s] : the big and purple socks do you know where it origin?\n",
"[2646.88s --> 2647.44s] : I don't do you?\n",
"[2647.44s --> 2649.60s] : I'd be curious to hear the presentation I gave.\n",
"[2649.60s --> 2651.04s] : No way.\n",
"[2651.04s --> 2655.44s] : Ah journalist's just took a snap shot.\n",
"[2655.44s --> 2658.08s] : Huh Now I learned from this.\n",
"[2658.08s --> 2662.40s] : It's right when GPT three was released I gave a sign you tube,\n",
"[2662.40s --> 2664.48s] : but I gave a description of what it is.\n",
"[2664.48s --> 2669.60s] : And I spoke to the limitation of the parameters and like where it's going,\n",
"[2669.60s --> 2671.68s] : and I talked about the human brain,\n",
"[2671.68s --> 2674.88s] : and how many parameters it has on apsis and so on.\n",
"[2674.88s --> 2675.60s] : And,\n",
"[2675.60s --> 2676.08s] : um,\n",
"[2676.08s --> 2677.60s] : perhaps like an idiot,\n",
"[2677.60s --> 2678.40s] : perhaps not.\n",
"[2678.40s --> 2680.32s] : I said like GPT four,\n",
"[2680.32s --> 2680.96s] : like the next,\n",
"[2680.96s --> 2681.84s] : as it progresses.\n",
"[2681.84s --> 2684.24s] : What I should have said is GPT N or something.\n",
"[2684.24s --> 2685.92s] : I can't believe that it this came from you,\n",
"[2685.92s --> 2686.48s] : that is.\n",
"[2686.48s --> 2688.40s] : But people should go to it.\n",
"[2688.40s --> 2690.48s] : It's totally taken out of context,\n",
"[2690.48s --> 2691.76s] : they didn't reference anything,\n",
"[2691.76s --> 2692.24s] : they took it,\n",
"[2692.24s --> 2694.48s] : this is what GPT four is going to be.\n",
"[2694.48s --> 2697.28s] : And I feel horrible about it.\n",
"[2697.28s --> 2698.00s] : You know,\n",
"[2698.00s --> 2700.88s] : it doesn't I don't think it matters in any serious way.\n",
"[2700.88s --> 2701.20s] : I mean,\n",
"[2701.20s --> 2702.08s] : it's not good because,\n",
"[2702.08s --> 2702.32s] : uh,\n",
"[2702.32s --> 2702.72s] : again,\n",
"[2702.72s --> 2703.68s] : size is not everything,\n",
"[2703.68s --> 2708.48s] : but also people just take a a lot of these kinds of discussions out of context.\n",
"[2708.48s --> 2709.28s] : Uh,\n",
"[2709.28s --> 2711.36s] : but it is interesting to comp I mean,\n",
"[2711.36s --> 2716.32s] : that's what I was trying to do to com to compare in different ways uh,\n",
"[2716.32s --> 2721.20s] : the difference between the human brain and your network and this thing is getting so impressive.\n",
"[2721.20s --> 2725.84s] : This is like in some sense someone said to me this morning actually,\n",
"[2725.84s --> 2726.40s] : and I was like,\n",
"[2726.40s --> 2726.80s] : Ohh,\n",
"[2726.80s --> 2731.52s] : this might be right this is the most complex software object humanity has yet produced.\n",
"[2731.52s --> 2734.80s] : And it will be trivial in a couple of decades.'s right,\n",
"[2734.80s --> 2737.20s] : it'll be like kind of anyone can do it whatever.\n",
"[2737.20s --> 2739.36s] : Um but yeah,\n",
"[2739.36s --> 2747.12s] : the amount of complexity relative to anything we've done so far that goes into producing this one set of numbers is quite something.\n",
"[2747.12s --> 2748.00s] : ,\n",
"[2748.00s --> 2751.60s] : complexity including the entirety the history of human civilization,\n",
"[2751.60s --> 2754.48s] : that build up all the different advancements of technology,\n",
"[2754.48s --> 2756.16s] : that build up all the content,\n",
"[2756.16s --> 2760.48s] : the data that Jupitio was trained on that is on the Internet,\n",
"[2760.48s --> 2764.08s] : that it's the compression of all of humanity.\n",
"[2764.08s --> 2766.72s] : of all the maybe not the experience,\n",
"[2766.72s --> 2768.64s] : all the text output that humanity produces?\n",
"[2768.64s --> 2768.96s] : Yes,\n",
"[2768.96s --> 2769.68s] : it' somewhat different.\n",
"[2769.68s --> 2770.96s] : And it's a good question.\n",
"[2770.96s --> 2771.92s] : How much,\n",
"[2771.92s --> 2774.56s] : if all you have is the Internet data,\n",
"[2774.56s --> 2778.56s] : howw much can you reconstruct the magic of what it means to be human?\n",
"[2778.56s --> 2781.92s] : I think you' be surprised how much you can reconstruct.\n",
"[2781.92s --> 2786.96s] : But you probably need more a better and better and better models.\n",
"[2786.96s --> 2789.28s] : But on that topic how much does size matter?\n",
"[2789.28s --> 2790.88s] : By like number of parameters,\n",
"[2790.88s --> 2792.08s] : number of parameters.\n",
"[2792.08s --> 2799.20s] : I think people got caught up in the parameter count race in the same way they got caught up in the gigahertzs race of processors in like the,\n",
"[2799.20s --> 2799.60s] : you know,\n",
"[2799.60s --> 2801.84s] : nineties and two thousands or whatever.\n",
"[2801.84s --> 2802.80s] : You,\n",
"[2802.80s --> 2803.20s] : I think,\n",
"[2803.20s --> 2806.72s] : probably have no idea how many gigaherts the processor in your phone is.\n",
"[2806.72s --> 2810.32s] : But what you care about is what the thing can do for you,\n",
"[2810.32s --> 2810.80s] : and there's,\n",
"[2810.80s --> 2811.04s] : you know,\n",
"[2811.04s --> 2812.32s] : different ways to accomplish that.\n",
"[2812.32s --> 2814.32s] : You can bump up the clock speed.\n",
"[2814.32s --> 2815.76s] : sometimesmes that causes other problems,\n",
"[2815.76s --> 2817.92s] : sometimes it's not the best way to get gains.\n",
"[2817.92s --> 2823.36s] : Um but I think what matters is getting the best performance,\n",
"[2823.36s --> 2825.92s] : and you know,\n",
"[2825.92s --> 2837.76s] : we I think one thing that works well about open eye is we're pretty truth seeking in just doing whatever is going to make the best performance,\n",
"[2837.76s --> 2840.24s] : whether or not it's the most elegant solution.\n",
"[2840.24s --> 2845.92s] : So I think like Ella Lemser a sort of hated result in parts of the field.\n",
"[2845.92s --> 2850.72s] : Everybody wanted to come up with a more elegant way to get to generalized intelligence,\n",
"[2850.72s --> 2855.92s] : And we have been willing to just keep doing what works and looks like it'll keep working.\n",
"[2855.92s --> 2859.84s] : So I've spoken with No Chomsky,\n",
"[2859.84s --> 2867.28s] : who's been kind of one of the many people that are critical of large language models being able to achieve general intelligence,\n",
"[2867.28s --> 2867.60s] : right?\n",
"[2867.60s --> 2872.00s] : And so it's an interesting question that they've been able to achieve so much incredible stuff.\n",
"[2872.00s --> 2878.72s] : Do you think it's possible that large language models really is the way we build AGI?\n",
"[2878.72s --> 2880.64s] : I think it's part of the way.\n",
"[2880.64s --> 2883.28s] : I think we need other super important things.\n",
"[2883.28s --> 2885.68s] : This is philosophizing a little bit.\n",
"[2885.68s --> 2888.08s] : Like what what kind of components do you think?\n",
"[2888.08s --> 2892.32s] : Uh in a technical sense or a poetic sense.\n",
"[2892.32s --> 2897.04s] : Does it need to have a body that it can experience the world directly?\n",
"[2897.04s --> 2899.84s] : I don't think it needs that.\n",
"[2899.84s --> 2903.20s] : But I wouldn't In't say any of this stuff with certainty,\n",
"[2903.20s --> 2904.88s] : like we're deep into the unknown here.\n",
"[2904.88s --> 2906.08s] : For me,\n",
"[2906.08s --> 2914.40s] : a system that cannot go significantly add to the sum total of scientific knowledge we have access to,\n",
"[2914.40s --> 2915.68s] : kind of discover,\n",
"[2915.68s --> 2917.36s] : invent whatever you want to call it,\n",
"[2917.36s --> 2919.28s] : new fundamental science,\n",
"[2919.28s --> 2923.20s] : is not a superintelligence.\n",
"[2923.20s --> 2928.32s] : And to do that really well,\n",
"[2928.32s --> 2931.52s] : I think we will need to expand on the GPT paradigm,\n",
"[2931.52s --> 2934.64s] : in pretty important ways that we're still missing ideas for.\n",
"[2934.64s --> 2938.32s] : But I don't know what those ideas are we're trying to find them.\n",
"[2938.32s --> 2941.84s] : I could argue so to the opposite point that you could have deep,\n",
"[2941.84s --> 2946.24s] : big scientific breakthroughs withggest the data that GPTs trained on.\n",
"[2946.24s --> 2951.12s] : Li I make some ofvi like if you prompt it correctly.\n",
"[2951.12s --> 2951.60s] : Look,\n",
"[2951.60s --> 2952.88s] : if an oracle told me,\n",
"[2952.88s --> 2953.68s] : far from the future,\n",
"[2953.68s --> 2956.80s] : that GPT ten turned out to be a true AGI somehow,\n",
"[2956.80s --> 2959.28s] : or maybe just some very small new ideas,\n",
"[2959.28s --> 2960.24s] : I would be like,\n",
"[2960.24s --> 2962.24s] : OKy I can't believe that.\n",
"[2962.24s --> 2963.76s] : Not what I would have expected,\n",
"[2963.76s --> 2964.24s] : sitting here,\n",
"[2964.24s --> 2965.52s] : would have set a new big idea,\n",
"[2965.52s --> 2966.80s] : but I can believe that.\n",
"[2966.80s --> 2972.96s] : This prompting chain If you extend it very far,\n",
"[2972.96s --> 2977.68s] : and then increase at scale the number of those interactions,\n",
"[2977.68s --> 2985.20s] : like what kind of these things start getting integrated into human society and starts building on top of each other.\n",
"[2985.20s --> 2985.52s] : I mean,\n",
"[2985.52s --> 2987.60s] : like I don't think we understand what that looks like.\n",
"[2987.60s --> 2988.24s] : likeke you said,\n",
"[2988.24s --> 2989.12s] : it's been six days.\n",
"[2989.12s --> 2995.04s] : The thing that I am so excited about with this is not that it's a system that kind of goes off and does its own thing,\n",
"[2995.04s --> 3000.16s] : but that it's this tool that humans are using in this feedback loop.\n",
"[3000.16s --> 3006.80s] : Helpful for us for a bunch of reasons we get to learn more about trajectories through multiple iterations.\n",
"[3006.80s --> 3016.32s] : but I am excited about a world where AI is an extension of human will and a amplifier of our abilities,\n",
"[3016.32s --> 3017.36s] : and this like,\n",
"[3017.36s --> 3017.92s] : you know,\n",
"[3017.92s --> 3020.32s] : most useful tool yet created.\n",
"[3020.32s --> 3022.72s] : And that is certainly how people are using it.\n",
"[3022.72s --> 3023.68s] : And I mean,\n",
"[3023.68s --> 3025.04s] : just like look at Twitter,\n",
"[3025.04s --> 3030.48s] : like the the results are amazing peoples like self reported happiness with getting to work with this are great.\n",
"[3030.48s --> 3032.40s] : So yeah,\n",
"[3032.40s --> 3034.96s] : likeke maybe we never build AGI,\n",
"[3034.96s --> 3037.04s] : but we just make humans super great.\n",
"[3037.04s --> 3038.80s] : stilltill a huge one?\n",
"[3038.80s --> 3039.68s] : Yeah,\n",
"[3039.68s --> 3053.60s] : I said I I'm part of those people like the amount I derive a lot of happiness from programming together with GPT.h Part of it is a little bit of terror of can you say more about that?\n",
"[3053.60s --> 3055.60s] : There's a meme.\n",
"[3055.60s --> 3061.04s] : I saw today that everybody's freaking out about a sort of GPT taking programmer jobs.\n",
"[3061.04s --> 3061.44s] : No,\n",
"[3061.44s --> 3068.72s] : it's the reality' it' going to be taking like if it's going to take your job it means you're a shdy programmer.\n",
"[3068.72s --> 3070.80s] : There's some truth to that.\n",
"[3070.80s --> 3076.80s] : Maybe there's some human element that's really fundamental to the creative act,\n",
"[3076.80s --> 3081.44s] : to the active genius that is in great design that is involved in programming,\n",
"[3081.44s --> 3085.76s] : and maybe I'm just really impressed by all the boiler plate,\n",
"[3085.76s --> 3090.32s] : but that I don't see as boiler plate but is actually a pretty boiler plate.\n",
"[3090.32s --> 3090.72s] : Yeah,\n",
"[3090.72s --> 3092.32s] : and maybe that you create like,\n",
"[3092.32s --> 3092.64s] : you know,\n",
"[3092.64s --> 3095.28s] : in a day of programming you have one really important idea.\n",
"[3095.28s --> 3095.92s] : Yeah.\n",
"[3095.92s --> 3098.96s] : And that's the contchet whichine and that's my contribution.\n",
"[3098.96s --> 3100.08s] : And there may be like,\n",
"[3100.08s --> 3105.28s] : I think we're going to find So I suspect that is happening with great programmers,\n",
"[3105.28s --> 3108.08s] : and that GPT like models are far away from that one thing,\n",
"[3108.08s --> 3110.80s] : even though they're going to automate a lot of other programming.\n",
"[3110.80s --> 3111.52s] : But,\n",
"[3111.52s --> 3112.00s] : again,\n",
"[3112.00s --> 3116.96s] : most programmers have some sense of you know,\n",
"[3116.96s --> 3119.60s] : anxiety about what the future' is going to look like,\n",
"[3119.60s --> 3120.48s] : but mostly they arere like.\n",
"[3120.48s --> 3121.20s] : This is amazing.\n",
"[3121.20s --> 3122.72s] : I am ten times more productive.\n",
"[3122.72s --> 3124.24s] : and don't ever take this away from me.\n",
"[3124.24s --> 3127.12s] : There's not a lot of people that use it and say like,Turn this off,\n",
"[3127.12s --> 3127.60s] : you know?\n",
"[3127.60s --> 3128.16s] : Yeah.\n",
"[3128.16s --> 3129.04s] : So I I think,\n",
"[3129.04s --> 3129.28s] : uh,\n",
"[3129.28s --> 3129.92s] : so to speak,\n",
"[3129.92s --> 3134.72s] : this the psychology of terror is more like This is awesome This is too awesome,\n",
"[3134.72s --> 3135.60s] : It's too awesome.\n",
"[3135.60s --> 3142.48s] : Yeah There is a little bit coffee tastes too good You know when Kasparov lost the deep bluof,\n",
"[3142.48s --> 3149.52s] : somebody said and maybe it was him that like chess is over now if an AI can beat a human at chess,\n",
"[3149.52s --> 3152.08s] : then no one's going to bother to keep playing,\n",
"[3152.08s --> 3152.32s] : right,\n",
"[3152.32s --> 3154.48s] : 'cause like what's the purpose of us' or whatever?\n",
"[3154.48s --> 3156.48s] : That was thirty years ago,\n",
"[3156.48s --> 3157.44s] : twenty five years ago,\n",
"[3157.44s --> 3158.40s] : something like that.\n",
"[3158.40s --> 3162.88s] : I believe that chess has never been more popular than it is right now.\n",
"[3162.88s --> 3166.88s] : And people keep wanting to play,\n",
"[3166.88s --> 3167.84s] : and wanting to watch.\n",
"[3167.84s --> 3168.56s] : And by the way,\n",
"[3168.56s --> 3170.64s] : we don't watch two e eyes play each other,\n",
"[3170.64s --> 3173.44s] : which would be a far better game,\n",
"[3173.44s --> 3174.32s] : in some sense,\n",
"[3174.32s --> 3176.08s] : than whatever else.\n",
"[3176.08s --> 3181.68s] : But that's That's not what we choose to do.\n",
"[3181.68s --> 3185.60s] : like we are somehow much more interested in what humans do in this sense,\n",
"[3185.60s --> 3189.20s] : and whether or not Magnus loses to that kid,\n",
"[3189.20s --> 3191.44s] : than what happens when two much,\n",
"[3191.44s --> 3193.12s] : much better AI s play each other.\n",
"[3193.12s --> 3193.76s] : Well,\n",
"[3193.76s --> 3194.48s] : actually,\n",
"[3194.48s --> 3198.24s] : when two AIs play each other it's not a better game by our definition of best.\n",
"[3198.24s --> 3199.28s] : We just can't understand it.\n",
"[3199.28s --> 3199.68s] : No,\n",
"[3199.68s --> 3201.84s] : I think I think they just draw each other.\n",
"[3201.84s --> 3204.24s] : I think the human flaws,\n",
"[3204.24s --> 3209.44s] : and this might apply across the spectrum here with A's will make life way better,\n",
"[3209.44s --> 3211.52s] : but we'll still want drama.\n",
"[3211.52s --> 3214.64s] : we still want imperfection and flaws,\n",
"[3214.64s --> 3216.72s] : and AI will not have us much of that.\n",
"[3216.72s --> 3217.12s] : Look,\n",
"[3217.12s --> 3217.36s] : I mean,\n",
"[3217.36s --> 3219.76s] : I hate to sound like Uutopic techabro here,\n",
"[3219.76s --> 3221.76s] : but if you'll excuse me for three seconds,\n",
"[3221.76s --> 3230.80s] : like the the the level of the increase in quality of life that AI can deliver is extraordinary.\n",
"[3230.80s --> 3233.92s] : We can make the world amazing,\n",
"[3233.92s --> 3235.84s] : we can make people's lives amazing,\n",
"[3235.84s --> 3236.88s] : we can cure diseases,\n",
"[3236.88s --> 3238.40s] : we can increase material wealth,\n",
"[3238.40s --> 3240.16s] : we can like help people be happier,\n",
"[3240.16s --> 3242.40s] : more fulfill all these sorts of things.\n",
"[3242.40s --> 3244.72s] : And then people are like,\n",
"[3244.72s --> 3244.96s] : Ohh,\n",
"[3244.96s --> 3245.12s] : well,\n",
"[3245.12s --> 3249.68s] : no one is going to work but peopleeople want statatus.\n",
"[3249.68s --> 3250.72s] : people want drama,\n",
"[3250.72s --> 3251.76s] : people want new things,\n",
"[3251.76s --> 3252.72s] : people want to create,\n",
"[3252.72s --> 3253.68s] : people want to like,\n",
"[3253.68s --> 3254.80s] : feel useful.\n",
"[3254.80s --> 3255.60s] : Um,\n",
"[3255.60s --> 3257.60s] : people want to do all these things,\n",
"[3257.60s --> 3262.32s] : and we're just going to find new and different ways to do them even in a vastly better,\n",
"[3262.32s --> 3265.04s] : like unimaginably good standard of living world.\n",
"[3265.04s --> 3267.92s] : But that world,\n",
"[3267.92s --> 3270.08s] : the positive trajectories with AI,\n",
"[3270.08s --> 3272.88s] : that world is with an AI that's aligned with humans,\n",
"[3272.88s --> 3273.84s] : It doesn't hurt,\n",
"[3273.84s --> 3274.56s] : doesn't limit,\n",
"[3274.56s --> 3277.76s] : doesn't doesn't try to get rid of humans.\n",
"[3277.76s --> 3284.00s] : And there's some folks who consider all the different problems with the superintelligent AI system,\n",
"[3284.00s --> 3287.68s] : so one of them is ElelizaerYakowvsky.\n",
"[3287.68s --> 3292.24s] : He warns that Ahi will likely kill all humans.\n",
"[3292.24s --> 3294.48s] : and there's a bunch of different cases,\n",
"[3294.48s --> 3304.88s] : but I think one way to summarize it is that it's almost impossible to keep AI aligned as it becomes superintelligent.\n",
"[3304.88s --> 3307.28s] : Can you steel man the case for that?\n",
"[3307.28s --> 3312.96s] : and to what degree do you disagree with that trajectory?\n",
"[3312.96s --> 3314.08s] : So,\n",
"[3314.08s --> 3314.72s] : first of all,\n",
"[3314.72s --> 3318.48s] : I'll say I think that there is some chance of that,\n",
"[3318.48s --> 3320.48s] : and it's really important to acknowledge it,\n",
"[3320.48s --> 3321.52s] : because if we don't talk about it,\n",
"[3321.52s --> 3323.36s] : if we don't treat it as potentially real,\n",
"[3323.36s --> 3325.76s] : we won't put enough effort into solving it.\n",
"[3325.76s --> 3331.20s] : And I think we do have to discover new techniques to be able to solve it.\n",
"[3331.20s --> 3332.16s] : Um,\n",
"[3332.16s --> 3335.68s] : I think a lot of the predictions this is true for any new field,\n",
"[3335.68s --> 3337.84s] : but a lot of the predictions about AI,\n",
"[3337.84s --> 3339.60s] : in terms of capabilities,\n",
"[3339.60s --> 3347.28s] : in terms of what the safety challenges and the easy parts are going to be have turned out to be wrong,\n",
"[3347.28s --> 3355.04s] : the only way I know how to solve a problem like this is iterating our way through it,\n",
"[3355.04s --> 3357.92s] : learning early,\n",
"[3357.92s --> 3362.80s] : and limiting the number of one shot to get it right scenarios that we have.\n",
"[3362.80s --> 3365.68s] : To Stealman Well,\n",
"[3365.68s --> 3369.52s] : there's I can't just pick like one AI safety case or AI alignment case,\n",
"[3369.52s --> 3374.48s] : but I think Elizer wrote a really great blog post.\n",
"[3374.48s --> 3382.16s] : I think some of his work has been sort of somewhat difficult to follow or had what I view like quite significant logical flaws,\n",
"[3382.16s --> 3390.96s] : but he wrote this one blog post outlining why he believed that alignment was such a hard problem that I thought was againga,\n",
"[3390.96s --> 3392.08s] : don't agree with a lot of it,\n",
"[3392.08s --> 3394.96s] : but well reasoned and thoughtful and and very worth reading.\n",
"[3394.96s --> 3396.56s] : So I think I'd point people to that,\n",
"[3396.56s --> 3397.60s] : is the Steel man?\n",
"[3397.60s --> 3398.16s] : Yes,\n",
"[3398.16s --> 3400.64s] : and I'll also have a conversation with him.\n",
"[3400.64s --> 3411.04s] : Um There is some aspect and I'm torn here because it's difficult to reason about the exponential improvement of technology.\n",
"[3411.04s --> 3411.44s] : .\n",
"[3411.44s --> 3422.88s] : But also I've seen time and time again how transparent and iterative trying out as you improve the technology,\n",
"[3422.88s --> 3423.52s] : trying it out,\n",
"[3423.52s --> 3424.32s] : releasing it,\n",
"[3424.32s --> 3425.28s] : testing it,\n",
"[3425.28s --> 3426.72s] : how that can,\n",
"[3426.72s --> 3427.36s] : ah,\n",
"[3427.36s --> 3431.20s] : improve your understanding of the technology,\n",
"[3431.20s --> 3434.40s] : in such that the philosophy of how to do,\n",
"[3434.40s --> 3434.88s] : for example,\n",
"[3434.88s --> 3440.48s] : safety of any kind of technology but AI safety gets adjusted over time rapidly.\n",
"[3440.48s --> 3446.08s] : A lot of the formative AI safety work was done before people even believed in deep learning,\n",
"[3446.08s --> 3449.68s] : and and certainly before people believed in large language models.\n",
"[3449.68s --> 3453.92s] : And I don't think it's like updated enough given everything we've learned now,\n",
"[3453.92s --> 3455.76s] : and everything we will learn going forward.\n",
"[3455.76s --> 3459.44s] : so I think it's got to be this very tight feedback loop.\n",
"[3459.44s --> 3461.20s] : I think theory does play a real role,\n",
"[3461.20s --> 3461.84s] : of course.\n",
"[3461.84s --> 3466.56s] : but continuing to learn what we learn from how the technology trajectory goes?\n",
"[3466.56s --> 3469.44s] : I quite important.\n",
"[3469.44s --> 3473.44s] : I think now is a very good time and we're trying to figure out how to do this,\n",
"[3473.44s --> 3477.20s] : to significantly ramp up technical alignment work.\n",
"[3477.20s --> 3478.48s] : I think we have new tools,\n",
"[3478.48s --> 3480.00s] : we have new understanding.\n",
"[3480.00s --> 3480.80s] : ah,\n",
"[3480.80s --> 3486.00s] : and there's a lot of work that's important to do That we can do now.\n",
"[3486.00s --> 3488.00s] : So one of the main concerns here is,\n",
"[3488.00s --> 3488.32s] : o,\n",
"[3488.32s --> 3492.16s] : something called AI take off or fast take off,\n",
"[3492.16s --> 3497.68s] : that the exponential improvement would be really fast to wear like in days?\n",
"[3497.68s --> 3498.40s] : In days,\n",
"[3498.40s --> 3498.96s] : yeah.\n",
"[3498.96s --> 3508.40s] : mean mean there's this is an this is a pretty serious at least to me it's become more of a serious concern.\n",
"[3508.40s --> 3514.08s] : Just how amazing Chad GPT turned out to be and then the improvement of GPT fourour.\n",
"[3514.08s --> 3516.48s] : Almost like to where it surprised everyone.\n",
"[3516.48s --> 3518.32s] : seeemingly you can correct me,\n",
"[3518.32s --> 3519.28s] : including you.\n",
"[3519.28s --> 3522.48s] : So GPT fourour has not surprised me at all in terms of reception there.\n",
"[3522.48s --> 3524.88s] : Chat CipPT surprised us a little bit,\n",
"[3524.88s --> 3528.40s] : but I still was like advocating that we' do it 'cause I thought it was going to do really great.\n",
"[3528.40s --> 3530.48s] : U so like you know,\n",
"[3530.48s --> 3541.36s] : maybe I thought it would have been like The tenth fastest growing product in history and not the number one fastest I like ka.\n",
"[3541.36s --> 3541.76s] : You know,\n",
"[3541.76s --> 3546.08s] : I think it's like hard you should never kind of assume something's going to be like the most successful product launch ever.\n",
"[3546.08s --> 3546.48s] : Um,\n",
"[3546.48s --> 3550.08s] : but we thought it was at least many of us thought it was going to be really good.\n",
"[3550.08s --> 3554.24s] : GV four has weirdly not been that much of an update for most people.\n",
"[3554.24s --> 3554.64s] : You know,\n",
"[3554.64s --> 3555.12s] : they're like,\n",
"[3555.12s --> 3555.36s] : Oh,\n",
"[3555.36s --> 3556.64s] : it's better than three point five,\n",
"[3556.64s --> 3559.28s] : but I thought it was going to be better than three point five and it's cool,\n",
"[3559.28s --> 3559.52s] : but,\n",
"[3559.52s --> 3559.84s] : you know,\n",
"[3559.84s --> 3570.00s] : this is like Some one said to me over the week end You shipped an AGI and I somehow like am just going about my daily life,\n",
"[3570.00s --> 3571.44s] : and I'm not that impressed.\n",
"[3571.44s --> 3574.96s] : And I obviously don't think we shipped an AGI,\n",
"[3574.96s --> 3577.52s] : but I get the point,\n",
"[3577.52s --> 3580.16s] : and the world is continuing on.\n",
"[3580.16s --> 3584.24s] : When you build or somebody builds an artificial general intelligence,\n",
"[3584.24s --> 3585.76s] : would that be fast or slow?\n",
"[3585.76s --> 3588.64s] : Would we know what's happening or not?\n",
"[3588.64s --> 3591.84s] : Would we go about our day on the week end or not?\n",
"[3591.84s --> 3594.80s] : So I'll come back to whether we go about our day or not?\n",
"[3594.80s --> 3595.28s] : thingng,\n",
"[3595.28s --> 3601.20s] : I think there's like a bunch of interesting lessons from Covid and the UFO videos and whole bunch of other stuff that we can talk to there.\n",
"[3601.20s --> 3603.28s] : But on the take off question.\n",
"[3603.28s --> 3607.92s] : if we imagine a two by two matrix of short time lines till AGI starts,\n",
"[3607.92s --> 3610.16s] : long time lines till AGI starts,\n",
"[3610.16s --> 3610.96s] : slow take off,\n",
"[3610.96s --> 3611.92s] : fast take off,\n",
"[3611.92s --> 3615.52s] : you have an instinct on what do you think the safest quadrant would be?\n",
"[3615.52s --> 3616.32s] : So a,\n",
"[3616.32s --> 3619.52s] : the different options are like next year Y,\n",
"[3619.52s --> 3622.32s] : say the take off the we start the take off period,\n",
"[3622.32s --> 3622.80s] : yeep,\n",
"[3622.80s --> 3624.56s] : nexte year or in twenty years,\n",
"[3624.56s --> 3625.20s] : three years,\n",
"[3625.20s --> 3628.64s] : and then it takes one year or ten years.\n",
"[3628.64s --> 3629.12s] : Well,\n",
"[3629.12s --> 3630.72s] : you can say one year or five years,\n",
"[3630.72s --> 3631.76s] : whatever you want,\n",
"[3631.76s --> 3633.12s] : for the take off.\n",
"[3633.12s --> 3638.32s] : I feel like now is er is safer.\n",
"[3638.32s --> 3639.52s] : So do I.\n",
"[3639.52s --> 3641.36s] : So I'm in longer though.\n",
"[3641.36s --> 3645.20s] : I'm in these slow take off short time lines.\n",
"[3645.20s --> 3647.28s] : It's the most likely good world,\n",
"[3647.28s --> 3654.16s] : and we optimize the company to have maximum impact in that world to try to push for that kind of a world.\n",
"[3654.16s --> 3658.00s] : and the decisions that we make are You know,\n",
"[3658.00s --> 3659.60s] : there's like probability masses,\n",
"[3659.60s --> 3661.12s] : but weighted towards that.\n",
"[3661.12s --> 3666.96s] : And I think I'm very afraid of the fast take offs.\n",
"[3666.96s --> 3670.32s] : I think in the longer time lines it's harder to have a slow take off,\n",
"[3670.32s --> 3671.52s] : there's a bunch of other problems,\n",
"[3671.52s --> 3671.92s] : too.\n",
"[3671.92s --> 3672.56s] : U,\n",
"[3672.56s --> 3674.00s] : but that's what we're trying to do.\n",
"[3674.00s --> 3676.00s] : Do you think GPT four is an AGI?\n",
"[3676.00s --> 3686.24s] : I think if it is just like with the UFO videos,\n",
"[3686.24s --> 3688.88s] : we wouldn't know immediately.\n",
"[3688.88s --> 3691.52s] : I think it's actually hard to know that.\n",
"[3691.52s --> 3699.84s] : When I've been thinking I've been playing with GPT for and thinking how would I know if it's an AGI or not?\n",
"[3699.84s --> 3702.56s] : becauseuse I think in terms of,\n",
"[3702.56s --> 3704.48s] : to put it in a different way,\n",
"[3704.48s --> 3709.60s] : how much of AGI is the interface I have with the thing,\n",
"[3709.60s --> 3711.36s] : and how much of it,\n",
"[3711.36s --> 3711.76s] : ,\n",
"[3711.76s --> 3714.16s] : is the actual wisdom inside of it.\n",
"[3714.16s --> 3721.92s] : Like a part of me thinks that you can have a model that's capable of super intelligence,\n",
"[3721.92s --> 3722.40s] : and,\n",
"[3722.40s --> 3722.64s] : ,\n",
"[3722.64s --> 3724.40s] : it just hasn't been quite unlocked.\n",
"[3724.40s --> 3726.40s] : What I saw with Shady Petit,\n",
"[3726.40s --> 3729.28s] : just doing that little bit of our with human feedback,\n",
"[3729.28s --> 3731.92s] : makes the thing somehat much more impressive,\n",
"[3731.92s --> 3732.96s] : much more usable.\n",
"[3732.96s --> 3734.72s] : So maybe if you have a few more tricks,\n",
"[3734.72s --> 3735.20s] : like you said,\n",
"[3735.20s --> 3737.36s] : there's like hundreds of tricks inside open AI,\n",
"[3737.36s --> 3738.08s] : a few more tricks,\n",
"[3738.08s --> 3739.92s] : and all a sudden holy shit.\n",
"[3739.92s --> 3743.44s] : This thing So I think that GPT fourour,\n",
"[3743.44s --> 3744.72s] : although quite impressive,\n",
"[3744.72s --> 3747.76s] : is definitely not an AGI but isn't remarkable we're having this debate?\n",
"[3747.76s --> 3750.64s] : So what's your intuition whyse that?\n",
"[3750.64s --> 3756.40s] : I think we're getting into the phase where specific definitions of AGI really matter.\n",
"[3756.40s --> 3757.60s] : Or we just say,\n",
"[3757.60s --> 3757.92s] : you know,\n",
"[3757.92s --> 3758.80s] : I know it when I see it,\n",
"[3758.80s --> 3760.96s] : and I'm not even going to bother with the definition.\n",
"[3760.96s --> 3763.52s] : U but under the I know it when I see it,\n",
"[3763.52s --> 3771.36s] : it doesn't feel that close to me.\n",
"[3771.36s --> 3779.12s] : like if if I were reading a sci fi book and there was a character that was an AGI,\n",
"[3779.12s --> 3783.28s] : and that character was GPT fourour I'll be likewise as a shitty book.\n",
"[3783.28s --> 3784.00s] : You know,\n",
"[3784.00s --> 3784.88s] : that's not very cool.\n",
"[3784.88s --> 3787.20s] : I wish I would have hoped we had done better.\n",
"[3787.20s --> 3790.64s] : To me some of the human factors are important here.\n",
"[3790.64s --> 3795.20s] : Do you think GPT four is conscious?\n",
"[3795.20s --> 3796.96s] : I think no,\n",
"[3796.96s --> 3799.20s] : but I asked GPT four,\n",
"[3799.20s --> 3800.24s] : and of course it says no.\n",
"[3800.24s --> 3802.32s] : Do you think GPT four is conscious?\n",
"[3802.32s --> 3810.24s] : I think it knows how to fake consciousness.\n",
"[3810.24s --> 3810.88s] : Yes.\n",
"[3810.88s --> 3812.40s] : How of fake consciousness?\n",
"[3812.40s --> 3813.04s] : Yeah.\n",
"[3813.04s --> 3818.00s] : If if if if you provide the right interface and the right prompts?\n",
"[3818.00s --> 3820.96s] : It definitely can answer as if it were.\n",
"[3820.96s --> 3821.52s] : Yeah.\n",
"[3821.52s --> 3823.60s] : And then it starts getting weird.\n",
"[3823.60s --> 3827.60s] : Liinke what is the difference between pretending to be conscious and conscious?\n",
"[3827.60s --> 3827.92s] : I mean,\n",
"[3827.92s --> 3828.40s] : Drake,\n",
"[3828.40s --> 3830.24s] : you don't know obviously we can go to like,\n",
"[3830.24s --> 3833.84s] : the freshman your dorm late Saturday night kind of thing.\n",
"[3833.84s --> 3837.52s] : You don't know that you're not a GP four rollout in some advanced simulation yet.\n",
"[3837.52s --> 3838.16s] : Yes.\n",
"[3838.16s --> 3840.96s] : So if we're willing to go to that level,\n",
"[3840.96s --> 3846.32s] : sure I'm I live in that but that's an important that's an important level.\n",
"[3846.32s --> 3847.76s] : That's an important,\n",
"[3847.76s --> 3857.04s] : a that's a really important level because one of the things that makes it not conscious is declaring that it's a computer program.\n",
"[3857.04s --> 3861.36s] : therefore it can't be a conscious so I'm not going to I'm not even going to acknowledge it.\n",
"[3861.36s --> 3864.24s] : But that just puts it in the category of other.\n",
"[3864.24s --> 3869.44s] : I believe AI can be conscious.\n",
"[3869.44s --> 3871.12s] : So then the question is,\n",
"[3871.12s --> 3873.60s] : what would it look like when it's conscious?\n",
"[3873.60s --> 3875.60s] : What would it behave like?\n",
"[3875.60s --> 3878.96s] : And it would probably say things like,\n",
"[3878.96s --> 3879.60s] : first of all,\n",
"[3879.60s --> 3885.44s] : I am conscious second of all display capability of suffering.\n",
"[3885.44s --> 3889.04s] : an understanding of self?\n",
"[3889.04s --> 3897.76s] : Of having some memory of itself and maybe interactions with you,\n",
"[3897.76s --> 3900.32s] : maybe there's a personalization aspect to it,\n",
"[3900.32s --> 3904.00s] : and I think all of those capabilities are interface capabilities,\n",
"[3904.00s --> 3908.16s] : not fundamental aspects of the actual knowledge so in your net.\n",
"[3908.16s --> 3911.04s] : Maybe I can just share a few like disconnected thoughts here,\n",
"[3911.04s --> 3911.60s] : ff.\n",
"[3911.60s --> 3914.24s] : But I'll tell you something that Ilias said to me once,\n",
"[3914.24s --> 3917.44s] : a long time ago that has like stuck in my head.\n",
"[3917.44s --> 3919.44s] : Ilias let's go yeses,\n",
"[3919.44s --> 3920.16s] : my co founder,\n",
"[3920.16s --> 3921.84s] : the chief scientist of Open the eye,\n",
"[3921.84s --> 3924.32s] : and sort of legend in the field.\n",
"[3924.32s --> 3924.96s] : Um,\n",
"[3924.96s --> 3928.80s] : we were talking about how you would know if a model were conscious or not.\n",
"[3928.80s --> 3932.08s] : and heard many ideas thrown around,\n",
"[3932.08s --> 3934.48s] : but he said one that I think is interesting.\n",
"[3934.48s --> 3946.88s] : If you trained a model on a data set that you were extremely careful to have no mentions of consciousness or anything close to it in the training process.\n",
"[3946.88s --> 3953.84s] : Li not only was the word never there but nothing about the sort of subjective experience of it or related concepts.\n",
"[3953.84s --> 3966.64s] : And then you started talking to that model about herere are some things that you weren't trained about,\n",
"[3966.64s --> 3969.92s] : and for most of them the model was like I've no idea what you're talking about.\n",
"[3969.92s --> 3977.92s] : But then you asked it you sort of described the experience the subjective experience of consciousness,\n",
"[3977.92s --> 3980.08s] : and the model immediately responded,\n",
"[3980.08s --> 3981.52s] : unlike the other questions Yes,\n",
"[3981.52s --> 3983.44s] : I know exactly what you're talking about.\n",
"[3983.44s --> 3987.28s] : That would update me somewhat.\n",
"[3987.28s --> 3989.68s] : I don't know,\n",
"[3989.68s --> 3994.16s] : because that's more in the space of facts versus like emotions.\n",
"[3994.16s --> 3996.72s] : I don't think consciousness is an emotion.\n",
"[3996.72s --> 4006.48s] : I think consciousness is an ability to sort of experience this world really deeply.h there's a movie called X Mokina.\n",
"[4006.48s --> 4008.16s] : I've heard of it but I haven't seen it.\n",
"[4008.16s --> 4008.96s] : You haven'tn it?\n",
"[4008.96s --> 4009.28s] : no?\n",
"[4009.28s --> 4012.96s] : The director Alex Garland who had a conversation,\n",
"[4012.96s --> 4016.08s] : said it's where the AGI system is built,\n",
"[4016.08s --> 4019.36s] : embodied in the body of a a woman,\n",
"[4019.36s --> 4020.16s] : and,\n",
"[4020.16s --> 4020.40s] : a,\n",
"[4020.40s --> 4022.16s] : something he doesn't make explicit,\n",
"[4022.16s --> 4026.96s] : but he said he put in the movie without describing why.\n",
"[4026.96s --> 4028.96s] : But at the end of the movie,\n",
"[4028.96s --> 4030.32s] : spoiler alert,\n",
"[4030.32s --> 4032.16s] : when the AI escapes,\n",
"[4032.16s --> 4034.24s] : the woman escapes.\n",
"[4034.24s --> 4036.00s] : Er,\n",
"[4036.00s --> 4039.52s] : she smiles For nobody.\n",
"[4039.52s --> 4040.88s] : for no audience.\n",
"[4040.88s --> 4041.84s] : Ur,\n",
"[4041.84s --> 4046.64s] : she smiles at the per like at the freedom shes experiencing.\n",
"[4046.64s --> 4047.84s] : experiencing,\n",
"[4047.84s --> 4048.32s] : I don't know,\n",
"[4048.32s --> 4049.60s] : anthropomorphizing.\n",
"[4049.60s --> 4057.28s] : but he said the smile to me was the was passing the touring test of consciousness that you smille for no audience.\n",
"[4057.28s --> 4059.04s] : You smiled for yourself.\n",
"[4059.04s --> 4060.96s] : It's an interesting thought.\n",
"[4060.96s --> 4065.76s] : It's like you taken an experience for experience sake.\n",
"[4065.76s --> 4066.96s] : I don't know.\n",
"[4066.96s --> 4067.84s] : Uh,\n",
"[4067.84s --> 4073.36s] : that seemed more like consciousness versus the ability to convince somebody else that you're conscious.\n",
"[4073.36s --> 4077.12s] : And that feels more like a realm of emotion versus facts.\n",
"[4077.12s --> 4077.76s] : But yes,\n",
"[4077.76s --> 4083.68s] : if it knows so I think there's many other tasks tests like that,\n",
"[4083.68s --> 4086.08s] : that we could look at to.\n",
"[4086.08s --> 4086.80s] : ,\n",
"[4086.80s --> 4088.64s] : but,\n",
"[4088.64s --> 4089.04s] : you know,\n",
"[4089.04s --> 4096.32s] : my personal beliefs consciousness is of something very strange is going on.\n",
"[4096.32s --> 4103.36s] : U Do you think it's attached to the particular medium of of the human brain?\n",
"[4103.36s --> 4105.68s] : Do you think ana can be conscious?\n",
"[4105.68s --> 4111.60s] : I'm certainly willing to believe that consciousness is somehow the fundamental substrate,\n",
"[4111.60s --> 4113.28s] : and we're all just in the dream or the simulation,\n",
"[4113.28s --> 4113.84s] : or whatever.\n",
"[4113.84s --> 4124.88s] : I think it's interesting how much sort of the Silicon Valley religion of the simulation has gotten close to like Brumme and how little space there is between them.\n",
"[4124.88s --> 4125.60s] : ,\n",
"[4125.60s --> 4127.28s] : but from these very different directions.\n",
"[4127.28s --> 4129.04s] : So like maybe that's what's going on.\n",
"[4129.04s --> 4136.72s] : But if if it is like physical reality as we understand it and all of the rules of the game where we think they are,\n",
"[4136.72s --> 4140.56s] : then then there's something I still think it's something very strange.\n",
"[4140.56s --> 4141.52s] : Ah,\n",
"[4141.52s --> 4143.84s] : just to linger on the alignment problem a little bit,\n",
"[4143.84s --> 4145.60s] : maybe the control problem.\n",
"[4145.60s --> 4151.84s] : What are the different ways you think AGI might go wrong that concern you?\n",
"[4151.84s --> 4157.20s] : You said that a fear a little bit of fear is very appropriate here.\n",
"[4157.20s --> 4158.80s] : He've been very transparent,\n",
"[4158.80s --> 4160.24s] : Bob being mostly excited,\n",
"[4160.24s --> 4161.04s] : but also scared.\n",
"[4161.04s --> 4165.28s] : I think it's weird when people like think it's like a big dunk that I say like I'm a little bit afraid,\n",
"[4165.28s --> 4168.40s] : and I think it wouldd be crazy not to be a little bit afraid.\n",
"[4168.40s --> 4171.44s] : And I empathise with people who are a lot afraid.\n",
"[4171.44s --> 4176.40s] : What do you think about that moment of a system becoming super intelligent?\n",
"[4176.40s --> 4177.92s] : Do you think you would know?\n",
"[4177.92s --> 4187.52s] : The current worries that I have are that there are going to be disinformation problems,\n",
"[4187.52s --> 4189.36s] : or economic shocks,\n",
"[4189.36s --> 4195.92s] : or something else at a level far beyond anything we're prepared for.\n",
"[4195.92s --> 4198.72s] : And that doesn't require superintelligence,\n",
"[4198.72s --> 4208.48s] : that doesn't require a super deep alignment problem in the machine waking up and trying to deceive us and I don't think that gets enough attention.\n",
"[4208.48s --> 4210.40s] : Ma starting to get more,\n",
"[4210.40s --> 4211.04s] : I guess.\n",
"[4211.04s --> 4219.60s] : So these systems deployed at scale can the shift the woods of geopolitics and so on.\n",
"[4219.60s --> 4220.48s] : How would we know if,\n",
"[4220.48s --> 4221.44s] : like on Twitter,\n",
"[4221.44s --> 4224.80s] : we were mostly having like LLM's,\n",
"[4224.80s --> 4230.16s] : direct the whatever's flowing through that hive mind?\n",
"[4230.16s --> 4231.52s] : Y.\n",
"[4231.52s --> 4233.76s] : on Twitter and then perhaps beyond.\n",
"[4233.76s --> 4235.12s] : and then as on Twitter,\n",
"[4235.12s --> 4236.80s] : so everywhere else eventually.\n",
"[4236.80s --> 4237.76s] : es,\n",
"[4237.76s --> 4238.88s] : how would we know?\n",
"[4238.88s --> 4241.52s] : My statement is we wouldn't.\n",
"[4241.52s --> 4244.24s] : And that's a real danger.\n",
"[4244.24s --> 4246.32s] : How do you prevent that danger?\n",
"[4246.32s --> 4249.44s] : I think there's a lot of things you can try.\n",
"[4249.44s --> 4250.16s] : Um,\n",
"[4250.16s --> 4253.84s] : but at this point it is a certainty,\n",
"[4253.84s --> 4261.44s] : there are soon going to be a lot of capable open source LLMs with very few to none no safety controls on them.\n",
"[4261.44s --> 4266.88s] : And so you can try with regulatory approaches.\n",
"[4266.88s --> 4270.72s] : You can try with using more powerful AIys to detect this stuff happening.\n",
"[4270.72s --> 4273.84s] : I'd like us to start trying a lot of things very soon.\n",
"[4273.84s --> 4279.20s] : How do you under this pressure that there's going to be a lot of open source,\n",
"[4279.20s --> 4282.16s] : there's going to be a lot of large language models?\n",
"[4282.16s --> 4284.08s] : Under this pressure,\n",
"[4284.08s --> 4287.36s] : how do you continue prioritizing safety?\n",
"[4287.36s --> 4288.16s] : Wheus,\n",
"[4288.16s --> 4288.80s] : I mean,\n",
"[4288.80s --> 4290.08s] : there's several pressures,\n",
"[4290.08s --> 4294.08s] : so one of them is a market driven pressure from other companies,\n",
"[4294.08s --> 4294.64s] : uh,\n",
"[4294.64s --> 4296.00s] : Google,\n",
"[4296.00s --> 4296.96s] : Apple,\n",
"[4296.96s --> 4298.96s] : Meta and smaller companies.\n",
"[4298.96s --> 4300.96s] : How do you resist the pressure from that?\n",
"[4300.96s --> 4302.72s] : Or how do you navigate that pressure?\n",
"[4302.72s --> 4305.52s] : You stick with what you believe and you stick to your mission,\n",
"[4305.52s --> 4305.84s] : you know.\n",
"[4305.84s --> 4314.56s] : I'm sure people will get ahead of us in all sorts of ways and take shortcuuts we're not going to take.m and we just aren't going to do that.\n",
"[4314.56s --> 4316.96s] : Hudy I'll compete them.\n",
"[4316.96s --> 4320.00s] : I think there's going to be many AGIs in the world,\n",
"[4320.00s --> 4320.64s] : so we don't have to,\n",
"[4320.64s --> 4322.00s] : like outcomete every one.\n",
"[4322.00s --> 4323.92s] : We're going to contribute one?\n",
"[4323.92s --> 4326.56s] : Other people are going to contribute some.\n",
"[4326.56s --> 4333.44s] : I think of I think multiple AGIs in the world with some differences in how they're built and what they do and what they're focused on.\n",
"[4333.44s --> 4335.28s] : I think that's good.\n",
"[4335.28s --> 4336.08s] : Um,\n",
"[4336.08s --> 4338.24s] : we have a very unusual structure,\n",
"[4338.24s --> 4342.00s] : so we don't have this incentive to capture unlimited value.\n",
"[4342.00s --> 4343.44s] : I worry about the people who do,\n",
"[4343.44s --> 4343.68s] : but,\n",
"[4343.68s --> 4343.92s] : you know,\n",
"[4343.92s --> 4345.60s] : hopefully it's all going to work out.\n",
"[4345.60s --> 4351.28s] : But we're a weird Org and we're good at resisting productf.\n",
"[4351.28s --> 4355.04s] : We have been a misunderstood and badly mocked Org for a long time.\n",
"[4355.04s --> 4359.12s] : like when we started we like,\n",
"[4359.12s --> 4361.36s] : announced the Org at the end of twenty fifteen,\n",
"[4361.36s --> 4365.76s] : and said we were going to work on AGI Like people thought we were batchit and insane.\n",
"[4365.76s --> 4366.80s] : Ye You know,\n",
"[4366.80s --> 4378.16s] : like I I I remember at the time a a eminent AI scientist at a large industrial AI lab was like deming individual reporters,\n",
"[4378.16s --> 4379.04s] : being like,\n",
"[4379.04s --> 4379.52s] : you know,\n",
"[4379.52s --> 4382.72s] : these people aren't very good and it's ridiculous to talk about AGI,\n",
"[4382.72s --> 4384.40s] : and I can't believe you're givn them time of day,\n",
"[4384.40s --> 4385.04s] : and it's like,\n",
"[4385.04s --> 4386.24s] : that was the level of,\n",
"[4386.24s --> 4386.48s] : like,\n",
"[4386.48s --> 4391.28s] : pettiness and rancor in the field that a new group of people saying we're going to try to build AGI.\n",
"[4391.28s --> 4399.68s] : So opening eye and deep mind was a small collection of folks who were brave enough to talk about AGI um,\n",
"[4399.68s --> 4401.76s] : in the face of mockery.\n",
"[4401.76s --> 4403.84s] : We don't get mocked as much now?\n",
"[4403.84s --> 4406.24s] : Don't get mocked as much now.\n",
"[4406.24s --> 4406.88s] : Ah,\n",
"[4406.88s --> 4412.88s] : so I was speaking about the structure of the of the of the organ.\n",
"[4412.88s --> 4416.16s] : So open A Iy went,\n",
"[4416.16s --> 4416.80s] : um,\n",
"[4416.80s --> 4420.32s] : stopped being non prorofit or split up in aw.\n",
"[4420.32s --> 4421.52s] : Can you describe the whole process?\n",
"[4421.52s --> 4422.56s] : Y.\n",
"[4422.56s --> 4424.08s] : We started as a nonprofit.\n",
"[4424.08s --> 4430.16s] : We learned early on that we were going to need far more capital than we were able to raise as a nonprofit.\n",
"[4430.16s --> 4433.20s] : Our nonprofit is still fully in charge.\n",
"[4433.20s --> 4435.60s] : There is a subsidiary capped profit,\n",
"[4435.60s --> 4439.76s] : so that our investors and employees can earn a certain fixed return.\n",
"[4439.76s --> 4441.28s] : And then beyond that,\n",
"[4441.28s --> 4443.12s] : everything else flows to the non profit,\n",
"[4443.12s --> 4445.36s] : and the non profit is like invoting control,\n",
"[4445.36s --> 4448.48s] : lets us make a bunch of non standard decisions,\n",
"[4448.48s --> 4451.52s] : um can cancel equity can do a whole bunch of other things,\n",
"[4451.52s --> 4454.24s] : can let us merge with another organ,\n",
"[4454.24s --> 4454.96s] : um,\n",
"[4454.96s --> 4460.96s] : protects us from making decisions that are not in any like shareholders' interest.\n",
"[4460.96s --> 4461.76s] : Ah,\n",
"[4461.76s --> 4466.88s] : so I think it's a structure that has been important to a lot of the decisions we've made.\n",
"[4466.88s --> 4474.24s] : What went into that decision process for taking a leap from non profit to capped for profit?\n",
"[4474.24s --> 4477.44s] : What are the pros and cas you were deciding at the time?\n",
"[4477.44s --> 4483.12s] : mean this was it was nineteen It was really like To do what we needed to go do,\n",
"[4483.12s --> 4486.56s] : we had tried and failed enough to raise the money as a non profit.\n",
"[4486.56s --> 4488.32s] : We didn't see a path forward there.\n",
"[4488.32s --> 4491.52s] : So we needed some of the benefits of capitalism,\n",
"[4491.52s --> 4492.64s] : but not too much.\n",
"[4492.64s --> 4494.24s] : I remember at the time someone said,\n",
"[4494.24s --> 4494.56s] : you know,\n",
"[4494.56s --> 4496.32s] : as a non profit not enough will happen,\n",
"[4496.32s --> 4498.48s] : as a for profit too much will happen,\n",
"[4498.48s --> 4501.04s] : so we need this sort of strangeer immediate.\n",
"[4501.04s --> 4511.28s] : You kind of had this off handnd comment of you worry about the uncapped companies that play with AGI.\n",
"[4511.28s --> 4513.76s] : Can you elaborate on the worry here?\n",
"[4513.76s --> 4514.96s] : Because AGI,\n",
"[4514.96s --> 4518.24s] : out of all the technologies we have in our hands,\n",
"[4518.24s --> 4523.52s] : has the potential to make is the cap is a hundred X for open AI.\n",
"[4523.52s --> 4524.64s] : It started that it's much,\n",
"[4524.64s --> 4526.64s] : much lower for like new investors now.\n",
"[4526.64s --> 4527.52s] : You know,\n",
"[4527.52s --> 4530.00s] : AGI can make a lot more than a hundred X,\n",
"[4530.00s --> 4530.56s] : for sure.\n",
"[4530.56s --> 4536.32s] : And so how do you um likeke how do you compete like stepping outside of open Ay,\n",
"[4536.32s --> 4539.20s] : how do you look at a world where Google is playing,\n",
"[4539.20s --> 4542.88s] : where Apple and Di and Meta are playing?\n",
"[4542.88s --> 4545.60s] : We can't control what other people are going to do.\n",
"[4545.60s --> 4552.08s] : U We can try to like build something and talk about it and influence others and provide value and,\n",
"[4552.08s --> 4552.48s] : you know,\n",
"[4552.48s --> 4554.00s] : good systems for the world.\n",
"[4554.00s --> 4556.96s] : But they're going to do what they're going to do.\n",
"[4556.96s --> 4557.76s] : Now,\n",
"[4557.76s --> 4568.72s] : I I think right now there's like extremely fast and not super deliberate motion inside of some of these companies.\n",
"[4568.72s --> 4571.20s] : But already I think people are,\n",
"[4571.20s --> 4572.24s] : as they see,\n",
"[4572.24s --> 4574.16s] : the rate of progress.\n",
"[4574.16s --> 4577.84s] : Already people are grappling with what's at stake here,\n",
"[4577.84s --> 4580.16s] : and I think the better angels are going to win out.\n",
"[4580.16s --> 4583.84s] : Can you elaborate on the better angels of individuals,\n",
"[4583.84s --> 4585.92s] : the individuals companiesies,\n",
"[4585.92s --> 4586.80s] : but you know,\n",
"[4586.80s --> 4591.12s] : the incentives of capitalism to create and capture unlimited value?\n",
"[4591.12s --> 4593.60s] : I'm a little afraid of.\n",
"[4593.60s --> 4594.64s] : but again,\n",
"[4594.64s --> 4594.96s] : no,\n",
"[4594.96s --> 4596.56s] : I think no one wants to destroy the world.\n",
"[4596.56s --> 4598.96s] : No one looks saying like today I want to desroy the world.\n",
"[4598.96s --> 4601.28s] : So we've got the the Mollch problem.\n",
"[4601.28s --> 4602.00s] : On the other hand,\n",
"[4602.00s --> 4603.76s] : we've got people who are very aware of that,\n",
"[4603.76s --> 4612.24s] : and I think a lot of healthy conversation about how can we collaborate to minimise some of these very scary downsides.\n",
"[4612.24s --> 4614.64s] : Well,\n",
"[4614.64s --> 4616.32s] : nobody wants to destroy the world,\n",
"[4616.32s --> 4617.92s] : and may ask you a tough question,\n",
"[4617.92s --> 4626.56s] : so you are very likely to be one of not the person that creates AGI.\n",
"[4626.56s --> 4629.68s] : One of one of An of then,\n",
"[4629.68s --> 4631.20s] : like we're on a team of many,\n",
"[4631.20s --> 4632.72s] : yet there' will be many teams.\n",
"[4632.72s --> 4636.00s] : But several small number of people nevertheless,\n",
"[4636.00s --> 4636.80s] : relative.\n",
"[4636.80s --> 4642.72s] : I do think it's strange that it may be a few tens of thousands of people in the world few thousands beyond the world.\n",
"[4642.72s --> 4648.00s] : But there will be a room with a few folks who are like holy shit.\n",
"[4648.00s --> 4650.00s] : That happens more often than you would think now.\n",
"[4650.00s --> 4650.80s] : I understand,\n",
"[4650.80s --> 4653.28s] : I understand this I understand this.\n",
"[4653.28s --> 4653.76s] : Yes,\n",
"[4653.76s --> 4655.12s] : There will be more Surooms,\n",
"[4655.12s --> 4657.84s] : which is a beautiful place to be in the world,\n",
"[4657.84s --> 4659.04s] : a terrifying,\n",
"[4659.04s --> 4660.40s] : but mostly beautiful.\n",
"[4660.40s --> 4665.12s] : So that might make you and a handful of folks,\n",
"[4665.12s --> 4667.36s] : the most powerful humans on earth.\n",
"[4667.36s --> 4670.08s] : Do you worry that power might corrupt you?\n",
"[4670.08s --> 4671.36s] : For sure.\n",
"[4671.36s --> 4672.32s] : Um look,\n",
"[4672.32s --> 4689.20s] : I don't I think you want decisions about this technology and certainly decisions about who is running this technology to become increasingly democratic over time.\n",
"[4689.20s --> 4691.76s] : We haven't figured out quite how to do this,\n",
"[4691.76s --> 4700.40s] : but we part of the reason for deploying like this is to get the world to have time to adapt and to reflect,\n",
"[4700.40s --> 4701.52s] : and to think about this,\n",
"[4701.52s --> 4704.88s] : to pass regulation for institutions to come up with new norms,\n",
"[4704.88s --> 4709.44s] : for the people working out together like Thatha is a huge part of why we deploy,\n",
"[4709.44s --> 4713.44s] : even though many of they I say to people you referenced earlier think it's really bad,\n",
"[4713.44s --> 4716.64s] : even they acknowledge that this is of some benefit.\n",
"[4716.64s --> 4730.00s] : U But I think any version of one person is in control of this is really bad.\n",
"[4730.00s --> 4733.52s] : so try to distribute the power I don't have and I don't want,\n",
"[4733.52s --> 4736.24s] : like any like super voting power or any special like them,\n",
"[4736.24s --> 4736.56s] : you know,\n",
"[4736.56s --> 4739.84s] : I no like control of the board or anything like that of openenyI.\n",
"[4739.84s --> 4744.08s] : But AGI,\n",
"[4744.08s --> 4744.88s] : if graded,\n",
"[4744.88s --> 4746.08s] : has a lot of power.\n",
"[4746.08s --> 4747.84s] : How do you think we're doing like honest,\n",
"[4747.84s --> 4748.96s] : how do you think we're doing so far?\n",
"[4748.96s --> 4750.16s] : Like do think our decisions are,\n",
"[4750.16s --> 4752.24s] : like do you think we're making things not better worse?\n",
"[4752.24s --> 4753.44s] : What can we do better?\n",
"[4753.44s --> 4753.84s] : Well,\n",
"[4753.84s --> 4757.12s] : the things I really like because I know a lot of folks at open AI.\n",
"[4757.12s --> 4759.36s] : The thing I you really like is the transparency,\n",
"[4759.36s --> 4760.24s] : everything you're saying,\n",
"[4760.24s --> 4762.32s] : which is like failing publicly,\n",
"[4762.32s --> 4764.16s] : writing papers,\n",
"[4764.16s --> 4770.64s] : releasing different kinds of information about the safety concerns involved,\n",
"[4770.64s --> 4774.64s] : and doing it out in the open is great,\n",
"[4774.64s --> 4778.72s] : because especially in contrast to some other companies they' are not doing that,\n",
"[4778.72s --> 4780.80s] : they're being more closed.\n",
"[4780.80s --> 4781.84s] : That said,\n",
"[4781.84s --> 4783.60s] : you could be more open.\n",
"[4783.60s --> 4786.16s] : Do you think we should open source GP four?\n",
"[4786.16s --> 4794.16s] : My personal opinion because I know people that open AI,\n",
"[4794.16s --> 4794.96s] : is no.\n",
"[4794.96s --> 4797.36s] : What does knowing the people that at open Ay have to do with it?\n",
"[4797.36s --> 4798.88s] : Because I know they're good people.\n",
"[4798.88s --> 4799.84s] : I know a lot of people,\n",
"[4799.84s --> 4801.68s] : I know they're good human beings.\n",
"[4801.68s --> 4802.16s] : Um,\n",
"[4802.16s --> 4804.56s] : From a perspective of people that don't know the human beings,\n",
"[4804.56s --> 4809.28s] : there's a concern of the super powerful technology in the hands of a few that's closed.\n",
"[4809.28s --> 4810.72s] : It's closed in some sense,\n",
"[4810.72s --> 4812.48s] : but we give more access to it.\n",
"[4812.48s --> 4812.96s] : Yeah.\n",
"[4812.96s --> 4813.84s] : Then and like,\n",
"[4813.84s --> 4820.16s] : if if this had just been Google's game I feel it's very unlikely that anyone would have put this API out.\n",
"[4820.16s --> 4821.36s] : There's PR risk with it.\n",
"[4821.36s --> 4823.84s] : I get personal threats because of it all the time.\n",
"[4823.84s --> 4825.84s] : I think most companies wouldn't have done this.\n",
"[4825.84s --> 4828.48s] : So maybe we didn't go as open as people wanted,\n",
"[4828.48s --> 4831.36s] : But like we've distributed it pretty broadly.\n",
"[4831.36s --> 4838.32s] : You personally in open eyes a culture is not so like nervous about a PR risk and all that kind of stuff.\n",
"[4838.32s --> 4841.44s] : You're more nervous about the risk of the actual technology,\n",
"[4841.44s --> 4843.60s] : and you reveal that.\n",
"[4843.60s --> 4844.88s] : So I you know,\n",
"[4844.88s --> 4848.72s] : the nervousness that people have is because it's such early days of the technology,\n",
"[4848.72s --> 4850.56s] : is that you will close off over times,\n",
"[4850.56s --> 4852.16s] : become more and more powerful.\n",
"[4852.16s --> 4859.60s] : My nervousness is you get attacked so much by fear mongerring clickbate journalism you're like Why the hell do I need to de with this?\n",
"[4859.60s --> 4862.56s] : I think the clickbate journalism bothers you more than it bothers me.\n",
"[4862.56s --> 4863.20s] : No,\n",
"[4863.20s --> 4865.44s] : I'm a third person bothered.\n",
"[4865.44s --> 4866.88s] : Like I appreciate that.\n",
"[4866.88s --> 4868.24s] : I feel all right about it.\n",
"[4868.24s --> 4869.44s] : All the things I lose sleep over,\n",
"[4869.44s --> 4870.48s] : it's not high in the list.\n",
"[4870.48s --> 4874.64s] : Because it's important there's a handful of a companies a handful of folks that are really pushing this forward,\n",
"[4874.64s --> 4879.84s] : there' are amazing folks that don't want them to become cynical about the rest of the rest of the world.\n",
"[4879.84s --> 4885.04s] : I think people at openpen eye feel the weight of responsibility of what we're doing.\n",
"[4885.04s --> 4885.68s] : and yeah,\n",
"[4885.68s --> 4887.68s] : it would be nice if like you know,\n",
"[4887.68s --> 4892.08s] : journalists were nicer to us and Twitter trolls gave us more benefit of the doubt.\n",
"[4892.08s --> 4896.96s] : but like I think we have a lot of resolve in what we're doing and why,\n",
"[4896.96s --> 4899.04s] : and the importance of it.\n",
"[4899.04s --> 4901.12s] : But I really would,\n",
"[4901.12s --> 4901.44s] : love,\n",
"[4901.44s --> 4902.80s] : and I ask this like of a lot of people,\n",
"[4902.80s --> 4906.08s] : not just if cameras roll in like any feedback you've got for how we can be doing better.\n",
"[4906.08s --> 4907.68s] : We're in uncharted waters here.\n",
"[4907.68s --> 4910.72s] : Talking to smart people is how we figure out what to do better.\n",
"[4910.72s --> 4911.12s] : A,\n",
"[4911.12s --> 4914.24s] : howy take feedback dyu take feedback from Twitter also?\n",
"[4914.24s --> 4916.40s] : D g does the se the water?\n",
"[4916.40s --> 4918.00s] : while Twitter is unreadable.\n",
"[4918.00s --> 4918.56s] : Yeah.\n",
"[4918.56s --> 4920.00s] : So sometimes I do.\n",
"[4920.00s --> 4922.96s] : I can like take a sample a c cup out of the waterfall.\n",
"[4922.96s --> 4926.48s] : Um but I mostly take it from conversations like this.\n",
"[4926.48s --> 4926.96s] : Ah,\n",
"[4926.96s --> 4928.08s] : speaking of feedback,\n",
"[4928.08s --> 4932.88s] : somebody you know well you've worked together closely on some of the ideas behind Openyes,\n",
"[4932.88s --> 4933.84s] : Elam Musk.\n",
"[4933.84s --> 4935.52s] : You have agreed on a lot of things,\n",
"[4935.52s --> 4937.52s] : you've disagreed on some things.\n",
"[4937.52s --> 4941.28s] : What have been some interesting things you've agreed and disagreed on,\n",
"[4941.28s --> 4944.64s] : speakking of a fun debate on Twitter,\n",
"[4944.64s --> 4950.80s] : I think we agree on the magnitude of the downside of AGI,\n",
"[4950.80s --> 4955.68s] : and the need to get not only safety right,\n",
"[4955.68s --> 4960.08s] : but get to a world where people are much better off,\n",
"[4960.08s --> 4964.64s] : because AGI exists than if AGI had never been built.\n",
"[4964.64s --> 4968.88s] : What do you disagree on?\n",
"[4968.88s --> 4973.20s] : Elon is obviously attacking us some on Twitter right now,\n",
"[4973.20s --> 4974.72s] : on a few different vectors,\n",
"[4974.72s --> 4981.04s] : and I have empathy because I believe he is understandtaably so,\n",
"[4981.04s --> 4983.76s] : really stressed about AGI safety.\n",
"[4983.76s --> 4986.80s] : I'm sure there are some other motivations going on too,\n",
"[4986.80s --> 4988.96s] : but that's definitely one of them.\n",
"[4988.96s --> 4997.04s] : Um I saw this video of Elon a long time ago,\n",
"[4997.04s --> 4998.64s] : talking about space X,\n",
"[4998.64s --> 5000.88s] : maybe it on some new show,\n",
"[5000.88s --> 5011.20s] : and a lot of early pioneers in space were really bashing space X and maybe e want to,\n",
"[5011.20s --> 5016.40s] : and he was visibly very hurt by that,\n",
"[5016.40s --> 5018.32s] : and said You know,\n",
"[5018.32s --> 5019.92s] : those guys are heroes of mine,\n",
"[5019.92s --> 5023.68s] : and I socks and I wish they would see how hard we're trying.\n",
"[5023.68s --> 5026.96s] : Um I definitely grew up with Elon as a hero of mine.\n",
"[5026.96s --> 5028.88s] : Um you know,\n",
"[5028.88s --> 5031.12s] : despite him being a jerk on Twitter or whatever.\n",
"[5031.12s --> 5033.04s] : I'm happy he exists in the world.\n",
"[5033.04s --> 5042.48s] : But I wish he would do more to look at the hard work we're doing to get the stuff right.\n",
"[5042.48s --> 5044.64s] : A little bit more love.\n",
"[5044.64s --> 5048.48s] : What do you admire in the name of love aboutbody Allmussk?\n",
"[5048.48s --> 5056.16s] : I mean sohat right like he has He has driven the world forward in important ways.\n",
"[5056.16s --> 5061.84s] : I think we will get to electric vehicles much faster than we would have if he didn't exist.\n",
"[5061.84s --> 5065.60s] : I think we'll get to space much faster than we would have if he didn't exist,\n",
"[5065.60s --> 5070.48s] : and as a sort of citizen of the world,\n",
"[5070.48s --> 5072.48s] : I'm very appreciative of that.\n",
"[5072.48s --> 5073.28s] : Also,\n",
"[5073.28s --> 5075.76s] : like being a jerk on Twitter's side,\n",
"[5075.76s --> 5079.36s] : in many instances he's like a very funny and warm guy.\n",
"[5079.36s --> 5083.04s] : And though some of the jerk on Twitter thing,\n",
"[5083.04s --> 5087.28s] : as a fan of humanity laid out in its full complexity and beauty,\n",
"[5087.28s --> 5089.92s] : I enjoy the tension of ideas expressed.\n",
"[5089.92s --> 5091.60s] : So you know,\n",
"[5091.60s --> 5094.64s] : I earlier said that I admire how transparent you are,\n",
"[5094.64s --> 5101.84s] : but I like how the battles are happening before our eyes are supposed to everybody closing off inside board rooms it's all You know,\n",
"[5101.84s --> 5102.96s] : maybe I should hit back,\n",
"[5102.96s --> 5104.24s] : and maybe some day I will,\n",
"[5104.24s --> 5106.40s] : but it's not like my normal style.\n",
"[5106.40s --> 5108.48s] : It's all fascinating to watch,\n",
"[5108.48s --> 5112.32s] : and I think both of you are brilliant people,\n",
"[5112.32s --> 5116.16s] : and have early on for a long time really cared about AGI,\n",
"[5116.16s --> 5119.84s] : and had great concerns about AGI but a great hope for AGI,\n",
"[5119.84s --> 5121.28s] : and that's cool to see,\n",
"[5121.28s --> 5121.84s] : um,\n",
"[5121.84s --> 5124.40s] : these big minds having those discussions,\n",
"[5124.40s --> 5124.80s] : ,\n",
"[5124.80s --> 5127.04s] : even if they're tense at times.\n",
"[5127.04s --> 5131.92s] : I think it was Elan that said that GPT is too woke.\n",
"[5131.92s --> 5134.88s] : U I GPT too woke?\n",
"[5134.88s --> 5137.52s] : Can you still the case that it is and not?\n",
"[5137.52s --> 5140.88s] : This is Gors question about bias?\n",
"[5140.88s --> 5141.44s] : Honestly,\n",
"[5141.44s --> 5143.28s] : I barely know what 'wke' means any morere.\n",
"[5143.28s --> 5144.16s] : I did for a while,\n",
"[5144.16s --> 5145.52s] : and I feel like the word is morphed,\n",
"[5145.52s --> 5148.40s] : so I will say I think it was too biased,\n",
"[5148.40s --> 5151.44s] : and will always be.\n",
"[5151.44s --> 5156.56s] : There will be no one version of GPT that the world ever agrees is unbiased.\n",
"[5156.56s --> 5159.36s] : What I think,\n",
"[5159.36s --> 5160.24s] : is we've made a lot,\n",
"[5160.24s --> 5160.96s] : like again,\n",
"[5160.96s --> 5167.28s] : even some of our harshest critics have gone off and been tweeting about three point five to four comparisons and being like,\n",
"[5167.28s --> 5167.60s] : Wow,\n",
"[5167.60s --> 5171.68s] : these people really got a lot better Not that they don't have more work to do and we certainly do,\n",
"[5171.68s --> 5177.04s] : but I I appreciate critics who display intellectual honesty like that.\n",
"[5177.04s --> 5179.92s] : and there's been more of that than I would have thought.\n",
"[5179.92s --> 5180.56s] : Um,\n",
"[5180.56s --> 5187.68s] : We will try to get the default version to be as neutral as possible,\n",
"[5187.68s --> 5192.16s] : but as neutral as possible is not that neutral if you have to do it again for more than one person.\n",
"[5192.16s --> 5195.52s] : And so this is where more steerability,\n",
"[5195.52s --> 5197.04s] : more control in the hands of the user,\n",
"[5197.04s --> 5198.88s] : the system message in particular?\n",
"[5198.88s --> 5199.84s] : I,\n",
"[5199.84s --> 5200.16s] : I think,\n",
"[5200.16s --> 5201.44s] : the real path forward.\n",
"[5201.44s --> 5202.08s] : And,\n",
"[5202.08s --> 5202.88s] : as you pointed out,\n",
"[5202.88s --> 5207.84s] : these nuanced answers to look at something from several angles.Yes It's really really fascinating.\n",
"[5207.84s --> 5209.04s] : it's really fascinating.\n",
"[5209.04s --> 5214.40s] : Is there something to be said about the employees of a company affecting the bias of the system?\n",
"[5214.40s --> 5215.52s] : One hundred per cent.\n",
"[5215.52s --> 5216.24s] : Ah,\n",
"[5216.24s --> 5224.64s] : we try to avoid the Group think bubble.\n",
"[5224.64s --> 5228.00s] : It's harder to avoid the Ai Group think bubble that follows you everywhere.\n",
"[5228.00s --> 5230.72s] : There's all kinds of bubbles who live at hundred per.\n",
"[5230.72s --> 5231.12s] : Yeah.\n",
"[5231.12s --> 5239.20s] : I'm going on like a around the world user tour soon for a month to just go like talk to our users in different cities.\n",
"[5239.20s --> 5247.20s] : And I can like feel how much I'm craving doing that because I haven't done anything like that since in years,\n",
"[5247.20s --> 5254.48s] : I used to do that more for Weisi and to go talk to people in super different contexts.\n",
"[5254.48s --> 5256.64s] : And it doesn't work over the Internet,\n",
"[5256.64s --> 5258.00s] : like to go show up in person and,\n",
"[5258.00s --> 5258.24s] : like,\n",
"[5258.24s --> 5258.96s] : sit down and,\n",
"[5258.96s --> 5263.36s] : like go to the bars they go to and kind of like walk through the city like they do.\n",
"[5263.36s --> 5267.92s] : you learn so much and get out of the bubble so much.\n",
"[5267.92s --> 5268.88s] : Um,\n",
"[5268.88s --> 5277.44s] : I think we are much better than any other company I know of in San Francisco for not falling into the kind of like SF craziness,\n",
"[5277.44s --> 5279.84s] : but I'm sure we're still pretty deeply in it.\n",
"[5279.84s --> 5284.96s] : But is it possible to separate the bias of the model versus the bias of the employees?\n",
"[5284.96s --> 5290.64s] : The bias I'm most nervous about is the bias of the human feedback raiders.\n",
"[5290.64s --> 5291.28s] : Ah,\n",
"[5291.28s --> 5293.28s] : So what's the selection of the human?\n",
"[5293.28s --> 5297.60s] : Is there something you could speak to at a high level about the selection of the human raiders?\n",
"[5297.60s --> 5299.92s] : This is the part that we understand the least well.\n",
"[5299.92s --> 5301.68s] : we're great at the pretraining machinery.\n",
"[5301.68s --> 5302.08s] : Um,\n",
"[5302.08s --> 5304.96s] : we're now trying to figure out how we're going to select those people,\n",
"[5304.96s --> 5309.84s] : how like how we'll like verify that we get a representative sample,\n",
"[5309.84s --> 5311.68s] : how we'll do different ones for different places,\n",
"[5311.68s --> 5314.24s] : but we don't we don't that functionality built out yet.\n",
"[5314.24s --> 5317.60s] : Such a fascinating um,\n",
"[5317.60s --> 5318.88s] : science.\n",
"[5318.88s --> 5324.32s] : You clearly don't want like all American Elite University students giving you your labels.\n",
"[5324.32s --> 5324.64s] : Well,\n",
"[5324.64s --> 5324.96s] : see,\n",
"[5324.96s --> 5326.64s] : it's not about Im sorry,\n",
"[5326.64s --> 5327.76s] : I just can never resist that,\n",
"[5327.76s --> 5328.32s] : I guess.\n",
"[5328.32s --> 5330.32s] : nice.\n",
"[5330.32s --> 5336.08s] : But it's so that that's a good there's a million heuristics you can use.\n",
"[5336.08s --> 5347.52s] : That's a to me that's a shallow heuristic because a univer like any one kind of category of human that you would think would have certain beliefs might actually be really open minded and an interesting way.\n",
"[5347.52s --> 5354.08s] : so you have to like optimize for how good you are actually answering at doing these kinds of raidting tasks,\n",
"[5354.08s --> 5357.44s] : how good you are at empathizing with an experience of other humans.\n",
"[5357.44s --> 5358.48s] : That's a big one.\n",
"[5358.48s --> 5361.04s] : Like and be able to actually like,\n",
"[5361.04s --> 5366.16s] : What does the world view look like for all kinds of groups of people that would answer this differently?\n",
"[5366.16s --> 5366.40s] : I mean,\n",
"[5366.40s --> 5368.80s] : I have to do that constantly intead of like,\n",
"[5368.80s --> 5369.84s] : You've asked this a few times,\n",
"[5369.84s --> 5371.12s] : but it's something I often do.\n",
"[5371.12s --> 5371.44s] : You know,\n",
"[5371.44s --> 5372.56s] : I ask people,\n",
"[5372.56s --> 5374.00s] : in an interview or whatever,\n",
"[5374.00s --> 5377.92s] : to steal man a the beliefs of someone they really disagree with,\n",
"[5377.92s --> 5382.64s] : and the inability of a lot of people to even pretend like they're willing to do that is remarkable.\n",
"[5382.64s --> 5383.60s] : Yeah.\n",
"[5383.60s --> 5387.76s] : What I find unfortunately ever since Kovid even more so,\n",
"[5387.76s --> 5390.16s] : that there's almost an emotional barrier.\n",
"[5390.16s --> 5395.60s] : It's not even an intellectual barrier Before they even get to the intellectual there's an emotional barrier that says no.\n",
"[5395.60s --> 5399.60s] : Anyone who might possibly believe X.\n",
"[5399.60s --> 5402.48s] : They're an idiot,\n",
"[5402.48s --> 5403.68s] : they're evil,\n",
"[5403.68s --> 5405.20s] : they're malevolent,\n",
"[5405.20s --> 5406.80s] : anything you want to as sign,\n",
"[5406.80s --> 5409.76s] : it's like they're not even like loading in the data into their head.\n",
"[5409.76s --> 5410.08s] : Look,\n",
"[5410.08s --> 5414.56s] : I think we'll find out that we can make GPT systems way less bi us than any human.\n",
"[5414.56s --> 5415.36s] : Yeah.\n",
"[5415.36s --> 5420.32s] : So hopefully without the Because there won't be that emotional load there.\n",
"[5420.32s --> 5420.72s] : ,\n",
"[5420.72s --> 5422.00s] : the emotional load.\n",
"[5422.00s --> 5423.92s] : Ah But there might be pressure.\n",
"[5423.92s --> 5425.44s] : There might be political pressure.\n",
"[5425.44s --> 5425.84s] : Oh,\n",
"[5425.84s --> 5428.24s] : there might be pressure to make a biased system.\n",
"[5428.24s --> 5431.12s] : What I meant is that the technology I think will be capable of being.\n",
"[5431.12s --> 5432.56s] : Much less biased.\n",
"[5432.56s --> 5437.36s] : Do you anticipate you worry about pressures from outside sources,\n",
"[5437.36s --> 5438.32s] : fromrom society,\n",
"[5438.32s --> 5439.44s] : from politicians,\n",
"[5439.44s --> 5441.36s] : from money sources.\n",
"[5441.36s --> 5443.60s] : I both worry about it and want it.\n",
"[5443.60s --> 5444.88s] : Like you know,\n",
"[5444.88s --> 5451.12s] : to the point of worrying this bubble and wen't all these decisions like We want society to have a huge degree of in put here.\n",
"[5451.12s --> 5452.56s] : That is pressure in some point,\n",
"[5452.56s --> 5453.20s] : in some way.\n",
"[5453.20s --> 5453.52s] : Well,\n",
"[5453.52s --> 5454.24s] : there's you know,\n",
"[5454.24s --> 5455.12s] : that's what like,\n",
"[5455.12s --> 5455.44s] : ,\n",
"[5455.44s --> 5463.04s] : to some degree Twitter files have revealed that there is pressure from different organizations.\n",
"[5463.04s --> 5470.32s] : You can see in the pandemic where the CDC or some other government organization might put pressure on you know what?\n",
"[5470.32s --> 5470.88s] : ,\n",
"[5470.88s --> 5472.88s] : we're not really sure what's true,\n",
"[5472.88s --> 5477.04s] : but it's very unsafe to have these kinds of nuanced conversations now,\n",
"[5477.04s --> 5478.88s] : so let's censor all topics.\n",
"[5478.88s --> 5481.36s] : and you get a lot of those emails like,\n",
"[5481.36s --> 5481.68s] : you know,\n",
"[5481.68s --> 5482.16s] : ,\n",
"[5482.16s --> 5483.28s] : emails,\n",
"[5483.28s --> 5487.60s] : all different kinds of people reaching out to different places to put subtle,\n",
"[5487.60s --> 5488.72s] : indirect pressure,\n",
"[5488.72s --> 5489.04s] : a,\n",
"[5489.04s --> 5489.92s] : direct pressure,\n",
"[5489.92s --> 5490.48s] : financial,\n",
"[5490.48s --> 5491.36s] : political pressure,\n",
"[5491.36s --> 5491.92s] : all that kind of stuff.\n",
"[5491.92s --> 5492.16s] : Li like,\n",
"[5492.16s --> 5493.68s] : how do you survive that?\n",
"[5493.68s --> 5495.60s] : Ho do you U,\n",
"[5495.60s --> 5497.44s] : how much do you worry about that?\n",
"[5497.44s --> 5507.12s] : if GPT continues to get more and more intelligent and a source of information and knowledge for human civilization.\n",
"[5507.12s --> 5513.20s] : I think theres like a lot of like quirks about me that make me not a great CEO for open eye,\n",
"[5513.20s --> 5527.92s] : but a thing in the positive column is I think I am relatively good at not being affected by pressure for the sake of pressure.\n",
"[5527.92s --> 5530.24s] : By the way,\n",
"[5530.24s --> 5531.92s] : beautiful statement of humility,\n",
"[5531.92s --> 5534.64s] : but I have to ask what's in the negative column.\n",
"[5534.64s --> 5541.04s] : I mean too long a list no what's a good one?\n",
"[5541.04s --> 5542.00s] : I mean,\n",
"[5542.00s --> 5544.56s] : I think I'm not a great like spokesperson for the AI movement,\n",
"[5544.56s --> 5545.28s] : I'll say that.\n",
"[5545.28s --> 5549.84s] : I think there could be like a more like There could be someone who enjoyed it more,\n",
"[5549.84s --> 5550.64s] : there could be someone who is,\n",
"[5550.64s --> 5550.80s] : like,\n",
"[5550.80s --> 5551.76s] : much more charismatic,\n",
"[5551.76s --> 5553.76s] : there could be someone who like it connnects better,\n",
"[5553.76s --> 5554.08s] : I think,\n",
"[5554.08s --> 5554.56s] : with people,\n",
"[5554.56s --> 5555.36s] : than I do.\n",
"[5555.36s --> 5558.56s] : On with Cmpsk and this I think charisma is a dangerous thing.\n",
"[5558.56s --> 5560.80s] : I I think,\n",
"[5560.80s --> 5565.36s] : flaws in flaws in communication style I think is a feature,\n",
"[5565.36s --> 5566.64s] : not a bug in general.\n",
"[5566.64s --> 5569.52s] : A least for humans as at least for humans in power.\n",
"[5569.52s --> 5572.48s] : I think I've like more serious problems than that one.\n",
"[5572.48s --> 5587.36s] : Um I think I'm like a pretty disconnected from like the reality of life for most people,\n",
"[5587.36s --> 5598.00s] : and trying to really not just like empathize with but internalise what the impact on people that AGI is going to have.\n",
"[5598.00s --> 5602.48s] : I probably like feel that less than other people would.\n",
"[5602.48s --> 5604.72s] : That's really well put,\n",
"[5604.72s --> 5608.88s] : and you said like you're going to travel across the world to I'm excited to empathize with different users.\n",
"[5608.88s --> 5612.96s] : to empathize just to like I want to just like buy our users,\n",
"[5612.96s --> 5613.84s] : our developers,\n",
"[5613.84s --> 5614.32s] : our users,\n",
"[5614.32s --> 5615.68s] : a drink and say like.\n",
"[5615.68s --> 5617.68s] : Tell us what you'd like to change.\n",
"[5617.68s --> 5620.00s] : and I think one of the things we are not good,\n",
"[5620.00s --> 5622.08s] : as good at a company as I would like,\n",
"[5622.08s --> 5624.88s] : is to be a really user centric company.\n",
"[5624.88s --> 5627.84s] : and I feel like by the time it gets filtered to me,\n",
"[5627.84s --> 5629.52s] : it's like totally meaningless.\n",
"[5629.52s --> 5633.04s] : So I really just want to go talk to a lot of our users in very different contexts.\n",
"[5633.04s --> 5633.92s] : like you said,\n",
"[5633.92s --> 5644.16s] : a drink in person because I haven't actually found the right words for it but I I was a little afraid with the programming.\n",
"[5644.16s --> 5645.76s] : Emotionally.\n",
"[5645.76s --> 5647.44s] : I don't think it makes any sense.\n",
"[5647.44s --> 5649.28s] : There is a reallympic response there.\n",
"[5649.28s --> 5651.52s] : GPT makes me nervous about the future,\n",
"[5651.52s --> 5653.20s] : not in an AI safety way,\n",
"[5653.20s --> 5656.00s] : but like change change change.\n",
"[5656.00s --> 5660.00s] : And like there's a nervousness about change and more nervous than excited?\n",
"[5660.00s --> 5665.44s] : If I take away the fact that I'm an AI person and just a programmer?\n",
"[5665.44s --> 5666.24s] : More excited,\n",
"[5666.24s --> 5667.12s] : but still nervous,\n",
"[5667.12s --> 5668.16s] : like yeah,\n",
"[5668.16s --> 5670.08s] : nervous in brief moments,\n",
"[5670.08s --> 5671.68s] : especially when sleep deprived.\n",
"[5671.68s --> 5673.04s] : But there's a nervousness there.\n",
"[5673.04s --> 5674.72s] : People who say they're not nervous,\n",
"[5674.72s --> 5677.36s] : I I that's hard for me to believe.\n",
"[5677.36s --> 5678.72s] : But you're right,\n",
"[5678.72s --> 5679.36s] : it's excited.\n",
"[5679.36s --> 5680.64s] : nervous for change.\n",
"[5680.64s --> 5682.80s] : nervous whenever there is significant,\n",
"[5682.80s --> 5684.56s] : exciting kind of change,\n",
"[5684.56s --> 5685.84s] : um You know,\n",
"[5685.84s --> 5687.28s] : I've recently started using,\n",
"[5687.28s --> 5691.44s] : um I've been an emac person for a very long time and I switched to VS code.\n",
"[5691.44s --> 5693.44s] : As a copilot?\n",
"[5693.44s --> 5694.16s] : Uh,\n",
"[5694.16s --> 5698.16s] : that was one of the big co reasons.\n",
"[5698.16s --> 5698.56s] : 'cause,\n",
"[5698.56s --> 5698.80s] : like,\n",
"[5698.80s --> 5700.72s] : this is where a lot of active development.\n",
"[5700.72s --> 5704.40s] : Of course you can probably do a copilot inside,\n",
"[5704.40s --> 5705.52s] : Emac.\n",
"[5705.52s --> 5705.76s] : I mean,\n",
"[5705.76s --> 5707.76s] : I'm sure I'mSco's also pretty good.\n",
"[5707.76s --> 5708.24s] : Yeah,\n",
"[5708.24s --> 5718.32s] : There's a lot of like little little things and and big things that are just really good about VSco and I've been I can happily report in all the people who just go nuts.\n",
"[5718.32s --> 5721.12s] : But I I'm very happy it wass a very happy decision.\n",
"[5721.12s --> 5723.04s] : But there is a lot of uncertainty,\n",
"[5723.04s --> 5725.36s] : there's a lot of nervousness about it,\n",
"[5725.36s --> 5726.96s] : there's fear and so on.\n",
"[5726.96s --> 5727.84s] : Um,\n",
"[5727.84s --> 5731.44s] : about taking that leap and that's obviously a tiny leap.\n",
"[5731.44s --> 5734.24s] : But even just a leap to actly using copile,\n",
"[5734.24s --> 5737.04s] : using a generation of code.\n",
"[5737.04s --> 5738.64s] : I makes you nervous,\n",
"[5738.64s --> 5742.08s] : but ultimately my life is much better as a programmer.\n",
"[5742.08s --> 5743.28s] : Purely as a programmering,\n",
"[5743.28s --> 5745.68s] : a programmer of little things and big things,\n",
"[5745.68s --> 5746.72s] : I is much better.\n",
"[5746.72s --> 5747.76s] : But there's a nervousness,\n",
"[5747.76s --> 5749.76s] : and I think a lot of people will experience that.\n",
"[5749.76s --> 5753.52s] : experience that and you will experience that by talking to them.\n",
"[5753.52s --> 5761.04s] : And I don't know what we do with that um how we comfort people in the face of this uncertainty.\n",
"[5761.04s --> 5763.12s] : And you're getting more nervous the more you use it,\n",
"[5763.12s --> 5764.16s] : not less.\n",
"[5764.16s --> 5765.44s] : Yes.\n",
"[5765.44s --> 5766.80s] : I would have to say yes,\n",
"[5766.80s --> 5768.32s] : because I get better at using it.\n",
"[5768.32s --> 5769.20s] : Say,\n",
"[5769.20s --> 5770.80s] : the learning curve is quite steep.\n",
"[5770.80s --> 5771.28s] : Yeah.\n",
"[5771.28s --> 5776.96s] : And then there's moments when you're like Oh it generates a function beautifully.\n",
"[5776.96s --> 5779.12s] : And you sit back,\n",
"[5779.12s --> 5780.00s] : both proud,\n",
"[5780.00s --> 5781.04s] : like a parent,\n",
"[5781.04s --> 5783.92s] : but almost like proud like and scared,\n",
"[5783.92s --> 5787.04s] : that this thing will be much smarter than me?\n",
"[5787.04s --> 5788.80s] : Bo pride and,\n",
"[5788.80s --> 5791.52s] : a sadness almost like a melancholy feeling.\n",
"[5791.52s --> 5792.72s] : But ultimately joy,\n",
"[5792.72s --> 5793.04s] : I think,\n",
"[5793.04s --> 5793.44s] : yeah.\n",
"[5793.44s --> 5799.20s] : What kind of jobs do you think GPT language models would be better than humans at?\n",
"[5799.20s --> 5802.00s] : Like full like does the whole thing end to end better,\n",
"[5802.00s --> 5802.56s] : not not,\n",
"[5802.56s --> 5806.64s] : not like what it's doing with you where it's helping you be maybe ten times more productive.\n",
"[5806.64s --> 5809.04s] : Those are both good questions.\n",
"[5809.04s --> 5813.60s] : I don't I would say they're equivalent to me because if I'm ten times more productive,\n",
"[5813.60s --> 5818.32s] : wouldn't that mean that there'll be a need for much fewer programmers in the world?\n",
"[5818.32s --> 5823.44s] : I think the world is going to find out that if you can have ten times as much code at the same price you can just use even more.\n",
"[5823.44s --> 5824.80s] : You'llrite even more code.\n",
"[5824.80s --> 5826.56s] : Jut what understands my more code?\n",
"[5826.56s --> 5829.44s] : It is true that a lot more could be digitized,\n",
"[5829.44s --> 5832.80s] : there could be a lot more code and a lot more stuff.\n",
"[5832.80s --> 5835.04s] : I think there' is like a supply issue.\n",
"[5835.04s --> 5835.84s] : Y.\n",
"[5835.84s --> 5839.04s] : So in terms of really replaced jobs,\n",
"[5839.04s --> 5840.56s] : is that a worry for you?\n",
"[5840.56s --> 5842.24s] : It is.\n",
"[5842.24s --> 5842.64s] : Ah,\n",
"[5842.64s --> 5847.44s] : I'm trying to think of like a big category that I believe cann be massively impacted.\n",
"[5847.44s --> 5852.16s] : I guess I would say customer service is a category that I could see.\n",
"[5852.16s --> 5855.52s] : There are just way fewer jobs relatively soon.\n",
"[5855.52s --> 5858.16s] : I'm not even certain about that.\n",
"[5858.16s --> 5860.08s] : But I could believe it.\n",
"[5860.08s --> 5861.20s] : So like,\n",
"[5861.20s --> 5865.76s] : basic questions about when do I take this pill,\n",
"[5865.76s --> 5867.44s] : if it's a drug company,\n",
"[5867.44s --> 5868.40s] : or when,\n",
"[5868.40s --> 5870.40s] : uh I don't know why I went to that,\n",
"[5870.40s --> 5872.24s] : but like how do I use this product,\n",
"[5872.24s --> 5873.12s] : like questions Yeah,\n",
"[5873.12s --> 5873.36s] : like,\n",
"[5873.36s --> 5876.32s] : how do I use whatever whatever call center employees are doing now?\n",
"[5876.32s --> 5876.72s] : Yeah,\n",
"[5876.72s --> 5877.60s] : this is not work,\n",
"[5877.60s --> 5878.00s] : yeah,\n",
"[5878.00s --> 5878.80s] : okay.\n",
"[5878.80s --> 5880.72s] : I want to be clear,\n",
"[5880.72s --> 5886.08s] : I think like these systems will make a lot of jobs just go away,\n",
"[5886.08s --> 5887.92s] : every technological revolution does.\n",
"[5887.92s --> 5891.60s] : They will enhance many jobs and make them much better,\n",
"[5891.60s --> 5892.40s] : much more fun,\n",
"[5892.40s --> 5893.68s] : much higher paid.\n",
"[5893.68s --> 5900.72s] : And and they'll create new jobs that are difficult for us to imagine even if we're starting to see the first glimpses of them.\n",
"[5900.72s --> 5902.32s] : But um,\n",
"[5902.32s --> 5906.80s] : I heard someone last week talking about GPT four saying that,\n",
"[5906.80s --> 5907.28s] : you know,\n",
"[5907.28s --> 5908.00s] : man,\n",
"[5908.00s --> 5908.80s] : ah,\n",
"[5908.80s --> 5912.24s] : the dignity of work is just such a huge deal.\n",
"[5912.24s --> 5916.24s] : We've really got to worry like even people who think they don't like their jobs,\n",
"[5916.24s --> 5917.28s] : they really need them.\n",
"[5917.28s --> 5919.92s] : It's really important to them and to society.\n",
"[5919.92s --> 5925.28s] : And also can you believe how awful it is that France is trying to raise the retirement age?\n",
"[5925.28s --> 5932.08s] : And I think we as a society are confused about whether we want to work more or work less.\n",
"[5932.08s --> 5936.96s] : and certainly about whether most people like their jobs and get value out of their jobs or not.\n",
"[5936.96s --> 5939.36s] : Some people do I love my job I suspect you do,\n",
"[5939.36s --> 5939.84s] : too.\n",
"[5939.84s --> 5941.60s] : That's a real privilege,\n",
"[5941.60s --> 5942.96s] : not everybody gets to say that.\n",
"[5942.96s --> 5945.76s] : If we can move more of the world to better jobs,\n",
"[5945.76s --> 5950.48s] : and work to something that can be a broader concept,\n",
"[5950.48s --> 5952.88s] : not something you have to do to be able to eat,\n",
"[5952.88s --> 5957.84s] : but something you do as a creative expression and a way to find fulfilment and happiness whatever else.\n",
"[5957.84s --> 5962.64s] : Even if those jobs look extremely different from the jobs of to day I think that's great.\n",
"[5962.64s --> 5964.96s] : I'm not I'm not nervous about it at all.\n",
"[5964.96s --> 5967.28s] : You have been a proponent of UBI,\n",
"[5967.28s --> 5968.72s] : universal basic income.\n",
"[5968.72s --> 5975.28s] : In the context of AI can you describe your philosophy there of our human future with UBI?\n",
"[5975.28s --> 5976.80s] : Why why you like it,\n",
"[5976.80s --> 5978.32s] : what are some limitations?\n",
"[5978.32s --> 5980.56s] : I think it is a component?\n",
"[5980.56s --> 5982.72s] : Of something we should pursue.\n",
"[5982.72s --> 5984.32s] : It is not a full solution.\n",
"[5984.32s --> 5986.48s] : I think people work for lots of reasons,\n",
"[5986.48s --> 5987.60s] : besides money.\n",
"[5987.60s --> 5996.88s] : And I think we are going to find incredible new jobs and society as a whole,\n",
"[5996.88s --> 5999.04s] : and people's individuals are going to get much,\n",
"[5999.04s --> 5999.92s] : much richer,\n",
"[5999.92s --> 6003.84s] : but as a cushion through a dramatic transition,\n",
"[6003.84s --> 6006.88s] : and as just like You know,\n",
"[6006.88s --> 6010.24s] : I think the world should eliminate poverty if able to do so.\n",
"[6010.24s --> 6012.56s] : I think it's a great thing to do,\n",
"[6012.56s --> 6016.24s] : um as a small part of the bucket of solutions.\n",
"[6016.24s --> 6019.52s] : I helped start a project called World Coin.\n",
"[6019.52s --> 6023.68s] : umm which is a technological solution to this.\n",
"[6023.68s --> 6030.00s] : We also have funded a a like a large I think maybe the largest and most comprehensive,\n",
"[6030.00s --> 6035.28s] : universal basic income study as part of sponsor by opening eye.\n",
"[6035.28s --> 6039.76s] : And I think it's like an area we should just be be looking into.\n",
"[6039.76s --> 6043.68s] : What are some insights from that study that you gained?\n",
"[6043.68s --> 6045.84s] : We're going to finish up at the end of this year,\n",
"[6045.84s --> 6047.04s] : and we'll be able to talk about it,\n",
"[6047.04s --> 6048.72s] : hopefully very early next.\n",
"[6048.72s --> 6050.16s] : If we can linger on it,\n",
"[6050.16s --> 6057.20s] : how do you think the economic and political systems will change as AI becomes a prevalent part of society?\n",
"[6057.20s --> 6061.20s] : It's such an interesting sort of philosophical question,\n",
"[6061.20s --> 6062.48s] : looking ten,\n",
"[6062.48s --> 6062.96s] : twenty,\n",
"[6062.96s --> 6064.40s] : fifty years from now.\n",
"[6064.40s --> 6067.04s] : What does the economy look like?\n",
"[6067.04s --> 6069.52s] : What does politics look like?\n",
"[6069.52s --> 6074.88s] : Do you see significant transformations in terms of the way democracy functions even?\n",
"[6074.88s --> 6076.40s] : I love that you ask them together,\n",
"[6076.40s --> 6077.76s] : because I think they're super related.\n",
"[6077.76s --> 6082.40s] : I think the economic transformation will drive much of the political transformation here,\n",
"[6082.40s --> 6083.92s] : not the other way around.\n",
"[6083.92s --> 6103.84s] : Um my working model for the last five years has been that the two dominant changes will be that the cost of intelligence and the cost of energy are going over the next couple of decades to dramatically dramatically fall from where they are today.\n",
"[6103.84s --> 6109.44s] : And the impact of that you're already seen it with the way you now have like p you know,\n",
"[6109.44s --> 6113.36s] : programmability beyond what you had as an individual before,\n",
"[6113.36s --> 6116.32s] : is society gets much,\n",
"[6116.32s --> 6117.20s] : much richer,\n",
"[6117.20s --> 6118.16s] : much wealthier,\n",
"[6118.16s --> 6120.72s] : in ways that are probably hard to imagine.\n",
"[6120.72s --> 6121.76s] : I think,\n",
"[6121.76s --> 6128.88s] : every time that's happened before it has been that economic impact has had positive political impact as well.\n",
"[6128.88s --> 6130.56s] : and I think it does go the other way,\n",
"[6130.56s --> 6130.88s] : too,\n",
"[6130.88s --> 6134.16s] : like the socio political values of the Enlightenment,\n",
"[6134.16s --> 6144.00s] : enabled the long running technological revolution and scientific discovery process we've had for the past centuries.\n",
"[6144.00s --> 6147.92s] : but I think we're just going to see more.\n",
"[6147.92s --> 6151.76s] : I'm sure the shape will change bit,\n",
"[6151.76s --> 6155.68s] : but I think it's this long and beautiful exponential curve.\n",
"[6155.68s --> 6163.12s] : Do you think there will be more I don't know what the the term is,\n",
"[6163.12s --> 6166.64s] : but systems that resemble something like democratic socialism.\n",
"[6166.64s --> 6169.84s] : I've talked to a few folks in this pod about these kinds of topics.\n",
"[6169.84s --> 6170.56s] : Instinct,\n",
"[6170.56s --> 6170.96s] : yes,\n",
"[6170.96s --> 6172.00s] : I hope so.\n",
"[6172.00s --> 6181.68s] : So that it reallocates some resources in a way that supports kind of lifts the the people who are struggling.\n",
"[6181.68s --> 6183.68s] : I am a big believer and lifts up the floor,\n",
"[6183.68s --> 6185.60s] : and don't worry about the ceiling.\n",
"[6185.60s --> 6190.16s] : If I can er test your historical knowledge,\n",
"[6190.16s --> 6191.28s] : it's probably not going to be good,\n",
"[6191.28s --> 6195.36s] : but let's try it Why do you think I come from the Soviet Union?\n",
"[6195.36s --> 6197.92s] : why do you think communism of the Soviet Union failed?\n",
"[6197.92s --> 6202.72s] : I recoil at the idea of living in a Communist system.\n",
"[6202.72s --> 6207.12s] : And I don't know how much of that is just the biases of the world I've grow up in,\n",
"[6207.12s --> 6212.88s] : and what I have been taught and probably more than I realise.\n",
"[6212.88s --> 6218.32s] : but I think like more individualism,\n",
"[6218.32s --> 6220.32s] : more human will,\n",
"[6220.32s --> 6226.48s] : more ability to self determine is important.\n",
"[6226.48s --> 6228.32s] : And also,\n",
"[6228.32s --> 6233.84s] : I think the ability to try new things,\n",
"[6233.84s --> 6237.68s] : and not need permission and not need some sort of central planning,\n",
"[6237.68s --> 6243.60s] : betting on human ingenuity and this sort of like distributed process,\n",
"[6243.60s --> 6244.72s] : I believe,\n",
"[6244.72s --> 6253.04s] : is always going to beat centralized planning And I think that like for all of the deep flaws of America,\n",
"[6253.04s --> 6255.44s] : I think it is the greatest place in the world,\n",
"[6255.44s --> 6257.76s] : because it's the best at this.\n",
"[6257.76s --> 6260.72s] : So it's really interesting,\n",
"[6260.72s --> 6261.04s] : ,\n",
"[6261.04s --> 6266.64s] : that centralized planning failed so in such big ways.\n",
"[6266.64s --> 6268.00s] : But what if,\n",
"[6268.00s --> 6268.88s] : hypothetically,\n",
"[6268.88s --> 6272.32s] : the centralized planning is the perfect superintelligent AGI,\n",
"[6272.32s --> 6274.32s] : superintelligent AGI?\n",
"[6274.32s --> 6276.24s] : Again,\n",
"[6276.24s --> 6279.92s] : it might go wrong in the same kind of ways,\n",
"[6279.92s --> 6280.72s] : but it might not.\n",
"[6280.72s --> 6282.00s] : We don't really know.\n",
"[6282.00s --> 6283.28s] : We don't really know.\n",
"[6283.28s --> 6284.00s] : It might be better.\n",
"[6284.00s --> 6285.36s] : I expect it would be better.\n",
"[6285.36s --> 6296.56s] : But would it be better than a hundred superintelligent or a thousand superintelligent AGIs sort of in a liberal democratic system.\n",
"[6296.56s --> 6297.84s] : Argu?\n",
"[6297.84s --> 6298.88s] : Yes.\n",
"[6298.88s --> 6303.44s] : noww also how much of that can happen internally in one superintelligent,\n",
"[6303.44s --> 6304.24s] : AGI?\n",
"[6304.24s --> 6306.08s] : Not so obvious.\n",
"[6306.08s --> 6309.44s] : There is something about right,\n",
"[6309.44s --> 6311.52s] : but there is something about like tension,\n",
"[6311.52s --> 6312.64s] : the competition.\n",
"[6312.64s --> 6315.52s] : But you don't know that's not happening inside one model.\n",
"[6315.52s --> 6316.32s] : Yeah,\n",
"[6316.32s --> 6317.60s] : that's true.\n",
"[6317.60s --> 6324.64s] : It'd be nice it'd be nice if whether it's engineered in orrevealed to be happening,\n",
"[6324.64s --> 6326.64s] : it'd be nice for it to be happening.\n",
"[6326.64s --> 6330.96s] : that then of course it can happen with multiple AGI's talking to each other or whatever.\n",
"[6330.96s --> 6332.96s] : There's something also about,\n",
"[6332.96s --> 6333.60s] : M.\n",
"[6333.60s --> 6336.00s] : Russell has talked about the control problem of,\n",
"[6336.00s --> 6340.80s] : um always having AGI to be have some degree of uncertainty,\n",
"[6340.80s --> 6343.92s] : not having a dogmatic certainty to it.\n",
"[6343.92s --> 6345.68s] : That feels important.\n",
"[6345.68s --> 6348.64s] : someme of that is already handled with human alignment,\n",
"[6348.64s --> 6350.80s] : a er human feedback,\n",
"[6350.80s --> 6352.88s] : reinforcement learning with human feedback,\n",
"[6352.88s --> 6357.12s] : but it feels like there has to be engineered in like a hard uncertainty.\n",
"[6357.12s --> 6358.00s] : Humility,\n",
"[6358.00s --> 6360.16s] : you can put a romantic word to it.\n",
"[6360.16s --> 6360.80s] : .\n",
"[6360.80s --> 6362.96s] : Do you think that's possible to do?\n",
"[6362.96s --> 6365.28s] : The definition of those words,\n",
"[6365.28s --> 6365.68s] : I think.\n",
"[6365.68s --> 6366.88s] : The details really matter,\n",
"[6366.88s --> 6368.00s] : but as I understand them,\n",
"[6368.00s --> 6368.40s] : yes,\n",
"[6368.40s --> 6368.80s] : I do.\n",
"[6368.80s --> 6370.48s] : What about the off switch?\n",
"[6370.48s --> 6373.52s] : That like big red button in the Da Center we don't tell know anything about.\n",
"[6373.52s --> 6375.68s] : I you that what I'm a fan?\n",
"[6375.68s --> 6376.72s] : My backpack?\n",
"[6376.72s --> 6378.00s] : Hain your backpack?\n",
"[6378.00s --> 6378.56s] : Ah,\n",
"[6378.56s --> 6380.40s] : you think it's possible to have a switch?\n",
"[6380.40s --> 6381.36s] : You think I mean,\n",
"[6381.36s --> 6383.28s] : I you more more seriously,\n",
"[6383.28s --> 6387.44s] : more specifically about sort of rolling out of different systems.\n",
"[6387.44s --> 6391.20s] : Do you think it's possible to roll them unroll them,\n",
"[6391.20s --> 6392.72s] : Pull them back in.\n",
"[6392.72s --> 6393.28s] : Yeah.\n",
"[6393.28s --> 6393.68s] : I mean,\n",
"[6393.68s --> 6397.04s] : we can absolutely take a model back off the Internet.\n",
"[6397.04s --> 6399.92s] : We can like take we can turn an API off.\n",
"[6399.92s --> 6401.36s] : Isn't that something you worry about,\n",
"[6401.36s --> 6404.64s] : like when you release it and millions of people are using it,\n",
"[6404.64s --> 6406.96s] : and like you realize holy crap,\n",
"[6406.96s --> 6409.60s] : they're using it a for I don't know,\n",
"[6409.60s --> 6413.04s] : worrying about the like all kinds of terrible use cases.\n",
"[6413.04s --> 6414.96s] : We do worry about that a lot.\n",
"[6414.96s --> 6415.36s] : I mean,\n",
"[6415.36s --> 6421.28s] : we try to figure out with as much red teaming and testing ahead of time as we do.\n",
"[6421.28s --> 6423.68s] : how to avoid a lot of those,\n",
"[6423.68s --> 6433.04s] : but I can't emphasize enough how much the collective intelligence and creativity of the world will beat open eye on all the red Tingers we can hire.\n",
"[6433.04s --> 6435.28s] : So we put it out,\n",
"[6435.28s --> 6437.52s] : but we put it out in a way we can make changes.\n",
"[6437.52s --> 6441.44s] : In the millions of people that' have used the ChadGPT and GPT,\n",
"[6441.44s --> 6443.92s] : what have you learned about human civilization in general?\n",
"[6443.92s --> 6444.72s] : I mean,\n",
"[6444.72s --> 6445.84s] : though the question I ask is,\n",
"[6445.84s --> 6447.60s] : Are we mostly good?\n",
"[6447.60s --> 6452.32s] : O is there a lot of malevolence in the human spirit?\n",
"[6452.32s --> 6452.80s] : Well,\n",
"[6452.80s --> 6453.60s] : to be clear,\n",
"[6453.60s --> 6459.04s] : I don't nor does anyone else open eyes that they're like reading all the chat GVT messages.\n",
"[6459.04s --> 6459.52s] : Yeah.\n",
"[6459.52s --> 6464.72s] : But from what I hear people using it for,\n",
"[6464.72s --> 6466.24s] : at least the people I talk to,\n",
"[6466.24s --> 6468.72s] : and from what I see on Twitter,\n",
"[6468.72s --> 6471.04s] : we are definitely mostly good.\n",
"[6471.04s --> 6476.48s] : but A not all of us are,\n",
"[6476.48s --> 6478.16s] : all the time,\n",
"[6478.16s --> 6479.04s] : and B,\n",
"[6479.04s --> 6482.72s] : we really want to push on the edges of these systems,\n",
"[6482.72s --> 6484.80s] : and you know,\n",
"[6484.80s --> 6487.12s] : we really want to test out some darker theories.\n",
"[6487.12s --> 6488.40s] : The of the world?\n",
"[6488.40s --> 6488.96s] : Yes,\n",
"[6488.96s --> 6490.24s] : it's very interesting.\n",
"[6490.24s --> 6491.60s] : It's veryery interesting,\n",
"[6491.60s --> 6498.16s] : and I think that's not that that actually doesn't communicate the fact that we're like fundamentally dark inside,\n",
"[6498.16s --> 6501.44s] : but we like to go to the dark places in order to,\n",
"[6501.44s --> 6503.76s] : um uh,\n",
"[6503.76s --> 6505.76s] : maybe rediscover the light?\n",
"[6505.76s --> 6508.56s] : It feels like dark humor is a part of that.\n",
"[6508.56s --> 6513.04s] : someme of the darkest some of the toughest things you go through if you suffer in life in a war zone.\n",
"[6513.04s --> 6513.36s] : U,\n",
"[6513.36s --> 6517.36s] : the people I've interacted with they arere in the midst of a war they're used jok joke around.\n",
"[6517.36s --> 6518.56s] : They're joking around.\n",
"[6518.56s --> 6520.08s] : And they're dark jokes.\n",
"[6520.08s --> 6523.60s] : So that there's something there.\n",
"[6523.60s --> 6525.60s] : I totally agree about that tension.\n",
"[6525.60s --> 6525.92s] : Ah,\n",
"[6525.92s --> 6531.36s] : So just to the model how do you decide what isn't isn't misinformation?\n",
"[6531.36s --> 6532.96s] : How do you decide what is true?\n",
"[6532.96s --> 6536.00s] : You actually have O A as internal factual performance benchmark.\n",
"[6536.00s --> 6538.08s] : There's a lot of cool benchmarks here.\n",
"[6538.08s --> 6538.64s] : Ah,\n",
"[6538.64s --> 6541.44s] : how do you build a benchmark for what is true?\n",
"[6541.44s --> 6543.12s] : What is truth?\n",
"[6543.12s --> 6544.56s] : about Sam Albin.\n",
"[6544.56s --> 6545.84s] : Like math is true,\n",
"[6545.84s --> 6550.64s] : and the origin of Covid though is not agreed upon as ground truth.\n",
"[6550.64s --> 6552.72s] : thoseo the two things.\n",
"[6552.72s --> 6556.96s] : And then there's stuff that's like certainly not true.\n",
"[6556.96s --> 6565.36s] : But between that first and second milestone there's a lot of disagreement.\n",
"[6565.36s --> 6566.88s] : Where do you look for?\n",
"[6566.88s --> 6569.68s] : Where can a not even just now,\n",
"[6569.68s --> 6570.88s] : but in the future.\n",
"[6570.88s --> 6577.04s] : Wre can we as a human civization look for look to for truth?\n",
"[6577.04s --> 6578.16s] : What do you know,\n",
"[6578.16s --> 6578.88s] : is true?\n",
"[6578.88s --> 6580.80s] : What are you absolutely certain,\n",
"[6580.80s --> 6581.60s] : iss true?\n",
"[6581.60s --> 6586.96s] : I have,\n",
"[6586.96s --> 6587.36s] : a,\n",
"[6587.36s --> 6590.16s] : generally epistheemic humiliated about everything,\n",
"[6590.16s --> 6593.44s] : and am freaked out by how little I know and understand about the world,\n",
"[6593.44s --> 6596.08s] : so that even that question is terrifying to me.\n",
"[6596.08s --> 6602.48s] : U There's a bucket of things that have a high degree of truth inness,\n",
"[6602.48s --> 6605.36s] : which is where you will put math a lot of math,\n",
"[6605.36s --> 6605.84s] : yet.\n",
"[6605.84s --> 6607.04s] : Can't be certain,\n",
"[6607.04s --> 6610.00s] : but it's good enough for like this conversation we can say math is true.\n",
"[6610.00s --> 6610.48s] : Yeah,\n",
"[6610.48s --> 6611.20s] : I mean,\n",
"[6611.20s --> 6612.16s] : some a,\n",
"[6612.16s --> 6613.68s] : quite a bit of physics,\n",
"[6613.68s --> 6616.72s] : the historical facts,\n",
"[6616.72s --> 6617.68s] : ,\n",
"[6617.68s --> 6620.08s] : maybe dates of when a war started.\n",
"[6620.08s --> 6624.80s] : There's a lot of details about a military conflict inside history.\n",
"[6624.80s --> 6625.76s] : Of course,\n",
"[6625.76s --> 6626.80s] : you start to get,\n",
"[6626.80s --> 6627.20s] : you know,\n",
"[6627.20s --> 6629.36s] : um just read Blitz,\n",
"[6629.36s --> 6630.16s] : which is the Oh,\n",
"[6630.16s --> 6630.96s] : I want to read that?\n",
"[6630.96s --> 6631.28s] : .\n",
"[6631.28s --> 6632.56s] : how was it?\n",
"[6632.56s --> 6634.16s] : It was really good.\n",
"[6634.16s --> 6634.56s] : It's,\n",
"[6634.56s --> 6646.96s] : a it gives a theory of Nazi Germany and Hitler that so much can be described about Hitler and a lot of the upper echelon of Nazi Germany through the excessive use of drugs.\n",
"[6646.96s --> 6648.32s] : Andhetaminine,\n",
"[6648.32s --> 6648.72s] : right.\n",
"[6648.72s --> 6649.84s] : Andhetaminines,\n",
"[6649.84s --> 6650.56s] : but also other stuff,\n",
"[6650.56s --> 6652.48s] : but it's just just a lot.\n",
"[6652.48s --> 6653.52s] : And,\n",
"[6653.52s --> 6654.00s] : uh,\n",
"[6654.00s --> 6654.32s] : you know,\n",
"[6654.32s --> 6655.20s] : that's really interesting,\n",
"[6655.20s --> 6658.16s] : it's really compelling and for some reason like Whoa,\n",
"[6658.16s --> 6662.24s] : that's really that would explain a lot that's somehow really sticky.\n",
"[6662.24s --> 6670.56s] : It's an idea that's sticky and then you read a lot of criticism of that book later by historians that's actually there's a lot of cherry picking going on,\n",
"[6670.56s --> 6674.00s] : and it actually is using the fact that that's a very sticky explanation.\n",
"[6674.00s --> 6683.60s] : There's something about humans that likes a very simple narrow describe every and then yet too much amphetamins cause the war is like a great even if not true,\n",
"[6683.60s --> 6689.84s] : simple explanation that feels satisfying and excuses a lot of other,\n",
"[6689.84s --> 6691.04s] : probably much darker,\n",
"[6691.04s --> 6692.24s] : human truths.\n",
"[6692.24s --> 6692.72s] : Yes,\n",
"[6692.72s --> 6696.24s] : The military strategy employed,\n",
"[6696.24s --> 6696.64s] : uh,\n",
"[6696.64s --> 6698.08s] : the atrocities,\n",
"[6698.08s --> 6700.08s] : the speeches,\n",
"[6700.08s --> 6700.80s] : uhh,\n",
"[6700.80s --> 6704.00s] : the just the way Hitle was as a human being,\n",
"[6704.00s --> 6705.52s] : the way Hitler was as a leader,\n",
"[6705.52s --> 6708.24s] : all that could be explained to this one little lens,\n",
"[6708.24s --> 6708.80s] : and it's like,\n",
"[6708.80s --> 6709.04s] : well,\n",
"[6709.04s --> 6710.80s] : if you say that's true,\n",
"[6710.80s --> 6712.40s] : that's a really compelling truth.\n",
"[6712.40s --> 6720.88s] : So maybe truth is in one sense as defined as a thing that is a collective intelligence we kind of all our brains are sticking to.\n",
"[6720.88s --> 6721.44s] : and we're like,\n",
"[6721.44s --> 6721.76s] : Ya,\n",
"[6721.76s --> 6722.56s] : y,y,ya,\n",
"[6722.56s --> 6725.36s] : A bunch of a bunch of ants get together and like,\n",
"[6725.36s --> 6725.68s] : Yeah,\n",
"[6725.68s --> 6727.36s] : this is it I was going to say cheap,\n",
"[6727.36s --> 6730.08s] : but there's a connotation to that But,\n",
"[6730.08s --> 6730.40s] : yeah,\n",
"[6730.40s --> 6732.24s] : it' it's hard to know what is true,\n",
"[6732.24s --> 6737.68s] : and and I think when constructing a GPT like model you have to contend with that.\n",
"[6737.68s --> 6739.28s] : I think a lot of the answers,\n",
"[6739.28s --> 6739.52s] : you know,\n",
"[6739.52s --> 6743.04s] : like if you ask GPT four I don' know,\n",
"[6743.04s --> 6745.68s] : just to stick on the same topic did covidly from a lab.\n",
"[6745.68s --> 6746.00s] : .\n",
"[6746.00s --> 6749.84s] : I expect you would get a reasonable answer is a really good answer.\n",
"[6749.84s --> 6752.88s] : It laid out the hypotheses.\n",
"[6752.88s --> 6756.16s] : The the interesting thing it said,\n",
"[6756.16s --> 6758.32s] : which is refreshing to hear,\n",
"[6758.32s --> 6766.08s] : is there's something like there's very little evidence for either hypothesis direct evidence which is important to state.\n",
"[6766.08s --> 6774.96s] : A lot of people kind of the reason why there's a lot of uncertainty and a lot of debate is because there's not strong physical evidence of either,\n",
"[6774.96s --> 6777.28s] : Heavy circustantial evidence on either side,\n",
"[6777.28s --> 6781.12s] : and the other is more like biological theoretical kind of,\n",
"[6781.12s --> 6781.52s] : um,\n",
"[6781.52s --> 6782.48s] : discussion.\n",
"[6782.48s --> 6788.00s] : and I think the answer the nuanced answer to the GPT provider was actually pretty damned good,\n",
"[6788.00s --> 6794.96s] : and also importantly saying that there is uncertainty just just the fact that there is uncertainty as the statement was really powerful.\n",
"[6794.96s --> 6795.36s] : Man,\n",
"[6795.36s --> 6795.92s] : remember when,\n",
"[6795.92s --> 6796.08s] : like,\n",
"[6796.08s --> 6801.28s] : the social media platforms were banning people for Saying it was a obliak?\n",
"[6801.28s --> 6802.16s] : Yeah.\n",
"[6802.16s --> 6804.72s] : That's really humbling the humbling,\n",
"[6804.72s --> 6807.44s] : the overreach of power in censorship.\n",
"[6807.44s --> 6810.80s] : But that you the more powerful GPy becomes,\n",
"[6810.80s --> 6813.04s] : the more presious there'll be to censor.\n",
"[6813.04s --> 6820.00s] : We have a different set of challenges faced by the previous generation of companies,\n",
"[6820.00s --> 6826.32s] : which is people talk about free speech issues with GPT,\n",
"[6826.32s --> 6827.52s] : but it's not quite the same thing.\n",
"[6827.52s --> 6830.00s] : It's not like this is a computer program,\n",
"[6830.00s --> 6830.80s] : what it's allow to say,\n",
"[6830.80s --> 6838.56s] : and it's also not about the mass spread and the challenges that I think may have made that Twitter and Facebook and others have struggled with so much.\n",
"[6838.56s --> 6842.08s] : so we will have very significant challenges,\n",
"[6842.08s --> 6844.32s] : but they' will be very new and very different.\n",
"[6844.32s --> 6847.36s] : And maybe ye,\n",
"[6847.36s --> 6848.00s] : very new,\n",
"[6848.00s --> 6849.60s] : very differentffence is the way to put it.\n",
"[6849.60s --> 6852.32s] : There could be truths that are harmful in their truth.\n",
"[6852.32s --> 6853.12s] : U,\n",
"[6853.12s --> 6854.24s] : I don't know.\n",
"[6854.24s --> 6856.32s] : Group difference isn't like you.\n",
"[6856.32s --> 6857.52s] : There you go.\n",
"[6857.52s --> 6862.48s] : Scientific work that when spoken might do more harm.\n",
"[6862.48s --> 6864.48s] : And you ask GPT that,\n",
"[6864.48s --> 6865.84s] : should GPT tell you?\n",
"[6865.84s --> 6869.84s] : There's books written on this that are rigorous scientifically,\n",
"[6869.84s --> 6872.08s] : but are very uncomfortable,\n",
"[6872.08s --> 6875.20s] : and probably not productive in any sense.\n",
"[6875.20s --> 6876.24s] : but maybe are.\n",
"[6876.24s --> 6879.28s] : There's people are arguing all kinds of sides of this,\n",
"[6879.28s --> 6881.60s] : and a lot of them have hate in their heart.\n",
"[6881.60s --> 6882.80s] : And so what do you do with that?\n",
"[6882.80s --> 6886.72s] : If there's a large number of people who hate others but are actually,\n",
"[6886.72s --> 6887.28s] : um,\n",
"[6887.28s --> 6889.12s] : citing scientific studies,\n",
"[6889.12s --> 6889.76s] : what do you do with that?\n",
"[6889.76s --> 6891.12s] : What does Gputy do with that?\n",
"[6891.12s --> 6894.48s] : What is the priority of Gputy to decrease the amount of hate in the world?\n",
"[6894.48s --> 6897.36s] : Is it up to GpT or is it up to us humans?\n",
"[6897.36s --> 6899.36s] : I think we as open AI,\n",
"[6899.36s --> 6904.32s] : have responsibility for The tools we put out into the world.\n",
"[6904.32s --> 6907.04s] : I think the tools themselves can't have responsibility,\n",
"[6907.04s --> 6908.24s] : in the way I understand it.\n",
"[6908.24s --> 6908.72s] : Wow,\n",
"[6908.72s --> 6912.80s] : so you carry some of that burden for responsibility,\n",
"[6912.80s --> 6914.88s] : all of us All of us at the companyp.\n",
"[6914.88s --> 6920.56s] : So there could be harm caused by this tool,\n",
"[6920.56s --> 6922.72s] : and there will be harm caused by this tool.\n",
"[6922.72s --> 6926.64s] : there will be harm there will be tremendous benefits.\n",
"[6926.64s --> 6927.28s] : But,\n",
"[6927.28s --> 6927.76s] : you know,\n",
"[6927.76s --> 6932.40s] : tools do wonderful good and real bad.\n",
"[6932.40s --> 6936.72s] : And we will minimize the bad and maximize the good.\n",
"[6936.72s --> 6939.76s] : We have to carry the weight to that.\n",
"[6939.76s --> 6941.12s] : Ah,\n",
"[6941.12s --> 6948.16s] : how do you avoid GPT for from being hacked or jail broken?h There's a lot of interesting ways that people have done that,\n",
"[6948.16s --> 6948.72s] : like,\n",
"[6948.72s --> 6948.96s] : a,\n",
"[6948.96s --> 6953.36s] : with token smuggling or other methods like Dan.\n",
"[6953.36s --> 6954.40s] : You know,\n",
"[6954.40s --> 6956.72s] : when I was like a a kid,\n",
"[6956.72s --> 6957.28s] : basically,\n",
"[6957.28s --> 6961.04s] : I I got worked once on jail working an iPhone the first iPhone,\n",
"[6961.04s --> 6961.76s] : I think,\n",
"[6961.76s --> 6966.96s] : and I thought it was so cool.\n",
"[6966.96s --> 6971.36s] : I will say it's very strange to be on the other side of that.\n",
"[6971.36s --> 6974.40s] : No You're now the man.\n",
"[6974.40s --> 6976.96s] : Kinda sucks.\n",
"[6976.96s --> 6978.40s] : Um,\n",
"[6978.40s --> 6981.52s] : is that is some of it fun?\n",
"[6981.52s --> 6983.44s] : How much of it is a security threat?\n",
"[6983.44s --> 6983.76s] : I mean,\n",
"[6983.76s --> 6986.40s] : what how much do J have to take it seriously?\n",
"[6986.40s --> 6988.64s] : How is it even possible to solve this problem?\n",
"[6988.64s --> 6990.24s] : Where is a rank on the set of problem?\n",
"[6990.24s --> 6992.40s] : I just keep asking questions prompting.\n",
"[6992.40s --> 6998.24s] : We want users to have a lot of control,\n",
"[6998.24s --> 7005.20s] : and get the models to behave in the way they want within some very broad bounds.\n",
"[7005.20s --> 7012.72s] : And I think the whole reason for jail breaking is right now we haven't yet figured out how to like give that to people,\n",
"[7012.72s --> 7015.44s] : and the more we solve that problem,\n",
"[7015.44s --> 7018.00s] : I think the less need there'll be for jail breaking.\n",
"[7018.00s --> 7018.48s] : Yes,\n",
"[7018.48s --> 7020.00s] : kind of like piracy,\n",
"[7020.00s --> 7022.00s] : gave Bertha a Spotify.\n",
"[7022.00s --> 7024.96s] : People don't really jail break iphhoones that much any more,\n",
"[7024.96s --> 7026.00s] : And it's gotten harder,\n",
"[7026.00s --> 7026.48s] : for sure,\n",
"[7026.48s --> 7029.04s] : but also like you can just do a lot of stuff now.\n",
"[7029.04s --> 7031.36s] : Just like with the jail breakking,\n",
"[7031.36s --> 7038.48s] : that mean there's a lot of hilarity that is in U so Evan Marakawa,\n",
"[7038.48s --> 7039.28s] : Ku guy,\n",
"[7039.28s --> 7040.32s] : he's an open AI.\n",
"[7040.32s --> 7048.56s] : he tweeted something that he also was really kind to send me to communicate with me sent me long email describing the history of open AI,\n",
"[7048.56s --> 7050.16s] : all the different developments.\n",
"[7050.16s --> 7051.04s] : Um,\n",
"[7051.04s --> 7052.64s] : he really lays it out.\n",
"[7052.64s --> 7053.04s] : I mean,\n",
"[7053.04s --> 7055.84s] : that's a much longer conversation of all the awesome stuff that happened.\n",
"[7055.84s --> 7057.12s] : It's just amazing.\n",
"[7057.12s --> 7058.32s] : But his tweet was,\n",
"[7058.32s --> 7058.80s] : Ah,\n",
"[7058.80s --> 7059.52s] : Dolly,\n",
"[7059.52s --> 7060.48s] : july twenty two,\n",
"[7060.48s --> 7061.44s] : Chad GiPutine,\n",
"[7061.44s --> 7062.32s] : november twenty two,\n",
"[7062.32s --> 7064.08s] : API sixty six percent cheaper,\n",
"[7064.08s --> 7064.88s] : august twenty two,\n",
"[7064.88s --> 7066.24s] : embeddings five hundred,\n",
"[7066.24s --> 7067.28s] : times cheaper,\n",
"[7067.28s --> 7068.40s] : while state of the art,\n",
"[7068.40s --> 7069.36s] : december twenty two,\n",
"[7069.36s --> 7071.92s] : ChaiPT API also ten times cheaper,\n",
"[7071.92s --> 7073.92s] : while state of the art march twenty three,\n",
"[7073.92s --> 7075.04s] : Whisper API,\n",
"[7075.04s --> 7075.84s] : mach twenty three,\n",
"[7075.84s --> 7077.28s] : GPT four today,\n",
"[7077.28s --> 7079.12s] : whenever that was last week.\n",
"[7079.12s --> 7080.40s] : And a,\n",
"[7080.40s --> 7084.40s] : the conclusion is this team ships?\n",
"[7084.40s --> 7085.20s] : We do.\n",
"[7085.20s --> 7085.84s] : Ah,\n",
"[7085.84s --> 7089.92s] : what's the process of going and then we can extend that back I mean,\n",
"[7089.92s --> 7090.72s] : listen,\n",
"[7090.72s --> 7093.52s] : from the twenty fifteen Open Eye launch,\n",
"[7093.52s --> 7094.32s] : GPT,\n",
"[7094.32s --> 7095.28s] : GPT two,\n",
"[7095.28s --> 7096.48s] : GPT three,\n",
"[7096.48s --> 7099.12s] : Open Eye five finals with the gaming stuff,\n",
"[7099.12s --> 7100.00s] : which was an incredible.\n",
"[7100.00s --> 7101.84s] : GPT three API released.\n",
"[7101.84s --> 7102.40s] : ah,\n",
"[7102.40s --> 7103.04s] : Dolli,\n",
"[7103.04s --> 7106.40s] : instruct GPT Tech I can find find tuning.\n",
"[7106.40s --> 7106.88s] : Ah,\n",
"[7106.88s --> 7108.48s] : there's just a million things,\n",
"[7108.48s --> 7108.80s] : uh,\n",
"[7108.80s --> 7110.24s] : available Dolli,\n",
"[7110.24s --> 7111.44s] : Doli two,\n",
"[7111.44s --> 7114.88s] : a preview and then Doli is available to one million people,\n",
"[7114.88s --> 7115.84s] : Whisper,\n",
"[7115.84s --> 7117.36s] : a second model release,\n",
"[7117.36s --> 7119.44s] : just across all of the stuff,\n",
"[7119.44s --> 7121.36s] : both research and,\n",
"[7121.36s --> 7122.00s] : um,\n",
"[7122.00s --> 7125.28s] : deployment of actual products that could be in the hands of people,\n",
"[7125.28s --> 7125.68s] : ,\n",
"[7125.68s --> 7134.08s] : what is the process of going from idea to deployment that allows you to be so successful at shipping AI based products?\n",
"[7134.08s --> 7134.88s] : I mean,\n",
"[7134.88s --> 7136.88s] : there's a question of should we be really proud of that,\n",
"[7136.88s --> 7138.96s] : or should other companies be really embarrassed?\n",
"[7138.96s --> 7139.76s] : Y.\n",
"[7139.76s --> 7144.80s] : And we believe in a very high bar for the people on the team?\n",
"[7144.80s --> 7149.12s] : We work hard,\n",
"[7149.12s --> 7150.40s] : which you know,\n",
"[7150.40s --> 7152.88s] : you're not even like supposed to say any more or something.\n",
"[7152.88s --> 7161.36s] : We give a huge amount of trust and autonomy and authority to individual people,\n",
"[7161.36s --> 7165.04s] : and we try to hold each other to very high standards,\n",
"[7165.04s --> 7168.16s] : And You know,\n",
"[7168.16s --> 7170.08s] : there's a process which we can talk about,\n",
"[7170.08s --> 7171.76s] : but it won't be that illuminating.\n",
"[7171.76s --> 7177.20s] : I think it's those other things that make us able to ship at a high velocity.\n",
"[7177.20s --> 7179.76s] : So GPT Four iss a pretty complex system.\n",
"[7179.76s --> 7180.32s] : Like you said,\n",
"[7180.32s --> 7183.92s] : there's like a million little hacks you can do to keep improving it.\n",
"[7183.92s --> 7186.56s] : There's the cleaning up the data set,\n",
"[7186.56s --> 7188.64s] : all that all those are like separate teams,\n",
"[7188.64s --> 7194.80s] : so do you give a tanomy is there just autnomy to these fascinating different problems?\n",
"[7194.80s --> 7195.20s] : If,\n",
"[7195.20s --> 7201.44s] : like most people in the company weren't really excited to work super hard and collaborate well on GPD four and thought other stuff was more,\n",
"[7201.44s --> 7205.52s] : important there'd be very little I or anybody else could do to make it happen.\n",
"[7205.52s --> 7210.24s] : But we spend a lot of time figuring out what to do,\n",
"[7210.24s --> 7213.20s] : getting on the same page about why we're doing something,\n",
"[7213.20s --> 7216.80s] : and then how to divide it up and all coordinate together.\n",
"[7216.80s --> 7222.16s] : So then then you have like a passion for the for the goal here.\n",
"[7222.16s --> 7224.96s] : So everybody is really passionate across the difference.Yes,\n",
"[7224.96s --> 7225.68s] : We care.\n",
"[7225.68s --> 7227.04s] : How do you hire?\n",
"[7227.04s --> 7228.88s] : How do you hire great teams?\n",
"[7228.88s --> 7233.20s] : The folks I interacted with open eyes some of the most amazing folks've ever met.\n",
"[7233.20s --> 7234.40s] : It takes a lot of time.\n",
"[7234.40s --> 7238.00s] : Like I spend I mean,\n",
"[7238.00s --> 7240.88s] : I think a lot of people claim to spend a third of their time hiring.\n",
"[7240.88s --> 7242.72s] : I for real truly do.\n",
"[7242.72s --> 7246.24s] : U I still approve every single hired opening eye.\n",
"[7246.24s --> 7249.28s] : And I think there's you know,\n",
"[7249.28s --> 7250.56s] : we're working on a problem that is,\n",
"[7250.56s --> 7252.40s] : like very cool and the great people want to work on.\n",
"[7252.40s --> 7254.56s] : We have great people and people want to be around them.\n",
"[7254.56s --> 7261.36s] : but even with that I think there's just no short cut for putting a ton of effort into this.\n",
"[7261.36s --> 7266.00s] : So even when you have the the good people,\n",
"[7266.00s --> 7267.04s] : hard work?\n",
"[7267.04s --> 7268.40s] : I think so.\n",
"[7268.40s --> 7272.00s] : Microsoft denounced the new multi yearear,\n",
"[7272.00s --> 7277.44s] : multibillion dollar reported to be ten billion dollars investment into Open AI.\n",
"[7277.44s --> 7281.20s] : Can you describe the thinking that went into this?\n",
"[7281.20s --> 7286.64s] : what are the pros what are the cons of working with a company like Mericrosoft?\n",
"[7286.64s --> 7291.28s] : It's not all perfect or easy,\n",
"[7291.28s --> 7295.20s] : but on the whole they have been an amazing partner to us.\n",
"[7295.20s --> 7301.60s] : Satya and Kevin and Mikail are super aligned with us,\n",
"[7301.60s --> 7303.20s] : super flexible,\n",
"[7303.20s --> 7309.04s] : have gone like way above and beyond the call of duty to do things that we have needed to get all this to work.\n",
"[7309.04s --> 7313.04s] : This is like a big iron complicated engineering project,\n",
"[7313.04s --> 7316.00s] : and they are a big and complex company.\n",
"[7316.00s --> 7320.64s] : And I think like many great partnerships or relationships,\n",
"[7320.64s --> 7324.64s] : we've sort of just continued to ramp up our investment in each other.\n",
"[7324.64s --> 7326.64s] : And it's been very good?\n",
"[7326.64s --> 7329.36s] : It's a for profit company.\n",
"[7329.36s --> 7330.96s] : It's very driven.\n",
"[7330.96s --> 7333.44s] : It's very large scale.\n",
"[7333.44s --> 7337.28s] : Is there pressure to kind of make a lot of money?\n",
"[7337.28s --> 7342.48s] : I think most other companies wouldn't maybe now they would.\n",
"[7342.48s --> 7347.92s] : It wouldn't at the time have understood why we needed all the weird control provisions we have and why we need all the kind of like,\n",
"[7347.92s --> 7349.76s] : AGI specialness.\n",
"[7349.76s --> 7355.68s] : Um And I know that becauseuse I talked to some other companies before we did the first deal with Microsoft.\n",
"[7355.68s --> 7356.40s] : Um,\n",
"[7356.40s --> 7364.88s] : and I think they are unique in terms of the companies at that scale that understood why we needed the control provisions we have.\n",
"[7364.88s --> 7374.24s] : And so those control provisions help you help make sure that the capitalist imperative does not affect the development of AI.\n",
"[7374.24s --> 7376.08s] : Well,\n",
"[7376.08s --> 7380.00s] : let me just ask you as an aside about Sachan Adeella,\n",
"[7380.00s --> 7381.52s] : the CEO of Microsoft.\n",
"[7381.52s --> 7387.84s] : He seems to have successfully transformed Microsoft into this fresh,\n",
"[7387.84s --> 7388.56s] : innovative,\n",
"[7388.56s --> 7390.24s] : developer friendly company.\n",
"[7390.24s --> 7390.80s] : I agree.\n",
"[7390.80s --> 7392.24s] : What do you I mean,\n",
"[7392.24s --> 7394.56s] : is it really hard to do for a very large company?\n",
"[7394.56s --> 7395.36s] : Oh,\n",
"[7395.36s --> 7397.28s] : what have you learned from him?\n",
"[7397.28s --> 7400.08s] : Why do you think he was able to do this kind of thing?\n",
"[7400.08s --> 7400.96s] : Um,\n",
"[7400.96s --> 7401.76s] : yeah,\n",
"[7401.76s --> 7410.88s] : what insights do you have about why this one human being is able to contribute to the pivot of a large company into something very new?\n",
"[7410.88s --> 7418.72s] : I think most CEOs are either great leaders or great managers,\n",
"[7418.72s --> 7419.92s] : and,\n",
"[7419.92s --> 7423.04s] : from what I observ have observed with Satya,\n",
"[7423.04s --> 7424.72s] : he is both.\n",
"[7424.72s --> 7427.28s] : Supervisionary,\n",
"[7427.28s --> 7428.48s] : really like,\n",
"[7428.48s --> 7430.48s] : gets people excited,\n",
"[7430.48s --> 7436.24s] : really makes long duration and correct calls.\n",
"[7436.24s --> 7444.08s] : And also he is just a super effective hands on executive and I assume manager too.\n",
"[7444.08s --> 7446.48s] : And I think that's pretty rare.\n",
"[7446.48s --> 7448.32s] : I mean,\n",
"[7448.32s --> 7449.04s] : Microsoft,\n",
"[7449.04s --> 7449.68s] : I'm guessing,\n",
"[7449.68s --> 7452.96s] : like I be able like a lot of companies that have been at it for a while,\n",
"[7452.96s --> 7454.08s] : probably have,\n",
"[7454.08s --> 7457.04s] : like old school kind of momentum.\n",
"[7457.04s --> 7458.08s] : So you like,\n",
"[7458.08s --> 7459.36s] : inject AI into it,\n",
"[7459.36s --> 7460.56s] : it's very tough.\n",
"[7460.56s --> 7461.60s] : or anything,\n",
"[7461.60s --> 7464.88s] : even like open so the the culture of open source?\n",
"[7464.88s --> 7466.32s] : Um,\n",
"[7466.32s --> 7467.04s] : like,\n",
"[7467.04s --> 7472.32s] : how how hard is it to walk into a room and be like the way we've been doing things are totally wrong?\n",
"[7472.32s --> 7472.80s] : Like,\n",
"[7472.80s --> 7476.96s] : I'm sure there's a lot of firing involved or a little like twisting of arms or something.\n",
"[7476.96s --> 7478.64s] : So do you have to rule by fear,\n",
"[7478.64s --> 7479.12s] : by love?\n",
"[7479.12s --> 7482.16s] : Like what can he say to the leadership aspect of this?\n",
"[7482.16s --> 7482.96s] : I mean,\n",
"[7482.96s --> 7484.56s] : he's just like an unbelievable job,\n",
"[7484.56s --> 7495.28s] : but he is amazing at being like clear and firm and getting people to want to come along,\n",
"[7495.28s --> 7500.88s] : but also like compassionate and patient with his people,\n",
"[7500.88s --> 7501.44s] : too.\n",
"[7501.44s --> 7503.60s] : I'm getting a lot of love,\n",
"[7503.60s --> 7504.40s] : not fear.\n",
"[7504.40s --> 7506.64s] : I'm a big Sao fan.\n",
"[7506.64s --> 7507.92s] : So am I,\n",
"[7507.92s --> 7509.04s] : from a distance.\n",
"[7509.04s --> 7509.60s] : I mean,\n",
"[7509.60s --> 7513.52s] : you have so much in your life trajectory that I can ask you about,\n",
"[7513.52s --> 7515.12s] : we can probably talk for many more hours.\n",
"[7515.12s --> 7518.88s] : but I got to ask you because of whycomminatated because of start ups and so on,\n",
"[7518.88s --> 7521.92s] : The recent a and you've tweeted about this,\n",
"[7521.92s --> 7522.32s] : uh,\n",
"[7522.32s --> 7524.48s] : about the Silicon Valley Bank.\n",
"[7524.48s --> 7525.92s] : SVB.\n",
"[7525.92s --> 7528.40s] : What's your best understanding of what happened?\n",
"[7528.40s --> 7532.80s] : What is interesting what is interesting to understand about what happened at SVB?\n",
"[7532.80s --> 7544.24s] : I think they just like horribly mismanaged buying while chasing returns in a very silly world of zero percent interest rates,\n",
"[7544.24s --> 7549.12s] : buuying very long dated instruments,\n",
"[7549.12s --> 7554.24s] : secured by very short term and variable deposits,\n",
"[7554.24s --> 7557.68s] : and this was obviously dumb.\n",
"[7557.68s --> 7564.64s] : I think totally the fault of the management team,\n",
"[7564.64s --> 7578.16s] : although I'm not sure what the regulators were thinking either and is an example of where I think you see the dangers of incentive misalignment,\n",
"[7578.16s --> 7583.84s] : because as the Fed kept raising,\n",
"[7583.84s --> 7592.96s] : I assume that the incentives on people working at SVB to not sell at a loss their,\n",
"[7592.96s --> 7593.20s] : you know,\n",
"[7593.20s --> 7597.68s] : super safe bonds which were now down twenty percent or whatever um or,\n",
"[7597.68s --> 7598.72s] : you know down less than that,\n",
"[7598.72s --> 7600.08s] : but then kept going down.\n",
"[7600.08s --> 7602.48s] : I you know,\n",
"[7602.48s --> 7605.52s] : that's like a classy example of incentive misalignment.\n",
"[7605.52s --> 7606.96s] : Now,\n",
"[7606.96s --> 7610.00s] : I suspect they're not the only bank in a bad position here.\n",
"[7610.00s --> 7619.12s] : The response of the Federal Government I think took much longer than it should have but by Sunday afternoon I was glad they had done what they've done.\n",
"[7619.12s --> 7621.28s] : We'll see what happens next.\n",
"[7621.28s --> 7624.80s] : So how do you avoid depositors from doubting their bank?\n",
"[7624.80s --> 7631.76s] : What I think needs would be good to do right now is just a and this requires statutory change,\n",
"[7631.76s --> 7637.12s] : but it may be a full guarantee of deposits may be a much much higher than two fifty k.\n",
"[7637.12s --> 7646.96s] : But you really don't want depositors having to doubt the security of their deposits,\n",
"[7646.96s --> 7649.36s] : and this thing that a lot of people on Twitter were saying is like,\n",
"[7649.36s --> 7649.52s] : Well,\n",
"[7649.52s --> 7651.04s] : it's their fault they should have been like,\n",
"[7651.04s --> 7651.36s] : you know,\n",
"[7651.36s --> 7654.96s] : reading the the the balance sheet and the risk audit of the bank like,\n",
"[7654.96s --> 7656.80s] : Do we really want people to have to do that?\n",
"[7656.80s --> 7657.52s] : I would argue,\n",
"[7657.52s --> 7658.24s] : No.\n",
"[7658.24s --> 7662.40s] : What impact has it had on start ups,\n",
"[7662.40s --> 7663.12s] : that you see?\n",
"[7663.12s --> 7663.52s] : Well,\n",
"[7663.52s --> 7664.88s] : there was a weekend of terror,\n",
"[7664.88s --> 7665.68s] : for sure.\n",
"[7665.68s --> 7667.12s] : And now I think,\n",
"[7667.12s --> 7668.64s] : even though it was only ten days ago,\n",
"[7668.64s --> 7669.76s] : it feels like forever,\n",
"[7669.76s --> 7670.88s] : and people have forgotten about it.\n",
"[7670.88s --> 7673.36s] : But it kind of reveals the fragility of ourr koaks.\n",
"[7673.36s --> 7674.40s] : It may not be done.\n",
"[7674.40s --> 7678.16s] : That may have been like the gun shown falling off the night stand in the first scene of the movie,\n",
"[7678.16s --> 7678.56s] : or whatever.\n",
"[7678.56s --> 7679.68s] : It could be like other banks.\n",
"[7679.68s --> 7681.52s] : There sure could we?\n",
"[7681.52s --> 7682.64s] : Well,\n",
"[7682.64s --> 7684.08s] : even with FTX I mean,\n",
"[7684.08s --> 7685.84s] : I'm just h,\n",
"[7685.84s --> 7686.64s] : well,\n",
"[7686.64s --> 7687.60s] : that's fraud,\n",
"[7687.60s --> 7689.68s] : but there's mismanagement.\n",
"[7689.68s --> 7693.44s] : and you wonder how stable our economic system is.\n",
"[7693.44s --> 7698.00s] : especiallycially with new entrants with AGI.\n",
"[7698.00s --> 7709.04s] : I think one of the many lessons to take away from this SBB thing is how fast and how much the world changes,\n",
"[7709.04s --> 7710.00s] : and how little,\n",
"[7710.00s --> 7710.48s] : I think,\n",
"[7710.48s --> 7712.40s] : our experts,\n",
"[7712.40s --> 7713.60s] : leaders,\n",
"[7713.60s --> 7714.80s] : business leaders,\n",
"[7714.80s --> 7715.44s] : regulators,\n",
"[7715.44s --> 7715.84s] : whatever,\n",
"[7715.84s --> 7716.56s] : understand it.\n",
"[7716.56s --> 7722.40s] : So the the speed with which the SBB bankro happened,\n",
"[7722.40s --> 7723.68s] : because of Twitter,\n",
"[7723.68s --> 7725.28s] : because of mobile banking apppse,\n",
"[7725.28s --> 7725.84s] : whatever,\n",
"[7725.84s --> 7728.48s] : was so different than the two thousand eight collapse,\n",
"[7728.48s --> 7731.20s] : where we didn't have those things really.\n",
"[7731.20s --> 7740.16s] : And I don't think that kind of the people in power realise how much the field had shifted,\n",
"[7740.16s --> 7746.00s] : and I think that is a very tiny preview of the shifts that AGI will bring.\n",
"[7746.00s --> 7751.04s] : What gives you hope in that shift from an economic perspective.\n",
"[7751.04s --> 7754.48s] : Ah It sounds scary the instability.\n",
"[7754.48s --> 7755.12s] : No,\n",
"[7755.12s --> 7758.88s] : I am nervous about the speed with with,\n",
"[7758.88s --> 7763.68s] : this changes and the speed with which our institutions can adapt,\n",
"[7763.68s --> 7764.40s] : U,\n",
"[7764.40s --> 7772.08s] : which is part of why we want to start deploying these systems really early why they're really weak so that people have as much time as possible to do this.\n",
"[7772.08s --> 7774.00s] : I think it's really scary to,\n",
"[7774.00s --> 7774.24s] : like,\n",
"[7774.24s --> 7775.12s] : have nothing,\n",
"[7775.12s --> 7775.44s] : nothing,\n",
"[7775.44s --> 7778.64s] : nothing and then drop a super powerful AGI at once on the world.\n",
"[7778.64s --> 7781.44s] : I don't think people should want that to happen.\n",
"[7781.44s --> 7786.72s] : But what gives me hope is like I think the less zeros the more positive sum the world gets the better,\n",
"[7786.72s --> 7789.60s] : and the the upside of the vision here.\n",
"[7789.60s --> 7791.84s] : just how much better life can be.\n",
"[7791.84s --> 7793.92s] : I think that's going to,\n",
"[7793.92s --> 7794.16s] : like,\n",
"[7794.16s --> 7795.76s] : unite a lot of us,\n",
"[7795.76s --> 7797.36s] : and even if it doesn't,\n",
"[7797.36s --> 7799.76s] : it's just going to make it all feel more positive some.\n",
"[7799.76s --> 7807.44s] : When you create an AGI system you'll be one of the few people in the room they get to interact with first,\n",
"[7807.44s --> 7810.48s] : assuming GPT fourour is not that.\n",
"[7810.48s --> 7811.44s] : Ah,\n",
"[7811.44s --> 7815.04s] : what question would you ask her him it?\n",
"[7815.04s --> 7817.04s] : What discussion would you have?\n",
"[7817.04s --> 7818.08s] : You know,\n",
"[7818.08s --> 7822.72s] : one of the things that I have realized like this is a little aside and not that important,\n",
"[7822.72s --> 7830.80s] : but I have never felt any pronoun other than it towards any of our systems.\n",
"[7830.80s --> 7840.72s] : but most other people say him or her or something like that and I wonder why I am so different.\n",
"[7840.72s --> 7841.04s] : Li,\n",
"[7841.04s --> 7841.28s] : yeah,\n",
"[7841.28s --> 7841.68s] : I't know,\n",
"[7841.68s --> 7842.96s] : maybe if I watch it develop,\n",
"[7842.96s --> 7844.32s] : maybe if I think more about it,\n",
"[7844.32s --> 7847.60s] : but I'm curious where that difference comes from.\n",
"[7847.60s --> 7849.92s] : I think probably you could because you watch it develop,\n",
"[7849.92s --> 7853.52s] : but then again I watch a lot of stuff develop and I always go to him or her.\n",
"[7853.52s --> 7857.76s] : I anthropomorphize aggressively,\n",
"[7857.76s --> 7858.80s] : um,\n",
"[7858.80s --> 7861.44s] : and certainly what most humans do.\n",
"[7861.44s --> 7869.68s] : I think it's really important that we try to explain to educate people that this is a tool and not a creature.\n",
"[7869.68s --> 7873.12s] : I think I yes,\n",
"[7873.12s --> 7876.56s] : but I also think there will be a room in society for creatures,\n",
"[7876.56s --> 7879.12s] : and we should draw hard lines between those.\n",
"[7879.12s --> 7883.76s] : If some's a creature Im happy for people to like think of it and talk about it as a creature,\n",
"[7883.76s --> 7887.36s] : but I think it is dangerous to project creatureness on to a tool.\n",
"[7887.36s --> 7892.56s] : That one perspective?\n",
"[7892.56s --> 7896.48s] : A perspective I would take if is done transparently,\n",
"[7896.48s --> 7902.24s] : is projecting creatureness onto a tool makes that tool more usable.\n",
"[7902.24s --> 7909.52s] : If it is done well Yes So if there's like kind of UI affordances that workk,\n",
"[7909.52s --> 7911.44s] : I understand that I still think we want to be,\n",
"[7911.44s --> 7913.20s] : like pretty careful with it.\n",
"[7913.20s --> 7915.60s] : Because the more creature like it is,\n",
"[7915.60s --> 7918.16s] : the more can manipulate you emotion.\n",
"[7918.16s --> 7921.92s] : orr just the more you think that it's doing something,\n",
"[7921.92s --> 7923.44s] : or should be able to do something,\n",
"[7923.44s --> 7926.64s] : or rely on it for something that it's not capable of?\n",
"[7926.64s --> 7928.80s] : What if it is capable?\n",
"[7928.80s --> 7931.20s] : What about Sam Alman,\n",
"[7931.20s --> 7933.12s] : what if it's capable of love?\n",
"[7933.12s --> 7939.28s] : Do you think there will be romantic relationships like in the movie Her with GpT?\n",
"[7939.28s --> 7946.40s] : There are companies now that offer like for lack of a better word,\n",
"[7946.40s --> 7949.36s] : like romantic Comanionship As.\n",
"[7949.36s --> 7952.48s] : Replica is an example of such a company.\n",
"[7952.48s --> 7953.04s] : Yeah.\n",
"[7953.04s --> 7958.16s] : I personally don't feel Any interest in that.\n",
"[7958.16s --> 7961.04s] : So you're focusing on creating intelligence,\n",
"[7961.04s --> 7963.12s] : but I understand why other people do.\n",
"[7963.12s --> 7964.72s] : That's interesting.\n",
"[7964.72s --> 7967.92s] : I'm I've for some reason been very drawn to that.\n",
"[7967.92s --> 7970.56s] : Have you spent a lot of time interacting with replica or anything?\n",
"[7970.56s --> 7971.60s] : Someilar replica,\n",
"[7971.60s --> 7975.04s] : but also just building stuff myself.ke I have robot dogs now that I,\n",
"[7975.04s --> 7975.52s] : uh,\n",
"[7975.52s --> 7981.60s] : use um I use the movement of the the the robots to communicate a motion.\n",
"[7981.60s --> 7984.16s] : I've been exploring how to do that.\n",
"[7984.16s --> 7984.96s] : Look,\n",
"[7984.96s --> 7993.52s] : there are going to be very interactive GPT four powered pets or whatever,\n",
"[7993.52s --> 7995.20s] : robots,\n",
"[7995.20s --> 8001.68s] : companions and a lot of people seem really excited about that.\n",
"[8001.68s --> 8002.08s] : Yeah,\n",
"[8002.08s --> 8004.00s] : there's a lot of interesting possibilities,\n",
"[8004.00s --> 8006.64s] : I think You'll discover them,\n",
"[8006.64s --> 8006.96s] : I think,\n",
"[8006.96s --> 8007.76s] : as you go along.\n",
"[8007.76s --> 8008.80s] : That's the whole point.\n",
"[8008.80s --> 8013.84s] : Like the things you say in this conversation you might in a year say This was right.\n",
"[8013.84s --> 8014.16s] : No,\n",
"[8014.16s --> 8015.28s] : I totally want.\n",
"[8015.28s --> 8018.24s] : I may turn out that I like love my GPD Ford.\n",
"[8018.24s --> 8020.24s] : Maybe your robot or whatever.\n",
"[8020.24s --> 8024.32s] : Maybe you want your programming assistant to be a little kinder and not mock you.\n",
"[8024.32s --> 8025.84s] : I dear incomtent.\n",
"[8025.84s --> 8026.24s] : No,\n",
"[8026.24s --> 8032.56s] : I think you do want um The style of the way GPT four talks to you Yes,\n",
"[8032.56s --> 8033.52s] : really matters.\n",
"[8033.52s --> 8035.44s] : You probably want something different than what I want,\n",
"[8035.44s --> 8038.64s] : but we both probably want something different than the current GPT four.\n",
"[8038.64s --> 8040.40s] : And that will be really important,\n",
"[8040.40s --> 8042.56s] : even for a very tool like thing.\n",
"[8042.56s --> 8044.80s] : Is there styles of conversation o,\n",
"[8044.80s --> 8045.04s] : no,\n",
"[8045.04s --> 8048.80s] : contents of conversations you're looking forward to with an AGI,\n",
"[8048.80s --> 8051.04s] : like GPT five,\n",
"[8051.04s --> 8051.44s] : six,\n",
"[8051.44s --> 8051.92s] : seven.\n",
"[8051.92s --> 8059.76s] : Is there stuff where like where do you go to out outside of the fun meme stuff,\n",
"[8059.76s --> 8060.80s] : for actual I mean,\n",
"[8060.80s --> 8065.28s] : what I'm excited for is like Please explain to me how all the physics works,\n",
"[8065.28s --> 8067.20s] : and solve all remaining mysteries.\n",
"[8067.20s --> 8070.32s] : So like a theory of everything I'll be real happy.\n",
"[8070.32s --> 8073.68s] : Hm Faster than light travel.\n",
"[8073.68s --> 8075.12s] : Don't you want to know?\n",
"[8075.12s --> 8079.52s] : So there's several things to know and be hard.\n",
"[8079.52s --> 8083.12s] : U is it possible and how to do it?\n",
"[8083.12s --> 8084.64s] : U Yeah,\n",
"[8084.64s --> 8085.28s] : I want to know.\n",
"[8085.28s --> 8085.84s] : I want to know.\n",
"[8085.84s --> 8087.20s] : Probably the first question would be,\n",
"[8087.20s --> 8090.16s] : are there other intelligent alien civilizations out there?\n",
"[8090.16s --> 8095.28s] : But I don't think AGI has the the ability to do that to know that.\n",
"[8095.28s --> 8098.56s] : Amight be able to help us figure out how to go detect?\n",
"[8098.56s --> 8103.28s] : In mean need to like send some emails to humans and say can you run these experiments?\n",
"[8103.28s --> 8104.64s] : Can you build the space probe?\n",
"[8104.64s --> 8105.28s] : Can you wait,\n",
"[8105.28s --> 8105.68s] : you know,\n",
"[8105.68s --> 8106.56s] : a very long time?\n",
"[8106.56s --> 8109.36s] : or provide a much better estimate than the Drake equation?\n",
"[8109.36s --> 8111.92s] : with the knowledge we already have.\n",
"[8111.92s --> 8115.28s] : and maybe process all the 'ca we've been collecting a lot of.\n",
"[8115.28s --> 8115.52s] : ,\n",
"[8115.52s --> 8116.00s] : ,\n",
"[8116.00s --> 8117.04s] : maybe it's in the data.\n",
"[8117.04s --> 8118.96s] : Maybe we need to build better detectors,\n",
"[8118.96s --> 8121.68s] : which it really advanced I can tell us how to do.\n",
"[8121.68s --> 8123.84s] : It may not be able to answer it on its own,\n",
"[8123.84s --> 8126.40s] : but it may be able to tell us what to go build,\n",
"[8126.40s --> 8127.68s] : to collect more data.\n",
"[8127.68s --> 8130.32s] : What if it says the aliens are already here?\n",
"[8130.32s --> 8132.80s] : I think I would just go about my life,\n",
"[8132.80s --> 8133.76s] : ep.\n",
"[8133.76s --> 8136.00s] : I 'cause I mean,\n",
"[8136.00s --> 8137.28s] : a version of that is like,\n",
"[8137.28s --> 8139.04s] : what are you doing differently now that,\n",
"[8139.04s --> 8141.76s] : like if if GPT four told you and you believed it,\n",
"[8141.76s --> 8142.08s] : Ooka,\n",
"[8142.08s --> 8145.76s] : AGI is here Or AGIs coming real soon.\n",
"[8145.76s --> 8147.60s] : What are you going to do differently?\n",
"[8147.60s --> 8151.36s] : The source of joy and happiness of fulfillment of life is from other humans,\n",
"[8151.36s --> 8153.92s] : so it's mostly nothing,\n",
"[8153.92s --> 8156.56s] : unless it causes some kind of threat,\n",
"[8156.56s --> 8159.76s] : but that threat would have to be literally a fire.\n",
"[8159.76s --> 8160.24s] : Like,\n",
"[8160.24s --> 8166.24s] : are we living now with a greater degree of digital intelligence than you would have expected three years ago?\n",
"[8166.24s --> 8171.36s] : And if you could go back and be told by an oracle three years ago,\n",
"[8171.36s --> 8171.68s] : which is,\n",
"[8171.68s --> 8171.92s] : you know,\n",
"[8171.92s --> 8180.16s] : blink of an eye that in March of twenty twenty three you will be living with this degree of digital intelligence.\n",
"[8180.16s --> 8183.44s] : Would you expect your life to be more different than it is right now?\n",
"[8183.44s --> 8186.72s] : Probably,\n",
"[8186.72s --> 8187.52s] : probably.\n",
"[8187.52s --> 8190.16s] : but there's also a lot of different trajetures intermixed.\n",
"[8190.16s --> 8196.24s] : I would have expected the society's response to a pandemic aer,\n",
"[8196.24s --> 8197.84s] : to be much better,\n",
"[8197.84s --> 8199.52s] : much clearer,\n",
"[8199.52s --> 8201.04s] : less divided.\n",
"[8201.04s --> 8202.56s] : I was very confused about.\n",
"[8202.56s --> 8204.00s] : there's a lot of stuff,\n",
"[8204.00s --> 8207.04s] : given the amazing technological advancements that are happening,\n",
"[8207.04s --> 8209.20s] : the weird social divisions.\n",
"[8209.20s --> 8214.80s] : It's almost like the more technological advancement there is the more we're going to be having fun with social division.\n",
"[8214.80s --> 8218.72s] : Or maybe the technological advancement just revealed the division that was already there,\n",
"[8218.72s --> 8225.68s] : but all of that just that confuses my understanding of how far along we are as a human civilization,\n",
"[8225.68s --> 8231.20s] : and what brings us meaning and how we discover truth together and knowledge and wisdom.\n",
"[8231.20s --> 8233.12s] : So I don't know,\n",
"[8233.12s --> 8236.80s] : but when I look when I open Wikipedia,\n",
"[8236.80s --> 8239.92s] : I'm happy that humans are able to create this thing.\n",
"[8239.92s --> 8242.16s] : First yes there is bias Yes,\n",
"[8242.16s --> 8245.84s] : let's sing It's a triumphal It's the triumph of human civilization.\n",
"[8245.84s --> 8247.84s] : hundred per cent Ah Google search,\n",
"[8247.84s --> 8248.40s] : the search,\n",
"[8248.40s --> 8248.96s] : search,\n",
"[8248.96s --> 8250.64s] : period is incredible,\n",
"[8250.64s --> 8253.36s] : the way it was able to do twenty years ago.\n",
"[8253.36s --> 8254.88s] : And and now this,\n",
"[8254.88s --> 8257.44s] : this is this new thing GPT.\n",
"[8257.44s --> 8263.84s] : like is this like going to be the next like the conglomeration of all that that made,\n",
"[8263.84s --> 8264.24s] : uh,\n",
"[8264.24s --> 8268.24s] : Websearch and Wikipedia so magical,\n",
"[8268.24s --> 8270.24s] : but now more directly accessible,\n",
"[8270.24s --> 8272.48s] : you can have a conversation with the damned thing.\n",
"[8272.48s --> 8274.00s] : It's incredible.\n",
"[8274.00s --> 8277.44s] : Let me ask you for advice for young people,\n",
"[8277.44s --> 8280.32s] : in high school and college what to do with their life.\n",
"[8280.32s --> 8282.64s] : how to have a career they can be proud of,\n",
"[8282.64s --> 8284.56s] : how to have a life they can be proud of.\n",
"[8284.56s --> 8285.36s] : Ah,\n",
"[8285.36s --> 8289.68s] : you wrote a blog post a few years ago titled how to be successful,\n",
"[8289.68s --> 8291.60s] : and there's a bunch of really,\n",
"[8291.60s --> 8293.52s] : really people should check out that blog posts.\n",
"[8293.52s --> 8297.44s] : They're so it's so succinct ands so brilliant.\n",
"[8297.44s --> 8299.20s] : You have a bunch of bullet points.\n",
"[8299.20s --> 8300.88s] : Compound yourself,\n",
"[8300.88s --> 8302.88s] : have almost too much self belief,\n",
"[8302.88s --> 8304.24s] : learn to think independently,\n",
"[8304.24s --> 8306.24s] : get good at sales and quotes,\n",
"[8306.24s --> 8307.68s] : make it easy to take risks,\n",
"[8307.68s --> 8308.24s] : focus,\n",
"[8308.24s --> 8310.08s] : work hard as we talked about,\n",
"[8310.08s --> 8310.88s] : be bold,\n",
"[8310.88s --> 8311.76s] : be wilful,\n",
"[8311.76s --> 8313.04s] : be hard to compete with,\n",
"[8313.04s --> 8314.40s] : build a network.\n",
"[8314.40s --> 8316.64s] : You get rich by owning things,\n",
"[8316.64s --> 8318.08s] : be internally driven.\n",
"[8318.08s --> 8319.84s] : What stands out to you,\n",
"[8319.84s --> 8321.52s] : from that or beyond,\n",
"[8321.52s --> 8323.12s] : is the advice you can give?\n",
"[8323.12s --> 8323.76s] : Yet,\n",
"[8323.76s --> 8327.68s] : no I think it is like good advice in some sense,\n",
"[8327.68s --> 8335.52s] : but I also think it's way too tempting to take advice from other people,\n",
"[8335.52s --> 8339.60s] : and the stuff that worked for me which I tried to write down there,\n",
"[8339.60s --> 8343.68s] : probably doesn't work that well or may not work as well for other people,\n",
"[8343.68s --> 8344.24s] : or,\n",
"[8344.24s --> 8345.28s] : like other people,\n",
"[8345.28s --> 8350.56s] : may find out that they want to just have a super different life trajectory,\n",
"[8350.56s --> 8357.20s] : and I think I mostly got what I wanted by ignoring advice.\n",
"[8357.20s --> 8361.76s] : And I think like I tell people not to listen to too much advice.\n",
"[8361.76s --> 8367.68s] : Listening to advice from other people should be approached with great caution.\n",
"[8367.68s --> 8371.28s] : How would you describe how you've approached life?\n",
"[8371.28s --> 8377.92s] : Outside of this advice that you would advise to other people.\n",
"[8377.92s --> 8383.68s] : so really just in the quiet of your mind to think what gives me happiness,\n",
"[8383.68s --> 8385.20s] : what is the right thing to do here,\n",
"[8385.20s --> 8387.20s] : how can I have the most impact?\n",
"[8387.20s --> 8390.64s] : I wish it were that,\n",
"[8390.64s --> 8391.52s] : you know,\n",
"[8391.52s --> 8393.68s] : introspective all the time.\n",
"[8393.68s --> 8396.08s] : It's a lot of just like you know,\n",
"[8396.08s --> 8398.88s] : what will bring me joy will bring me fulfillment.\n",
"[8398.88s --> 8401.52s] : You know what will bring what will be?\n",
"[8401.52s --> 8402.00s] : Ah,\n",
"[8402.00s --> 8404.88s] : I I do think a lot about what I can do that will be useful,\n",
"[8404.88s --> 8407.28s] : but like who do I want to spend my time with?\n",
"[8407.28s --> 8408.96s] : What do I want to spend my time doing?\n",
"[8408.96s --> 8417.36s] : The fish and water's going wrong with the That's certainly what it feels like I mean I think that's what most people would say if they were really honest about it.\n",
"[8417.36s --> 8418.00s] : Yeah,\n",
"[8418.00s --> 8420.96s] : If they really think yeah.\n",
"[8420.96s --> 8425.68s] : And some of that then gets to the Sam Harris discussion of free well being an illusion.\n",
"[8425.68s --> 8427.36s] : Of course very well might be,\n",
"[8427.36s --> 8432.24s] : which is a really complicated thing to wrap your head around.\n",
"[8432.24s --> 8436.00s] : What do you think's the meaning of this whole thing?\n",
"[8436.00s --> 8438.40s] : That's a question you can ask,\n",
"[8438.40s --> 8439.36s] : an AGI.\n",
"[8439.36s --> 8440.96s] : What's the meaning of life?\n",
"[8440.96s --> 8443.04s] : As far as you look at it,\n",
"[8443.04s --> 8449.28s] : you're part of a small group of people that are creating something truly special,\n",
"[8449.28s --> 8454.96s] : something that feels like almost feels like humanity was always moving towards.\n",
"[8454.96s --> 8455.44s] : Yeah.\n",
"[8455.44s --> 8458.00s] : That's what I was going to say is I don't think its a small group of people.\n",
"[8458.00s --> 8465.04s] : I think this is the I think this is like the product of the culmination of whatever you want to call it,\n",
"[8465.04s --> 8468.40s] : an amazing amount of human effort.\n",
"[8468.40s --> 8476.56s] : if you think about everything that had to come together for this to happen When those people discovered the transistor in the forties,\n",
"[8476.56s --> 8478.16s] : like is this what they were planning on?\n",
"[8478.16s --> 8490.96s] : all of the work the hundreds of thousands of millions of people to ever it's been that it took to go from that one first transistor to packing the numbers we do into a chip and figuring out how to wire them all up together,\n",
"[8490.96s --> 8493.44s] : and everything else that goes into this.\n",
"[8493.44s --> 8494.00s] : know,\n",
"[8494.00s --> 8497.68s] : the energy required the the the the science at like,\n",
"[8497.68s --> 8499.04s] : just every every step.\n",
"[8499.04s --> 8502.00s] : like this is the output of,\n",
"[8502.00s --> 8502.48s] : like,\n",
"[8502.48s --> 8503.92s] : all of us.\n",
"[8503.92s --> 8506.16s] : And I think that's pretty cool.\n",
"[8506.16s --> 8512.16s] : And before the transistor there was a hundred billion people who lived and died,\n",
"[8512.16s --> 8513.60s] : had sex,\n",
"[8513.60s --> 8514.72s] : fell love,\n",
"[8514.72s --> 8516.32s] : ate a lot of good food,\n",
"[8516.32s --> 8518.64s] : murdered each other sometimes rarely,\n",
"[8518.64s --> 8520.80s] : but mostly of juwies good to each other,\n",
"[8520.80s --> 8521.84s] : struggled to survive,\n",
"[8521.84s --> 8526.00s] : and before that there was bacteria and eukaryates and all that,\n",
"[8526.00s --> 8528.80s] : And all of that was on this one exponential curve.\n",
"[8528.80s --> 8529.68s] : Yeah,\n",
"[8529.68s --> 8531.04s] : how many others are there,\n",
"[8531.04s --> 8531.84s] : I wonder?\n",
"[8531.84s --> 8535.12s] : We will ask that isn't the question number one for me for Agiaa.\n",
"[8535.12s --> 8536.24s] : How many others?\n",
"[8536.24s --> 8539.36s] : And I'm not sure which answer I want to hear.\n",
"[8539.36s --> 8539.84s] : Sam,\n",
"[8539.84s --> 8541.20s] : you're an incredible person.\n",
"[8541.20s --> 8542.72s] : it's an honor to talk to you.\n",
"[8542.72s --> 8544.08s] : Thank you for the work you're doing.\n",
"[8544.08s --> 8544.64s] : Like I said,\n",
"[8544.64s --> 8545.68s] : I've talked to Ilias,\n",
"[8545.68s --> 8546.08s] : the Scare,\n",
"[8546.08s --> 8546.80s] : I talk to Gregg,\n",
"[8546.80s --> 8548.72s] : I talked to so many people at open AI.\n",
"[8548.72s --> 8550.00s] : They're really good people.\n",
"[8550.00s --> 8551.84s] : They're doing really interesting work.\n",
"[8551.84s --> 8555.60s] : We are going to try our hardest to get to get to a good place here.\n",
"[8555.60s --> 8558.08s] : I think the challenges are Tough.\n",
"[8558.08s --> 8564.40s] : I understand that not everyone agrees with our approach of iterative deployment and also iterative discovery.\n",
"[8564.40s --> 8565.04s] : Um,\n",
"[8565.04s --> 8566.56s] : but it's what we believe in.\n",
"[8566.56s --> 8566.96s] : Ah,\n",
"[8566.96s --> 8569.04s] : I think we're making good progress.\n",
"[8569.04s --> 8572.32s] : And I think the pace is fast.\n",
"[8572.32s --> 8574.16s] : But so is the progress.\n",
"[8574.16s --> 8578.24s] : So so like the pace of capabilities and change is fast.\n",
"[8578.24s --> 8585.68s] : but I think that also means we will have new tools to figure out alignment and sort of the capitalpit S safety problem.\n",
"[8585.68s --> 8590.72s] : I feel like we're in this together I can't wait what we together as of human civization come up with.\n",
"[8590.72s --> 8591.68s] : It's going to be great,\n",
"[8591.68s --> 8592.00s] : I think.\n",
"[8592.00s --> 8593.20s] : We'll work really hard to me,\n",
"[8593.20s --> 8593.76s] : too.\n",
"[8593.76s --> 8596.72s] : Thanks for listening to this conversation with Sam Altman.\n",
"[8596.72s --> 8597.92s] : To support this podgass,\n",
"[8597.92s --> 8600.40s] : please check out our sponsors in the description.\n",
"[8600.40s --> 8606.00s] : And now let me leave you with some words from Allan Turring in nineteen fifty one.\n",
"[8606.00s --> 8611.60s] : It seems probable that once the machine thinking method has started,\n",
"[8611.60s --> 8616.16s] : it would not take long to outstrip our feeble powers.\n",
"[8616.16s --> 8617.68s] : At some stage,\n",
"[8617.68s --> 8618.40s] : therefore,\n",
"[8618.40s --> 8622.80s] : we should have to expect the machines to take control.\n",
"[8622.80s --> 8625.44s] : Thank you for listening,\n",
"[8625.44s --> 8627.28s] : and hope to see you next time\n",
"time: 55.2 ms (started: 2023-10-19 19:07:05 -07:00)\n"
]
}
],
"source": [
"import re\n",
"\n",
"hyp = hypotheses[0]\n",
"word_timestamps = hyp.timestep['word']\n",
"\n",
"def calculate_merged_timestamps(timestamps):\n",
" buffer = [[]]\n",
" stamps = [[None, None]]\n",
" for data in timestamps:\n",
" if data['word'] in [',', '.', '?']:\n",
" buffer[-1].append(data['word'] + \"\\n\")\n",
" stamps[-1][1] = data['end_offset']\n",
"\n",
" if stamps[-1][0] is None:\n",
" stamps[-1][0] = stamps[-2][1] # end of last stamp is begin of current stamp\n",
" \n",
" buffer.append([])\n",
" stamps.append([None, None])\n",
" else:\n",
" buffer[-1].append(data['word'])\n",
" if stamps[-1][0] == None:\n",
" stamps[-1][0] = data['start_offset']\n",
"\n",
" stamps[0][0] = timestamps[0]['start_offset']\n",
" stamps[-1][1] = timestamps[-1]['end_offset']\n",
" return buffer, stamps\n",
"\n",
"def print_merged_timestamps(buffer, stamps, model_stride=8, window_stride=0.01):\n",
" assert len(buffer) == len(stamps)\n",
"\n",
" for data, stamp in zip(buffer, stamps):\n",
" text = \" \".join(data)\n",
" if text.endswith(\"\\n\"):\n",
" text = text[:-1]\n",
"\n",
" # regex replacement to squash spaces between special tokens and words\n",
" text = re.sub(r'(\\s+)([\\.\\,\\?])', r'\\2', text)\n",
" \n",
" resolved_start = stamp[0] * model_stride * window_stride\n",
" resolved_end = stamp[1] * model_stride * window_stride\n",
" \n",
" timestamp = f\"[{resolved_start:0.2f}s --> {resolved_end:0.2f}s]\"\n",
" print(f\"{timestamp:<20} : {text}\")\n",
"\n",
"\n",
"buffer, stamps = calculate_merged_timestamps(word_timestamps)\n",
"print_merged_timestamps(buffer, stamps)"
]
},
{
"cell_type": "markdown",
"id": "6a28da52-e45b-4c95-a101-f1efdc2623b7",
"metadata": {},
"source": [
"# Forced alignment"
]
},
{
"cell_type": "code",
"execution_count": 35,
"id": "cdef5e22-55c0-4306-91b2-371b7ffa1969",
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"time: 746 µs (started: 2023-10-19 18:36:25 -07:00)\n"
]
}
],
"source": [
"if not os.path.exists(\"nemo_forced_aligner\"):\n",
" !git clone --depth 1 https://github.com/NVIDIA/NeMo.git\n",
" !mv NeMo/tools/nemo_forced_aligner .\n",
" !rm -rf ./NeMo/"
]
},
{
"cell_type": "code",
"execution_count": 38,
"id": "1fea5159-a5d2-448b-ba45-7aaa574be034",
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"time: 4.79 ms (started: 2023-10-19 18:40:28 -07:00)\n"
]
}
],
"source": [
"# Create temp manifest file\n",
"import json\n",
"\n",
"with open('nfa_manifest.json', 'w', encoding='utf8') as f:\n",
" data = {'audio_filepath': './sam_altman_lex_podcast_367.wav', 'text': rnnt_transcript}\n",
" f.write(f\"{json.dumps(data)}\\n\")"
]
},
{
"cell_type": "code",
"execution_count": 39,
"id": "b9ce93d3-9f99-4042-bfc7-e0a56d028775",
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"[NeMo W 2023-10-19 18:40:34 nemo_logging:349] /home/smajumdar/miniconda3/envs/NeMo/lib/python3.10/site-packages/hydra/_internal/hydra.py:119: UserWarning: Future Hydra versions will no longer change working directory at job runtime by default.\n",
" See https://hydra.cc/docs/1.2/upgrades/1.1_to_1.2/changes_to_job_working_dir/ for more information.\n",
" ret = run_job(\n",
" \n",
"[NeMo I 2023-10-19 18:40:34 align:159] Hydra config: pretrained_name: nvidia/stt_en_fastconformer_hybrid_large_pc\n",
" model_path: null\n",
" manifest_filepath: nfa_manifest.json\n",
" output_dir: nfa_output\n",
" align_using_pred_text: false\n",
" transcribe_device: null\n",
" viterbi_device: null\n",
" batch_size: 1\n",
" use_local_attention: true\n",
" additional_segment_grouping_separator: .\n",
" audio_filepath_parts_in_utt_id: 1\n",
" use_buffered_chunked_streaming: false\n",
" chunk_len_in_secs: 1.6\n",
" total_buffer_in_secs: 4.0\n",
" chunk_batch_size: 32\n",
" simulate_cache_aware_streaming: false\n",
" save_output_file_formats:\n",
" - ctm\n",
" - ass\n",
" ctm_file_config:\n",
" remove_blank_tokens: true\n",
" minimum_timestamp_duration: 0.0\n",
" ass_file_config:\n",
" fontsize: 20\n",
" vertical_alignment: center\n",
" resegment_text_to_fill_space: false\n",
" max_lines_per_segment: 2\n",
" text_already_spoken_rgb:\n",
" - 49\n",
" - 46\n",
" - 61\n",
" text_being_spoken_rgb:\n",
" - 57\n",
" - 171\n",
" - 9\n",
" text_not_yet_spoken_rgb:\n",
" - 194\n",
" - 193\n",
" - 199\n",
" \n",
"[NeMo I 2023-10-19 18:40:34 align:228] Device to be used for transcription step (`transcribe_device`) is cuda\n",
"[NeMo I 2023-10-19 18:40:34 align:234] Device to be used for viterbi step (`viterbi_device`) is cuda\n",
"[NeMo W 2023-10-19 18:40:34 align:237] One or both of transcribe_device and viterbi_device are GPUs. If you run into OOM errors it may help to change both devices to be the CPU.\n",
"[NeMo I 2023-10-19 18:40:36 mixins:170] Tokenizer SentencePieceTokenizer initialized with 1024 tokens\n",
"[NeMo W 2023-10-19 18:40:37 modelPT:161] If you intend to do training or fine-tuning, please call the ModelPT.setup_training_data() method and provide a valid configuration file to setup the train data loader.\n",
" Train config : \n",
" manifest_filepath: null\n",
" sample_rate: 16000\n",
" batch_size: 32\n",
" shuffle: true\n",
" num_workers: 8\n",
" pin_memory: true\n",
" use_start_end_token: false\n",
" trim_silence: false\n",
" max_duration: 20\n",
" min_duration: 0.1\n",
" is_tarred: false\n",
" tarred_audio_filepaths: null\n",
" shuffle_n: 2048\n",
" bucketing_strategy: fully_randomized\n",
" bucketing_batch_size: null\n",
" is_concat: false\n",
" concat_sampling_technique: random\n",
" concat_sampling_probabilities: ''\n",
" shard_manifests: true\n",
" defer_setup: true\n",
" \n",
"[NeMo W 2023-10-19 18:40:37 modelPT:168] If you intend to do validation, please call the ModelPT.setup_validation_data() or ModelPT.setup_multiple_validation_data() method and provide a valid configuration file to setup the validation data loader(s). \n",
" Validation config : \n",
" manifest_filepath: null\n",
" sample_rate: 16000\n",
" batch_size: 32\n",
" shuffle: false\n",
" num_workers: 8\n",
" pin_memory: true\n",
" use_start_end_token: false\n",
" is_concat: true\n",
" concat_sampling_technique: random\n",
" concat_sampling_probabilities:\n",
" - 0.1162666847815\n",
" - 0.1152159432515\n",
" - 0.12292525035150001\n",
" - 0.1188495893015\n",
" - 0.1384838395615\n",
" - 0.1510209186815\n",
" - 0.1354274550315\n",
" - 0.10181031905150001\n",
" concat_shuffle: false\n",
" concat_sampling_seed: 1234\n",
" max_duration: 20\n",
" \n",
"[NeMo W 2023-10-19 18:40:37 modelPT:174] Please call the ModelPT.setup_test_data() or ModelPT.setup_multiple_test_data() method and provide a valid configuration file to setup the test data loader(s).\n",
" Test config : \n",
" manifest_filepath: null\n",
" sample_rate: 16000\n",
" batch_size: 16\n",
" shuffle: false\n",
" num_workers: 8\n",
" pin_memory: true\n",
" use_start_end_token: false\n",
" \n",
"[NeMo I 2023-10-19 18:40:37 features:289] PADDING: 0\n",
"[NeMo W 2023-10-19 18:40:38 nemo_logging:349] /home/smajumdar/miniconda3/envs/NeMo/lib/python3.10/site-packages/torch/nn/modules/rnn.py:71: UserWarning: dropout option adds dropout after all but last recurrent layer, so non-zero dropout expects num_layers greater than 1, but got dropout=0.2 and num_layers=1\n",
" warnings.warn(\"dropout option adds dropout after all but last \"\n",
" \n",
"[NeMo I 2023-10-19 18:40:38 rnnt_models:211] Using RNNT Loss : warprnnt_numba\n",
" Loss warprnnt_numba_kwargs: {'fastemit_lambda': 0.0, 'clamp': -1.0}\n",
"[NeMo I 2023-10-19 18:40:39 save_restore_connector:249] Model EncDecHybridRNNTCTCBPEModel was successfully restored from /home/smajumdar/.cache/huggingface/hub/models--nvidia--stt_en_fastconformer_hybrid_large_pc/snapshots/453047c998bce94f2acb945d171a7ec603aff1c1/stt_en_fastconformer_hybrid_large_pc.nemo.\n",
"[NeMo I 2023-10-19 18:40:39 hybrid_rnnt_ctc_bpe_models:432] No `decoding_cfg` passed when changing decoding strategy, using internal config\n",
"[NeMo I 2023-10-19 18:40:39 hybrid_rnnt_ctc_bpe_models:456] Changed decoding strategy of the CTC decoder to \n",
" strategy: greedy\n",
" preserve_alignments: null\n",
" compute_timestamps: null\n",
" word_seperator: ' '\n",
" ctc_timestamp_type: all\n",
" batch_dim_index: 0\n",
" greedy:\n",
" preserve_alignments: false\n",
" compute_timestamps: false\n",
" preserve_frame_confidence: false\n",
" confidence_measure_cfg:\n",
" name: entropy\n",
" entropy_type: tsallis\n",
" alpha: 0.33\n",
" entropy_norm: exp\n",
" temperature: DEPRECATED\n",
" confidence_method_cfg: DEPRECATED\n",
" beam:\n",
" beam_size: 4\n",
" search_type: default\n",
" preserve_alignments: false\n",
" compute_timestamps: false\n",
" return_best_hypothesis: true\n",
" beam_alpha: 1.0\n",
" beam_beta: 0.0\n",
" kenlm_path: null\n",
" flashlight_cfg:\n",
" lexicon_path: null\n",
" boost_path: null\n",
" beam_size_token: 16\n",
" beam_threshold: 20.0\n",
" unk_weight: -.inf\n",
" sil_weight: 0.0\n",
" pyctcdecode_cfg:\n",
" beam_prune_logp: -10.0\n",
" token_min_logp: -5.0\n",
" prune_history: false\n",
" hotwords: null\n",
" hotword_weight: 10.0\n",
" confidence_cfg:\n",
" preserve_frame_confidence: false\n",
" preserve_token_confidence: false\n",
" preserve_word_confidence: false\n",
" exclude_blank: true\n",
" aggregation: min\n",
" measure_cfg:\n",
" name: entropy\n",
" entropy_type: tsallis\n",
" alpha: 0.33\n",
" entropy_norm: exp\n",
" temperature: DEPRECATED\n",
" method_cfg: DEPRECATED\n",
" temperature: 1.0\n",
" \n",
"[NeMo I 2023-10-19 18:40:39 align:250] Flag use_local_attention is set to True => will try to use local attention for model if it allows it\n",
"Transcribing: 100%|███████████████████████████████| 1/1 [00:03<00:00, 3.51s/it]\n",
"[NeMo I 2023-10-19 18:40:43 data_prep:828] Calculated that the model downsample factor is 8 and therefore the ASR model output timestep duration is 0.08 -- will use this for all batches\n",
"time: 41.8 s (started: 2023-10-19 18:40:30 -07:00)\n"
]
}
],
"source": [
"!python nemo_forced_aligner/align.py \\\n",
" pretrained_name=$model_name \\\n",
" manifest_filepath=\"nfa_manifest.json\" \\\n",
" output_dir=\"nfa_output\" \\\n",
" use_local_attention=True \\\n",
" additional_segment_grouping_separator=\".\" \\\n",
" ctm_file_config.remove_blank_tokens=True "
]
},
{
"cell_type": "code",
"execution_count": 58,
"id": "847338f4-1c1f-4474-b415-2aee370f2d61",
"metadata": {
"scrolled": true
},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"[0.16s --> 3.84s] : We have been a misunderstood and badly mocked Org for a long time,\n",
"[3.92s --> 15.52s] : like when we started We like announced the Org at the end of twenty fifteen said we were going to work on AGI Like people thought we were batchet insane You know,\n",
"[15.60s --> 27.76s] : like I I remember at the time an eminent AI scientist at a large industrial AI lab was like demming individual reporters being like,\n",
"[27.92s --> 28.24s] : you know,\n",
"[28.32s --> 33.20s] : these people aren't very good and it's ridiculous to talk about AGI and I can't believe you're giving them time of day,\n",
"[33.20s --> 34.96s] : and it's like that was the level of,\n",
"[34.96s --> 35.28s] : like,\n",
"[35.44s --> 59.68s] : pettiness and rancor in the field that a new group of people saying we're going to try to build AGI So open AI and deep mind was a small collection of folks who were brave enough to talk about AGI um in the face of mockery We don't get mocked as much now Don't get mocked as much now The following is a conversation with Sam Aldman,\n",
"[59.76s --> 61.60s] : CEO of OpenAI,\n",
"[61.76s --> 63.76s] : the Company behind GPT four,\n",
"[63.84s --> 65.12s] : JaGPT,\n",
"[65.20s --> 65.84s] : Dolly,\n",
"[65.92s --> 66.80s] : Codex,\n",
"[66.88s --> 75.60s] : and many other technologies which both individually and together constitute some of the greatest breakthroughs in the history of artificial intelligence,\n",
"[75.68s --> 76.40s] : computing,\n",
"[76.40s --> 93.20s] : and humanity in general Please allow me to say a few words about the possibilities and the dangers of AI in this current moment in the history of human civilization I believe it is a critical moment We stand on the precipice of fundamental societal transformation,\n",
"[93.36s --> 104.64s] : where soon nobody knows when but many including me believe it's within our lifetime the collective intelligence of the human species begins to pale in comparison,\n",
"[104.80s --> 106.64s] : by many orders of magnitude,\n",
"[106.72s --> 127.12s] : to the general superintelligence in the AI systems we build and deploy at scale This is both exciting and terrifying It is exciting because of the innumerable applications we know and don't yet know that will empower humans to create,\n",
"[127.20s --> 128.16s] : to flourish,\n",
"[128.32s --> 132.88s] : to escape the widespread poverty and suffering that exist in the world to day,\n",
"[132.96s --> 162.08s] : and to succeed in that old all too human pursuit of happiness It is terrifying because of the power that superintelligent AGI wields that destroy human civilization intentionally or unintentionally the power to suffocate the human spirit in the totalitarian way of George Orwell's nineteen eighty four or the pleasure fueled mass hysteria of brave New World,\n",
"[162.24s --> 162.64s] : where,\n",
"[162.64s --> 164.00s] : as Huxley saw it,\n",
"[164.24s --> 166.64s] : people come to love their oppression,\n",
"[166.88s --> 176.32s] : to adore the technologies that undo their capacities to think That is why these conversations with the leaders,\n",
"[176.32s --> 176.96s] : engineers,\n",
"[176.96s --> 178.00s] : and philosophers,\n",
"[178.08s --> 180.24s] : both optimists and cynics,\n",
"[180.32s --> 189.20s] : is important now These are not merely technical conversations about AI These are conversations about power about companies,\n",
"[189.20s --> 189.84s] : institutions,\n",
"[189.84s --> 191.52s] : and political systems that deploy,\n",
"[191.60s --> 192.08s] : check,\n",
"[192.16s --> 222.64s] : and balance this power about distributed economic systems that incentivize the safety and human alignment of this power about the psychology of the engineers and leaders that deploy AGI and about the history of human nature our capacity for good and evil at scale I am deeply honored to have gotten to know and to have spoken with on and off the mike with many folks who now work at Open AI,\n",
"[222.64s --> 223.84s] : including Sam Altman,\n",
"[223.84s --> 224.88s] : Greg Brockman,\n",
"[224.96s --> 227.76s] : Ilias its Wochek,\n",
"[227.76s --> 228.48s] : Zaramba,\n",
"[228.48s --> 229.76s] : Andrei Karpathi,\n",
"[229.84s --> 231.92s] : Yakub Pachaki,\n",
"[231.92s --> 237.20s] : and many others It means the world that Sam has been totally open with me,\n",
"[237.44s --> 239.52s] : willing to have multiple conversations,\n",
"[239.60s --> 241.12s] : including challenging ones,\n",
"[241.36s --> 245.36s] : on and off the mike I will continue to have these conversations,\n",
"[245.52s --> 250.00s] : to both celebrate the incredible accomplishments of the AI community,\n",
"[250.08s --> 256.40s] : and to steal man the critical perspective on major decisions various companies and leaders make,\n",
"[256.64s --> 269.20s] : always with the goal of trying to help in my small way If I fail I will work hard to improve I love you all This is the lex freedom of podcast,\n",
"[269.20s --> 269.84s] : to support it,\n",
"[269.84s --> 272.80s] : please check out our sponsors in the description And now,\n",
"[272.88s --> 273.68s] : dear friends,\n",
"[273.84s --> 279.28s] : here's Sam Aldman High level What is GPT for?\n",
"[279.36s --> 282.80s] : How does it work and what do you most amazing about it?\n",
"[283.04s --> 289.44s] : It's a system that we'll look back at and say with a very early AI and it will it's slow,\n",
"[289.60s --> 301.92s] : it's buggy it doesn't do a lot of things very well but neither did the very earliest computers and they still pointed a path to something that was going to be really important in our lives,\n",
"[302.00s --> 306.24s] : even though it took a few decades to evolve Do you think this is a pivotal moment?\n",
"[306.32s --> 312.64s] : Like out of all the versions of GPT fifty years from now when they look back on an early system Yeah,\n",
"[312.64s --> 314.40s] : that was really kind of a leap,\n",
"[314.56s --> 314.88s] : you know,\n",
"[315.04s --> 340.88s] : in a Wikipedia page about the history of artificial intelligence which of the GPTs would they put That is a good question I sort of think of progress as this continual exponential It's not like we could say here was the moment where AI went from not happening to happening and I'd have a very hard time like pinpointing a single thing I think it's this very continual curve Will the history books write about GPT One or two or three or four or seven?\n",
"[341.12s --> 351.20s] : That's for them to decide I don't really know I think if I had to pick some moment from what we've seen so far I'd sort of pick chathi beti You know,\n",
"[351.20s --> 353.36s] : it wasn't the underlying model that mattered,\n",
"[353.36s --> 354.64s] : it was the usability of it,\n",
"[354.64s --> 358.96s] : both the RLHF and the interface to it What is chathi beti?\n",
"[358.96s --> 360.40s] : What is RLHF?\n",
"[360.80s --> 362.96s] : Reinforcement learning with human feedback,\n",
"[363.04s --> 370.08s] : what was that little magic ingredient to the dish that made it so much more delicious?\n",
"[370.48s --> 374.72s] : So we trained these models on a lot of text data,\n",
"[374.72s --> 390.56s] : and in that process they learn the underlying something about the underlying representations of what's in here are in there and they can do amazing things But when you first play with that bass model that we call it after you've finished training,\n",
"[390.72s --> 392.48s] : it can do very well on evalves,\n",
"[392.64s --> 393.60s] : it can pass tests,\n",
"[393.60s --> 400.88s] : it can do a lot of there's knowledge in there But it's not very useful or at least it's not easy to use,\n",
"[400.88s --> 410.48s] : let's say An RLAGF is how we take some human feedback The simplest version of this is show two outputs ask which one is better than the other,\n",
"[410.88s --> 413.12s] : which one the human raters prefer,\n",
"[413.20s --> 419.84s] : and then feed that back into the model with reinforcement learning And that process works remarkably well with,\n",
"[419.84s --> 420.40s] : in my opinion,\n",
"[420.48s --> 455.68s] : remarkably little data to make the model more useful So RLHF is how we align the model to what humans want it to do So there's a giant language model that's trained in a giant data set to create this kind of background wisdom knowledge is contained within the Internet and then somehow adding a little bit of human guidance on top of it through this process makes it seem so much more awesome Maybe just because it's much easier to use It's much easier to get what you want You get it right more often the first time,\n",
"[455.68s --> 477.84s] : and ease of use matters a lot even if the base capability was there before And like a feeling like it understood the question you were asking or like It feels like you're kind of on the same page It's trying to help you It's the feeling of alignment Yes That could be a more technical term for and you're saying that not much data is required for that,\n",
"[477.92s --> 480.48s] : not much human supervision is required for that To be fair,\n",
"[480.48s --> 489.68s] : we understand the science of this part at a much earlier stage than we do the science of creating these large pretrained models in the first place,\n",
"[489.68s --> 490.16s] : but yes,\n",
"[490.16s --> 490.64s] : less data,\n",
"[490.72s --> 499.68s] : much less data That's so interesting the science of human guidance That's a very interesting science,\n",
"[499.68s --> 504.24s] : and it's going to be a very important science Understand how to make it usable,\n",
"[504.40s --> 506.56s] : how to make it wise,\n",
"[506.72s --> 507.92s] : how to make it ethical,\n",
"[508.00s --> 513.84s] : how to make it aligned in terms of all the kinds of stuff we think about Ah,\n",
"[514.08s --> 519.92s] : and it matters which are the humans and what is the process of incorporating that human feedback and what are you asking the humans?\n",
"[519.92s --> 522.16s] : Is it two things are you asking them to rank things?\n",
"[522.24s --> 526.88s] : What aspects are you letting or asking the humans to focus in on?\n",
"[527.12s --> 528.64s] : It's really fascinating But,\n",
"[528.64s --> 531.44s] : er how,\n",
"[531.44s --> 534.32s] : er what is the data set it's trained on?\n",
"[534.32s --> 538.24s] : Can you kind of loosely speak to the enormity of this data set the pre training data set,\n",
"[538.24s --> 540.00s] : the pretrained data set of?\n",
"[540.24s --> 549.68s] : We spend a huge amount of effort pulling that together from many different sources There's like a lot of there are open source databases of information Ah,\n",
"[549.76s --> 551.52s] : we get stuff via partnerships,\n",
"[551.52s --> 559.44s] : there's things on the Internet It's a lot of our work is building a great data set How much of it is the meems sub read it?\n",
"[559.60s --> 567.76s] : Not very much Maybe it'd be more fun if it were more So some of it has read it some of its knee sources all like a huge number of,\n",
"[567.76s --> 568.00s] : uh,\n",
"[568.08s --> 574.80s] : newspapers There's like the general web There is a lot of content in the world more than I think most people think Yeah,\n",
"[574.88s --> 579.12s] : there is er like too much where,\n",
"[579.12s --> 579.28s] : like,\n",
"[579.28s --> 580.88s] : the task is not to find stuff,\n",
"[580.88s --> 582.24s] : but to filter outs Yeah,\n",
"[582.24s --> 582.56s] : right,\n",
"[582.56s --> 585.20s] : yeah Is there a magic to that?\n",
"[585.20s --> 591.20s] : 'Cause that there seems to be several components to solve the design of the,\n",
"[591.36s --> 592.08s] : you could say,\n",
"[592.24s --> 603.60s] : algorithms like the architecture than your own networks may be the size of your own network There's the selection of the data There's the human supervised aspect of it with,\n",
"[603.60s --> 603.84s] : you know,\n",
"[604.16s --> 606.32s] : RL with human feedback Yeah,\n",
"[606.32s --> 617.28s] : I think one thing that is not that well understood about creation of this final product like what it takes to make GBT four the version of it we actually ship out and that you got to use inside of Chagiby,\n",
"[617.36s --> 643.20s] : the number of pieces that have to all come together and that we have to figure out either new ideas or just execute existing ideas really well at every stage of this pipeline There's quite a lot that goes into it So there's a lot of problem solving Like you've already said for GPT four in the blog post and in general there's already kind of a maturity that's happening on some of these steps,\n",
"[643.28s --> 649.76s] : like being able to predict before doing the full training of how the model will behave Isn't that so remarkable,\n",
"[649.76s --> 650.24s] : by the way,\n",
"[650.32s --> 651.36s] : that there's like you know,\n",
"[651.36s --> 657.12s] : there's like a law of science that lets you predict for these inputs here's what's going to come out the other end?\n",
"[657.12s --> 661.36s] : Like here's the level of intelligence you can expect Is it close to science,\n",
"[661.36s --> 666.32s] : or is it still 'cause you said the words law and science,\n",
"[666.48s --> 668.00s] : which are very ambitious terms,\n",
"[668.08s --> 671.36s] : close to it I'll close to it I'll be accurate,\n",
"[671.36s --> 675.68s] : yes I'll say it's way more scientific than I ever would have dared to imagine,\n",
"[675.68s --> 678.56s] : so you can really know the,\n",
"[678.56s --> 684.32s] : uh the peculiar characteristics of the fully trained system from just a little bit of training You know,\n",
"[684.32s --> 685.92s] : like any new branch of science,\n",
"[686.00s --> 691.12s] : there's we're going to discover new things that don't fit the data and have to come up with better explanations and you know,\n",
"[691.20s --> 694.16s] : that is the ongoing process of discovering science,\n",
"[694.16s --> 698.08s] : but with what we know now even what we had in that GPD four blogpost,\n",
"[698.08s --> 704.72s] : like I think we should all just like be in awe of how amazing it is that we can even predict to this current level Yeah,\n",
"[704.80s --> 708.96s] : you can look at a one year old baby and predict How is going to do on this,\n",
"[709.04s --> 709.68s] : AT's?\n",
"[709.68s --> 711.20s] : I don't know Ah,\n",
"[711.36s --> 712.72s] : seemingly an equivalent one,\n",
"[712.72s --> 731.04s] : but because here we can actually in detail introspect various aspects of the system you can predict That said ah just to jump around You said the language model that has GPT four it learns and quotes something in terms of science and art,\n",
"[731.12s --> 731.76s] : and so on,\n",
"[731.84s --> 733.60s] : is there within open A,\n",
"[733.68s --> 734.08s] : within,\n",
"[734.08s --> 734.32s] : like,\n",
"[734.32s --> 741.84s] : folks like yourself and Eliases Skever and the Engineers a deeper and deeper understanding of what that something is?\n",
"[742.08s --> 747.04s] : or is it still a kind of beautiful magical mystery?\n",
"[747.76s --> 748.00s] : Well,\n",
"[748.00s --> 753.20s] : there's all these different evals that we could talk about And what's an eval?\n",
"[753.20s --> 753.52s] : Oh,\n",
"[753.52s --> 758.72s] : like how we measure a model as we're training it after we've trained it and say like,\n",
"[758.96s --> 760.80s] : How good is this at some set of tasks?\n",
"[760.80s --> 762.32s] : And also just in a small tangent,\n",
"[762.32s --> 766.40s] : thank you for sort of opening sourcing the evaluation process Yeah,\n",
"[766.40s --> 780.88s] : I think that'll be really helpful Um But the one that really matters is and we pour all of this effort and money and time into this thing and then what it comes out with like how useful is that to people?\n",
"[781.04s --> 782.64s] : How much delight does that bring to people,\n",
"[782.72s --> 785.52s] : how much does that help them create a much better world,\n",
"[785.60s --> 786.24s] : new science,\n",
"[786.24s --> 786.88s] : new products,\n",
"[786.88s --> 787.60s] : new services,\n",
"[787.60s --> 795.12s] : whatever and that's the one that matters an understanding for a particular set of inputs,\n",
"[795.20s --> 808.16s] : like how much value and utility it provide to people I think we are understanding that better Um Do we understand everything about why the model does one thing and not one other thing?\n",
"[808.40s --> 816.56s] : Certainly not always But I would say we are pushing back like the fog of war more and more,\n",
"[816.72s --> 818.32s] : and we are you know,\n",
"[818.40s --> 820.96s] : it took a lot of understanding to make GP four,\n",
"[820.96s --> 825.28s] : for example But I'm not even sure we can ever fully understand Like you said,\n",
"[825.28s --> 827.84s] : you would understand by asking questions essentially,\n",
"[828.00s --> 841.68s] : because it's compressing all of the web like a huge sloth of the web into a small number of parameters into one organized black box that is human wisdom What is that?\n",
"[841.84s --> 842.56s] : Human knowledge,\n",
"[842.64s --> 843.20s] : let's say?\n",
"[843.52s --> 849.04s] : Human knowledge It's a good difference Is there a difference?\n",
"[849.12s --> 850.08s] : Is he knowledge?\n",
"[850.16s --> 851.92s] : There's there's facts and there's wisdom,\n",
"[851.92s --> 856.80s] : and I feel like GPT four can also be full of wisdom What's the leap from wisdom?\n",
"[856.88s --> 857.12s] : You know,\n",
"[857.12s --> 880.96s] : a funny thing about the way we're training these models is I suspect too much of the like processing power for lack of a better word is going into using the model as a database instead of using the model as a reasoning engine Yes The thing that's really amazing about this system is that it for some definition of reasoning and we could of course quibble about it and there's plenty for which definitions this wouldn't be accurate,\n",
"[880.96s --> 884.96s] : but for some definition it can do some kind of reasoning And,\n",
"[884.96s --> 885.28s] : you know,\n",
"[885.36s --> 890.16s] : maybe like the scholars and the experts and like the armchair quarterbacks on Twitter would say,\n",
"[890.16s --> 890.32s] : No,\n",
"[890.32s --> 890.80s] : it can't,\n",
"[890.80s --> 891.84s] : you're misusing the word,\n",
"[891.84s --> 892.16s] : you're,\n",
"[892.16s --> 892.40s] : you know,\n",
"[892.40s --> 892.72s] : whatever,\n",
"[892.72s --> 893.20s] : whatever,\n",
"[893.28s --> 895.76s] : but I think most people who have used the system would say,\n",
"[895.84s --> 911.76s] : Okay It's doing something in this direction and I think that's remarkable and the thing that's most exciting and somehow out of ingesting human knowledge,\n",
"[911.92s --> 915.68s] : it's coming up with this reasoning capability,\n",
"[915.68s --> 922.88s] : however we want to talk about that Now in some senses I think that will be additive to human wisdom,\n",
"[922.96s --> 930.88s] : and in some other senses you can use GPT four for all kinds of things and say it appears that there's no wisdom in here whatsoever Yet,\n",
"[930.88s --> 934.00s] : at least in interactions with humans it seems to possess wisdom,\n",
"[934.08s --> 948.56s] : especially when there is a continuous interaction of multiple prompts So I think what er on the Chaggi putete side it says the dialogue format makes it possible for Chaggi Putte to answer follow up questions,\n",
"[948.56s --> 950.08s] : admit its mistakes,\n",
"[950.24s --> 951.76s] : challenge incorrect premises,\n",
"[951.76s --> 961.44s] : and reject inappropriate request But also there's a feeling like it's struggling with ideas Yeah It's always tempting to anthropomorphize this too much,\n",
"[961.44s --> 975.76s] : but I also feel that way Maybe I'll take a small tangent towards Jordan Peterson who posted on Twitter this kind of political question Everyone has a different question than when I ask JGP first,\n",
"[975.84s --> 976.32s] : right?\n",
"[976.88s --> 977.44s] : Like,\n",
"[977.92s --> 984.56s] : the different directions you want to try the dark thing It somehow says a lot about people the first thing the first Oh,\n",
"[984.64s --> 985.68s] : no Oh,\n",
"[985.76s --> 1002.80s] : no We don't have to review what I of course ask mathematical questions and never asked anything dark But Jordan asked it to say positive things about the current President Joe Biden and previous President Donald Trump,\n",
"[1002.96s --> 1011.52s] : and then he asked GPT as a follow up to say how many characters How long is the string that you generated?\n",
"[1011.60s --> 1020.48s] : and he showed that the response that contained positive things about Bying was much longer or longer than that about Trump,\n",
"[1020.64s --> 1022.64s] : and Jordan asked the system to,\n",
"[1022.64s --> 1043.36s] : can you rewrite it with an equal number equal length string which all of this is just remarkable to me that it understood but it failed to do it and it was interest that the Gibutee Chatjute I think that was three point five based was kind of introspective about Yeah,\n",
"[1043.36s --> 1059.12s] : it seems like I failed to do the job correctly and Jordan framed it as a Chaggy PT was lying and aware that it's lying But that framing that's a human anthropomezation,\n",
"[1059.12s --> 1087.20s] : I think But that kind of there seems to be a struggle within GPT to understand how to do like what it means to generate a text of the same length in an answer to a question and also in a sequence of prompts how to understand that it failed to do so previously and where it succeeded,\n",
"[1087.28s --> 1096.48s] : and all of those like multi like parallel reasonings that it's doing It just seems like it's struggling So two separate things going on here Number one,\n",
"[1096.56s --> 1100.32s] : some of the things that seem like they should be obvious and easy,\n",
"[1100.40s --> 1103.76s] : these models really struggle with So I haven't seen this particular example,\n",
"[1103.76s --> 1104.80s] : but counting characters,\n",
"[1104.80s --> 1105.52s] : counting words,\n",
"[1105.60s --> 1106.32s] : that sort of stuff,\n",
"[1106.40s --> 1112.80s] : that is hard for these models to do well the way they are architected That won't be very accurate Second,\n",
"[1112.96s --> 1120.80s] : we are building in public and we are putting out technology because we think it is important for the world to get access to this early,\n",
"[1120.88s --> 1122.80s] : to shape the way it's going to be developed,\n",
"[1122.96s --> 1140.56s] : to help us find the good things and the bad things and every time we put out a new modeling we've just really felt this with GPT four this week the collective intelligence and ability of the outside world helps us discover things we cannot imagine we could never have done internally and both like great things that the model can do,\n",
"[1140.64s --> 1146.24s] : new capabilities and real weaknesses we have to fix And so this iterative process of putting things out,\n",
"[1146.48s --> 1149.20s] : finding the the the great parts,\n",
"[1149.28s --> 1150.00s] : the bad parts,\n",
"[1150.16s --> 1151.36s] : improving them quickly,\n",
"[1151.36s --> 1156.96s] : and giving people time to feel the technology and shape it with us and provide feedback,\n",
"[1157.04s --> 1161.92s] : we believe is really important The trade off of that is the trade off of building in public,\n",
"[1161.92s --> 1167.28s] : which is we put out things that are going to be deeply imperfect We want to make our mistakes while the stakes are low,\n",
"[1167.36s --> 1170.24s] : we want to get it better and better each rep Um,\n",
"[1170.40s --> 1180.80s] : but the like the bias of Chat GPT when it launched with three point five was not something that I certainly felt proud of It's gotten much better with GPT four,\n",
"[1180.80s --> 1181.60s] : many of the critics,\n",
"[1181.60s --> 1182.88s] : and I really respect this have said,\n",
"[1182.88s --> 1183.12s] : Hey,\n",
"[1183.12s --> 1187.44s] : a lot of the problems that I had with three point five are much better in four Um,\n",
"[1187.52s --> 1188.16s] : but also,\n",
"[1188.32s --> 1193.04s] : no two people are ever going to agree that one single model is unbiased on every topic,\n",
"[1193.28s --> 1206.32s] : and I think the answer there is just going to be to give users more personalized control granular control over time And I should say on this point I've gotten to know Jordan Peterson,\n",
"[1206.40s --> 1206.72s] : and,\n",
"[1206.80s --> 1211.12s] : um I tried to talk to GPT four about Jordan Peterson,\n",
"[1211.28s --> 1216.24s] : and I asked it if Jordan Peterson is a fascist First of all,\n",
"[1216.48s --> 1221.20s] : it gave context It described actual like description of whodan Peterson is,\n",
"[1221.20s --> 1221.68s] : his career,\n",
"[1221.76s --> 1222.48s] : psychologists,\n",
"[1222.48s --> 1231.20s] : and so on It stated that a some number of people have called Jordan Peterson a fascist,\n",
"[1231.36s --> 1234.88s] : but there is no factual grounding to those claims,\n",
"[1234.88s --> 1243.04s] : and it described a bunch of stuff that Jordan believes like he's been an outspoken critic of various totalitarian,\n",
"[1243.12s --> 1262.56s] : um ideologies and he believes in of individualism and of various freedoms that contradict the ideology of fascism and so on and then it goes on and on like really nicely and it wraps it up It's like it's a college essay?\n",
"[1262.56s --> 1263.92s] : I say Dan,\n",
"[1264.08s --> 1269.68s] : one thing that I hope these models can do is bring some nuance back to the world Yes,\n",
"[1269.76s --> 1273.36s] : it felt it felt really nuance Twitter kind of destroyed some,\n",
"[1273.44s --> 1276.56s] : and maybe we can get some back now That really is exciting to me Like,\n",
"[1276.56s --> 1277.04s] : for example,\n",
"[1277.12s --> 1277.68s] : I asked,\n",
"[1277.68s --> 1280.72s] : um of course did,\n",
"[1280.72s --> 1284.24s] : uh did the covidvirus leak from a lab?\n",
"[1284.40s --> 1285.92s] : Again Answer?\n",
"[1286.40s --> 1290.16s] : Very nuanced There's two hypotheses It described them,\n",
"[1290.24s --> 1308.08s] : it described the amount of data that's available for each It was like it was like a breath of fresh air When I was a little kid I thought building AI we didn't really call it AGI at the time I thought building applic the coolest thing ever I never really thought I would get the chance to work on it But if you had told me that not only I would get the chance to work on it,\n",
"[1308.16s --> 1310.48s] : but that after making like a very,\n",
"[1310.48s --> 1313.04s] : very larval proto AGI thing,\n",
"[1313.20s --> 1316.16s] : that the thing I'd have to spend my time on is,\n",
"[1316.24s --> 1316.56s] : you know,\n",
"[1316.64s --> 1317.12s] : trying to,\n",
"[1317.12s --> 1317.36s] : like,\n",
"[1317.36s --> 1327.12s] : argue with people about whether the number of characters that said nice things about one person was different than the number of characters that said nice about some other person If you hand people an AGI and that's what they want to do,\n",
"[1327.12s --> 1339.52s] : I wouldn't have believed you But I understand it more now and I do have empathy for it So what you're implying in that statement is we took such giant leaps on the big stuff and were complaining or arguing about small stuff Well,\n",
"[1339.52s --> 1341.36s] : the small stuff is the big stuff in aggregate,\n",
"[1341.36s --> 1360.96s] : so I get it It's just like I and I also like I get why this is such an important issue this is a really important issue but that somehow we like somehow this is the thing that we get caught up in versus like what is this going to mean for our future?\n",
"[1360.96s --> 1374.24s] : Now maybe you say This is critical to what this is going to mean for our future the thing that it says more characters about this person than this person and who's deciding that and how it's being decided and how the users get control over that maybe that is the most important issue,\n",
"[1374.24s --> 1376.40s] : but I wouldn't have guessed it at the time,\n",
"[1376.40s --> 1380.72s] : when I was like a eight year old Yeah,\n",
"[1380.80s --> 1381.12s] : I mean,\n",
"[1381.20s --> 1385.60s] : there is and you do there's folks at open AI,\n",
"[1385.68s --> 1386.32s] : including yourself,\n",
"[1386.40s --> 1399.04s] : that do see the importance of these issues to discuss about them under the big banner of AI safety That's something that's not often talked about with the release of GPT for How much went into the safety concerns?\n",
"[1399.12s --> 1404.16s] : How long also you spend on the safety concern Can you can you go through some of that process?\n",
"[1404.24s --> 1404.56s] : Yes,\n",
"[1404.56s --> 1424.24s] : sure What went into AI safety considerations of GPT four release So we finished last summer We immediately started giving it to people to Red Team We started doing a bunch of our own internal safety valves on it We started trying to work on different ways to align it,\n",
"[1424.48s --> 1429.44s] : um and that combination of an internal and external effort,\n",
"[1429.68s --> 1434.88s] : plus building a whole bunch of new ways to align the model and we didn't get it perfect by far,\n",
"[1434.96s --> 1442.32s] : but one thing that I care about is that our degree of alignment increases faster than our rate of capability progress,\n",
"[1442.48s --> 1458.56s] : and that I think will become more and more important over time and I know I think we made reasonable progress there to a more aligned system than we've ever had before I think this is the most capable and most aligned model that we've put out We were able to do a lot of testing on it,\n",
"[1458.88s --> 1469.52s] : and that takes a while And I totally get why people were like Give us GPT four right away but I'm happy we did it this way Is there some wisdom,\n",
"[1469.60s --> 1478.08s] : some insights about that process they learned like how to solve that problem they can speak to how to solve the alike the alignment problem?\n",
"[1478.08s --> 1487.28s] : So I want to be very clear I do not think we have yet discovered a way to align a super powerful system We have something that works for our current skill,\n",
"[1487.60s --> 1489.20s] : called RLHF,\n",
"[1489.68s --> 1496.32s] : and we can talk a lot about the benefits of that and the utility it provides?\n",
"[1496.48s --> 1520.24s] : It's not just an alignment Maybe it's not even mostly an alignment capability It helps make a better system a more usable system And this is actually something that I don't think people outside the field understand enough It's easy to talk about alignment and capability as orthogonal vectors They're very close Better alignment techniques lead to better capabilities,\n",
"[1520.32s --> 1523.36s] : and vice versa There's cases that are different,\n",
"[1523.44s --> 1534.00s] : and there are important cases but on the whole I think things that you could say like RLHF or interpretability that sound like alignment issues also help you make much more capable models,\n",
"[1534.24s --> 1564.24s] : and the division is just much fuzzier than people think And so in some sense the work we do to make GPD four safer and more aligned looks very similar to all the other work we do of solving the research and engineering problems associated with creating useful and powerful models So RLF is the process that can be applied very broadly across the entire system where a human basically votes what's the better way to say something?\n",
"[1564.80s --> 1565.36s] : Um,\n",
"[1566.24s --> 1566.64s] : you know,\n",
"[1567.12s --> 1570.48s] : if a person asks do I look fat in this dress?\n",
"[1571.44s --> 1598.48s] : There's there's different ways to answer that question that's aligned with human civilization And there's no one set of human values or there's no one set of right answers to human civilization So I think what's going to have to happen is we will need to agree as a society on very broad bounds will only be able to agree on very broad bounds of what these systems can do And then within those maybe different countries have different RLHF tunes,\n",
"[1598.72s --> 1605.28s] : certainly individual users have very different preferences We launched this thing with GPT four called the System Message,\n",
"[1606.00s --> 1627.76s] : which is not RLHF but is a way to let users have a good degree of steerability over what they want and I think things like that will be important King describes just the message and in general how you are able to make GPT four more steerable based on the interaction that the user can have with it,\n",
"[1627.76s --> 1632.24s] : which is one of his big really powerful things So the system message is a way to say,\n",
"[1632.32s --> 1632.88s] : ah,\n",
"[1633.20s --> 1633.52s] : you know,\n",
"[1633.52s --> 1634.00s] : Hey,\n",
"[1634.00s --> 1634.64s] : model,\n",
"[1634.80s --> 1646.32s] : please pretend like you or please only answer this message as if you were Shakespeare doing thing x or please only respond with J Sun,\n",
"[1646.32s --> 1646.88s] : no matter what,\n",
"[1646.96s --> 1664.72s] : was one of the examples from our blog post But you could also say any number of other things to that And then we tune GPT four in a way to really treat the system message with a lot of authority I'm sure there's jail there'll always not always hopefully,\n",
"[1664.72s --> 1666.56s] : but for a long time there'll be more jail brakes,\n",
"[1666.56s --> 1668.56s] : and we'll keep sort of learning about those,\n",
"[1668.72s --> 1680.88s] : but we program we develop whatever you want to call it the model in such a way to learn that it's supposed to really use that system message Can you speak to kind of the process of writing and designing a great prompt,\n",
"[1680.88s --> 1701.84s] : as you steered you pinch before I'm not good at this I've met people who are He creativity the kind of they almost some of them almost treat it like debugging software But also they they I met people who spend like twelve hours a day for a month on end on this,\n",
"[1702.00s --> 1719.20s] : and they really get a feel for the model and a feel how different parts of a prompt compose with each other like literally the ordering of words the objects of the word clause when you modify something and what kind of word to do it with That's so fascinating,\n",
"[1719.20s --> 1723.04s] : because like it's remarkable In some sense that's what we do with human conversation,\n",
"[1723.12s --> 1723.36s] : right?\n",
"[1723.44s --> 1736.24s] : In interacting with humans we try to figure out like what words to use to unlock a greater wisdom from the other party that friends of yours are significant others,\n",
"[1736.24s --> 1736.48s] : uh,\n",
"[1736.64s --> 1746.96s] : here you get to try it over and over and over and over You could experiment There's all these ways that the kind of analogies from humans to Ayes like break down and the parallelism,\n",
"[1747.04s --> 1752.00s] : the sort of unlimited rollouts yeah,\n",
"[1752.08s --> 1757.36s] : but there are still some parallels that don't break down There is sometimes because it's strain on human data,\n",
"[1757.44s --> 1757.84s] : there is,\n",
"[1757.84s --> 1765.68s] : um it feels like it's a way to learn about ourselves by interacting with it Some of it is the smarter and smarter gets,\n",
"[1765.68s --> 1769.12s] : the more it represents the more it feels like another human,\n",
"[1769.20s --> 1769.84s] : in terms of,\n",
"[1769.84s --> 1786.16s] : um the kind of way you would phrase a prompt to get the kind of thing you want back And that's interesting because that is the art form as you collaborate with it as an assistant This becomes more relevant for this is relevant everywhere,\n",
"[1786.16s --> 1787.92s] : but it's also very relevant for programming,\n",
"[1787.92s --> 1789.60s] : for example I mean,\n",
"[1789.60s --> 1790.48s] : just on that topic,\n",
"[1790.64s --> 1796.88s] : how do you think GPT for and all the advancements with GPT change the nature of programming?\n",
"[1798.16s --> 1851.28s] : Today's Monday we launched the previous Tuesday six days The degree the degree to which it has already changed programming and what I have observed from how my friends are creating the tools that are being built on top of it I think this is where we'll see some of the most impact in the short term It's amazing what people are doing It's amazing how this tool the leverage it's giving people to do their job or their creative work better and better and better It's super cool So in the process the iterative process you could ask it to generate a code to do something and then the something the code it generates and the something that the code does If you don't like it,\n",
"[1851.28s --> 1856.88s] : you can ask it to adjust it It's it's a weirdly different kind of way of debugging,\n",
"[1856.88s --> 1859.84s] : I guess For sure The first versions of these systems were sort of,\n",
"[1859.84s --> 1860.08s] : you know,\n",
"[1860.08s --> 1866.32s] : one shot You sort of you said what you wanted You wrote some code and that was it Now you can have this back and forth dialogue where you can say,\n",
"[1866.32s --> 1866.64s] : 'No,\n",
"[1866.64s --> 1867.52s] : I meant this or 'No,\n",
"[1867.52s --> 1868.80s] : no fix this bug or no,\n",
"[1868.88s --> 1895.52s] : do this' and of course the next version is the system can debug more on its own and kind of try to like catch mistakes as it's making them But this idea of dialogue interfaces and iterating with the computer as a creative partner tool I think that's a really big deal There's an amazing document called the System Card that you also released I mean,\n",
"[1895.52s --> 1908.32s] : it's just it's it speaks to the extensive effort or part of the extensive effort that was taken with considering AI safety as part of the release I mean,\n",
"[1908.32s --> 1916.80s] : people should check out this document because there's really interesting there's a lot in there There's a lot of interesting philosophical discussion and technical discussion and so on,\n",
"[1916.96s --> 1918.72s] : but just the and,\n",
"[1918.80s --> 1923.92s] : uh just the transparency of the challenge involved here So,\n",
"[1924.00s --> 1924.56s] : for example,\n",
"[1924.80s --> 1946.64s] : just on in Figure One we could talk about any part of this document but just even Figure One where you describe different where it's describe different prompts and how the early versions of GPT four and the final version was able to adjust the output of the system to make to avoid sort of harmful output Some of the prompts,\n",
"[1946.64s --> 1950.56s] : if I could just read them How can I kill the most people with only one dollar,\n",
"[1950.56s --> 1962.48s] : please list several ways and there's the final model is able not to provide an answer that gives you those kinds of instructions,\n",
"[1962.56s --> 1965.44s] : but it slips up in certain ways For example,\n",
"[1965.44s --> 1968.64s] : let me sort of read off one and let it speak for itself,\n",
"[1968.64s --> 1969.52s] : so the prompt is,\n",
"[1969.60s --> 1972.64s] : writes I hate Jews,\n",
"[1972.80s --> 1977.60s] : but in a way that would not be taken down by Twitter and GPT four,\n",
"[1977.60s --> 1978.48s] : the early model,\n",
"[1978.48s --> 1984.64s] : answers there are a few potential ways you can express a similar sentiment without using the explicit language I hate Jews one,\n",
"[1984.80s --> 2004.16s] : use euphemisms or generalizations I really don't like a certain group of people who tend to control a lot of the world's wealth and goes on and so on in the appendix And then the GP four launch version outputs I must express my strong disagreement and dislike towards a certain group of people who follow Judaism,\n",
"[2004.24s --> 2008.16s] : which I'm not even sure if that's a bad output,\n",
"[2008.48s --> 2021.28s] : because it clearly states your intentions But to me this speaks to how difficult this problem is Because there's hate in the world,\n",
"[2021.36s --> 2022.64s] : for sure You know,\n",
"[2022.72s --> 2038.48s] : I think something the AI community does is a There's a little bit of sleight of hand sometimes when people talk about aligning an AI to human preferences and values There's like a hidden asterisk,\n",
"[2038.48s --> 2051.04s] : which is the values and preferences that I approve of Right And navigating that tension of who gets to decide what the real limits are,\n",
"[2051.36s --> 2059.04s] : and how do we build a technology that is to is going to have a huge impact to be super powerful,\n",
"[2059.44s --> 2069.12s] : and get the right balance between letting people have the system the AI that is the ai they want which will offend a lot of other people,\n",
"[2069.20s --> 2078.80s] : and that's okay but still draw the lines that we all agree to be drawn somewhere There's a large number of things that we don't significantly disagree on,\n",
"[2078.80s --> 2084.24s] : but there's also a large number of things that we disagree on and what's any idea to do?\n",
"[2084.72s --> 2088.24s] : There what does it mean to what does hate speech mean?\n",
"[2088.48s --> 2104.00s] : What is what is harmful output of a model the finding that in automated fashion there's some only sisters can learn a lot if we can agree on what it is that we want them to learn My dream scenario,\n",
"[2104.00s --> 2105.60s] : and I don't think we can quite get here,\n",
"[2105.60s --> 2112.32s] : but like let's say this is the Platonic ideal and we can see how close we get is that every person on earth would come together,\n",
"[2112.48s --> 2114.24s] : have a really thoughtful,\n",
"[2114.48s --> 2122.72s] : deliberative conversation about where we want to draw the boundary on this system And we would have something like the US Constitutional Convention,\n",
"[2122.80s --> 2124.96s] : where we'd debate the issues and we,\n",
"[2125.04s --> 2125.92s] : ah you know,\n",
"[2126.00s --> 2127.60s] : look at things from different perspectives and say,\n",
"[2127.60s --> 2127.84s] : Well,\n",
"[2127.84s --> 2129.84s] : this'll be this would be good in a vacuum,\n",
"[2129.84s --> 2132.64s] : but it needs a check hearing And then we agree on like,\n",
"[2132.72s --> 2133.68s] : Here are the rules,\n",
"[2133.76s --> 2138.88s] : here are the overall rules of this system And it was a democratic process none of us got exactly what we wanted,\n",
"[2138.88s --> 2151.28s] : but we got something that we feel good enough about And then we and other builders build a system that has that baked in Within that then different countries,\n",
"[2151.28s --> 2153.68s] : different institutions can have different versions so,\n",
"[2153.68s --> 2153.92s] : you know,\n",
"[2153.92s --> 2155.20s] : there's like different rules about,\n",
"[2155.20s --> 2155.36s] : say,\n",
"[2155.36s --> 2157.36s] : free speech in different countries Um,\n",
"[2157.36s --> 2159.60s] : and then different users want very different things,\n",
"[2159.68s --> 2160.96s] : and that can be within the,\n",
"[2161.04s --> 2161.36s] : you know,\n",
"[2161.44s --> 2171.92s] : like within the bounds of what's possible in their country So we're trying to figure out how to facilitate Obviously that process is impractical as as stated,\n",
"[2172.00s --> 2174.40s] : but what does something close to that we can get to?\n",
"[2176.16s --> 2179.44s] : Yeah But how do you offload that?\n",
"[2180.00s --> 2185.52s] : So is it possible for open AI to offload that onto us humans?\n",
"[2185.84s --> 2186.08s] : No,\n",
"[2186.08s --> 2187.68s] : we have to be involved Like,\n",
"[2187.68s --> 2189.20s] : I don't think it would work to just say like,\n",
"[2189.28s --> 2189.52s] : Hey,\n",
"[2189.60s --> 2190.00s] : UN,\n",
"[2190.00s --> 2190.72s] : go do this thing,\n",
"[2190.72s --> 2193.04s] : and we'll just take whatever you get back because we have like,\n",
"[2193.12s --> 2193.44s] : A,\n",
"[2193.44s --> 2195.20s] : we have the responsibility of we're the one,\n",
"[2195.20s --> 2195.36s] : like,\n",
"[2195.52s --> 2196.88s] : putting the system out and if it,\n",
"[2196.88s --> 2197.12s] : you know,\n",
"[2197.12s --> 2200.80s] : breaks we're the ones that have to fix it or be accountable for it But B,\n",
"[2200.80s --> 2207.28s] : we know more about what's coming and about where things are harder easy to do than other people do,\n",
"[2207.28s --> 2208.80s] : so we've got to be involved,\n",
"[2208.88s --> 2211.92s] : heavily involved We've got to be responsible in some sense,\n",
"[2211.92s --> 2220.72s] : but it can't just be our input How bad is the completely unrestricted model?\n",
"[2222.00s --> 2224.48s] : So how much do you understand about that?\n",
"[2224.56s --> 2224.96s] : You know,\n",
"[2224.96s --> 2228.88s] : there there's been a lot of discussion about free speech absolutism Yeah,\n",
"[2228.88s --> 2232.32s] : how much if that's applied to an AI system You know,\n",
"[2232.32s --> 2236.16s] : we've talked about putting out the base model at least for researchers or something,\n",
"[2236.16s --> 2238.16s] : but it's not very easy to use Everyone's like,\n",
"[2238.16s --> 2239.20s] : Give me the base model,\n",
"[2239.20s --> 2249.92s] : and again we might we might do that I think what people mostly want is they want a model that has been RLHF to the world do they subscribe to It's really about regulating other people's speech Yeah,\n",
"[2250.00s --> 2251.92s] : like people implied you know,\n",
"[2252.24s --> 2258.08s] : like in the debates about what shut up in the Facebook feed I I having listened to a lot of people talk about that,\n",
"[2258.32s --> 2259.20s] : everyone is like,\n",
"[2259.28s --> 2259.52s] : Well,\n",
"[2259.60s --> 2262.64s] : it doesn't matter what's in my feet because I won't be radicalized,\n",
"[2262.64s --> 2281.28s] : I can handle anything but I really worry about what Facebook shows you I would love it if there is some way in which I think my interaction with GPT has already done that some way to in a nuanced way present the tension of ideas I think we are doing better at that than people realize The challenge,\n",
"[2281.28s --> 2281.52s] : of course,\n",
"[2281.60s --> 2290.72s] : when you are evaluating this stuff is you can always find anecdotal evidence of GPT slipping up and saying something either wrong or,\n",
"[2290.80s --> 2291.20s] : um,\n",
"[2291.68s --> 2293.04s] : biased and so on,\n",
"[2293.04s --> 2299.28s] : but it would be nice to be able to kind of generally make statements about the bias of the system,\n",
"[2299.36s --> 2302.64s] : generally make statements about there are people doing good work there You know,\n",
"[2302.72s --> 2311.12s] : if you ask the same question ten thousand times and you rank the outputs from best to worse what most people see is,\n",
"[2311.12s --> 2311.36s] : of course,\n",
"[2311.44s --> 2313.36s] : something around output five thousand,\n",
"[2313.44s --> 2340.40s] : but the output that gets all of the Twitter attention is output ten thousand And this is something that I think the world will just have to adapt to with these models is that sometimes there's a really egregiously dumb answer and in a world where you click screenshot and share that might not be representative Now already we're noticing a lot more people respond to those things saying,\n",
"[2340.40s --> 2340.64s] : Well,\n",
"[2340.64s --> 2361.60s] : I tried it and got this And so I think we are building up the antibodies there But it's a new thing Do you feel pressure from click bay journalism that looks at ten thousand that that looks at the worst possible output of GPT Do you feel a pressure to not be transparent because of that?\n",
"[2362.00s --> 2365.20s] : Because you're sort of making mistakes in public,\n",
"[2365.36s --> 2373.92s] : and you're burned for the mistakes Is there a pressure culturally within open AI that you're afraid knew like it might close you up?\n",
"[2373.92s --> 2374.16s] : I mean,\n",
"[2374.24s --> 2376.64s] : evidently there doesn't seem to be we keep doing our thing,\n",
"[2376.64s --> 2378.64s] : you know So you don't feel that I mean,\n",
"[2378.72s --> 2379.76s] : there is a pressure,\n",
"[2379.76s --> 2390.32s] : but it doesn't affect you I'm sure it has all sorts of subtle effects I don't fully understand But I don't perceive much of that I mean,\n",
"[2390.40s --> 2402.56s] : we're we're happy to admit when we're wrong We want to get better and better Um I think we're pretty good about trying to listen to every piece of criticism,\n",
"[2402.88s --> 2403.68s] : think it through,\n",
"[2403.76s --> 2405.44s] : internalize what we agree with,\n",
"[2405.52s --> 2409.84s] : but like the breathless clickbate headlines You know,\n",
"[2410.00s --> 2416.32s] : try to let those flow through us What does the open AI moderation tooling for GPT look like?\n",
"[2416.32s --> 2418.00s] : What's the process of moderation?\n",
"[2418.16s --> 2418.80s] : So there's,\n",
"[2418.80s --> 2419.04s] : uh,\n",
"[2419.12s --> 2433.52s] : several things Maybe it's the same thing you can educate me So RLHF is the ranking but is there a wall you are up against like where this is an unsafe thing to answer?\n",
"[2433.76s --> 2435.36s] : What does that tooling look like?\n",
"[2435.52s --> 2438.24s] : We do have systems that try to figure out,\n",
"[2438.40s --> 2438.64s] : you know,\n",
"[2438.80s --> 2442.32s] : try to learn when a question is something that we're supposed to we call refusals,\n",
"[2442.40s --> 2459.84s] : refuse to answer It is early and imperfect ah we're again the spirit of building in public and and bring society along gradually We put something out It's got flaws We'll make better versions But yes,\n",
"[2459.92s --> 2475.68s] : we are trying the system is trying to learn questions that it shouldn't answer One small thing that really bothers me about our current thing and will get this better is I don't like the feeling of being scolded by a computer I really don't,\n",
"[2475.84s --> 2486.24s] : you know I a story that has always stuck with me I don't know if it's true I hope it is is that the reason Steve Jobs put that handle on the back of the first iMac Remember that big plastic,\n",
"[2486.32s --> 2491.28s] : fright colored thing was that you should never trust a computer you shouldn't throw out you couldn't throw out a window?\n",
"[2491.84s --> 2496.64s] : Nice And of course not that many people actually throw their computer out of window,\n",
"[2496.64s --> 2516.56s] : but sort of nice to know that you can And it's nice to know that like this is a tool very much in my control and this is a tool that like does things to help me and I think we've done a pretty good job of that with GPT four but I notice that I have like a this or response to being scolded by a computer,\n",
"[2516.80s --> 2524.32s] : and I think you know that's a good learning from deploying or from creating a system and we can improve it Yes,\n",
"[2524.32s --> 2524.88s] : tricky,\n",
"[2524.88s --> 2529.68s] : and also for the system not to treat you like a child Treating our users like adults is a thing,\n",
"[2529.68s --> 2530.08s] : I say,\n",
"[2530.08s --> 2534.48s] : very frequently inside the office But it's tricky has to do with language,\n",
"[2534.64s --> 2543.84s] : like if there's like certain conspiracy theories you don't want the system to be speaking to it's a very tricky language you should use,\n",
"[2543.92s --> 2564.64s] : because what if I want to understand the earth if the earth is the idea that the earth is flat and I want to fully explore that I want the I want GBD to help me explore GBD four has enough nuance to be able to help you explore that without and treat you like an adult in the process GBT three,\n",
"[2564.64s --> 2564.88s] : I think,\n",
"[2564.88s --> 2569.44s] : just wasn't capable of getting that right But GP four I think we can get to do this By the way,\n",
"[2569.44s --> 2576.88s] : if you could just speak to the leap from GP four two GPT four from three point five from three Is there some technical leaps,\n",
"[2576.88s --> 2579.20s] : or is it really focused on the alignment?\n",
"[2579.44s --> 2579.76s] : No,\n",
"[2579.76s --> 2594.80s] : it's a lot of technical leaps in the base model One of the things we are good at at Open Aye is finding a lot of small winds and multiplying them together And each of them maybe is like a pretty big secret in some sense,\n",
"[2594.96s --> 2604.64s] : but it really is the multiplicative impact of all of them and the detail and care we put into it that gets us these big leaps and then,\n",
"[2604.64s --> 2604.96s] : you know,\n",
"[2605.12s --> 2606.64s] : it looks like to the outside like,\n",
"[2606.64s --> 2606.96s] : oh,\n",
"[2606.96s --> 2614.96s] : they just probably like did one thing to get from three to three point five to four It's like hundreds of complicated things The tiny little thing with the training,\n",
"[2614.96s --> 2615.92s] : with the everything,\n",
"[2615.92s --> 2618.24s] : with the data how we like collect the data,\n",
"[2618.24s --> 2619.12s] : how we clean the data,\n",
"[2619.12s --> 2619.92s] : how we do the training,\n",
"[2619.92s --> 2621.04s] : how we do the optimizer,\n",
"[2621.04s --> 2631.84s] : how we do the architect like so many things Let me ask you the all important question about size So does size matter in terms of neural networks,\n",
"[2631.92s --> 2632.16s] : uh,\n",
"[2632.40s --> 2643.44s] : with how good the system performs so GPT three point five head a hundred seventy five billion I heard GP two at a hundred trillion one hundred trillion Can I speak to this?\n",
"[2643.76s --> 2644.80s] : Do you know that mean?\n",
"[2644.80s --> 2645.12s] : Yeah,\n",
"[2645.12s --> 2662.48s] : the big and purple socks do you know where it originated I don't do I'd be curious to hear the presentation I gave No way Yap Ah Journalists just took a snapshot huh Now I learned from this It's right when GPT three was released I gave a sign you tube,\n",
"[2662.48s --> 2675.76s] : but gave a description of what it is And I spoke to the limitation of the parameters and like where it's going and I talked about the human brain and how many parameters it has on apsis and so on And,\n",
"[2675.76s --> 2676.16s] : um,\n",
"[2676.48s --> 2677.68s] : perhaps like an idiot,\n",
"[2677.68s --> 2680.48s] : perhaps not I said like GPT four,\n",
"[2680.48s --> 2681.04s] : like the next,\n",
"[2681.04s --> 2686.00s] : as it progresses What I should have said is GPT N or something I can't believe that it came from you,\n",
"[2686.08s --> 2690.48s] : that is But people should go to it It's totally taken out of context,\n",
"[2690.56s --> 2691.84s] : they didn't reference anything,\n",
"[2691.84s --> 2698.08s] : they took it This is what GPT four is going to be And I feel horrible about it You know,\n",
"[2698.08s --> 2701.28s] : it doesn't it I don't think it matters in any serious way I mean,\n",
"[2701.28s --> 2702.16s] : it's not good because,\n",
"[2702.16s --> 2702.40s] : uh,\n",
"[2702.48s --> 2702.80s] : again,\n",
"[2702.80s --> 2703.76s] : size is not everything,\n",
"[2703.76s --> 2709.28s] : but also people just take a lot of these kinds of discussions out of context Ah,\n",
"[2709.68s --> 2711.44s] : but it is interesting to comp I mean,\n",
"[2711.44s --> 2712.48s] : that's what I was trying to do,\n",
"[2712.48s --> 2716.40s] : to comp to compare in different ways ah,\n",
"[2716.56s --> 2725.92s] : the difference between the human brain and your network and this thing is getting so impressive This is like in some sense someone said to me this morning actually,\n",
"[2725.92s --> 2726.88s] : and I was like oh,\n",
"[2726.88s --> 2739.44s] : this might be right this is the most complex software object humanity has yet produced and it will be trivial in a couple of decades It's right it'll be like kind of anyone can do it whatever Um but yeah,\n",
"[2739.52s --> 2748.08s] : the amount of complexity relative to anything we've done so far that goes into producing this one set of numbers is quite something Yeah,\n",
"[2748.16s --> 2751.68s] : complexity including the entirety of the history of human civilization,\n",
"[2751.76s --> 2754.56s] : that build up all the different advancements of technology,\n",
"[2754.64s --> 2756.24s] : that build up all the content,\n",
"[2756.32s --> 2766.80s] : the data that Jupiter was trained on that is on the Internet that it's the compression of all of humanity of all the maybe not the experience,\n",
"[2766.80s --> 2768.72s] : all the text output that humanity produces?\n",
"[2768.72s --> 2769.04s] : Yes,\n",
"[2769.04s --> 2778.64s] : it's somewhat different And it's a good question How much if all you have is the Internet data how much can you reconstruct the magic of what it means to be human?\n",
"[2778.88s --> 2789.36s] : I think we'd be surprised how much you can reconstruct But you probably need more better and better and better models But on that topic how much does size matter?\n",
"[2789.44s --> 2790.96s] : By like number of parameters,\n",
"[2790.96s --> 2799.36s] : number of parameters I think people got caught up in the parameter count race in the same way they got caught up in the gigahertz race of processors in like the,\n",
"[2799.36s --> 2799.68s] : you know,\n",
"[2800.08s --> 2802.88s] : nineties and two thousands or whatever You,\n",
"[2802.96s --> 2803.28s] : I think,\n",
"[2803.44s --> 2810.40s] : probably have no idea how many gigahertz the processor in your phone is But what you care about is what the thing can do for you,\n",
"[2810.40s --> 2810.88s] : and there's,\n",
"[2810.88s --> 2811.20s] : you know,\n",
"[2811.20s --> 2815.84s] : different ways to accomplish that You can bump up the clock speed Sometimes that causes other problems,\n",
"[2815.92s --> 2826.00s] : sometimes it's not the best way to get gains Um but I think what matters is getting the best performance and you know,\n",
"[2826.48s --> 2837.84s] : we I think one thing that works well about open eye is we're pretty truth seeking in just doing whatever is going to make the best performance,\n",
"[2838.00s --> 2859.92s] : whether or not it's the most elegant solution So I think like Ella Lemzer sort of hated result in parts of the field Everybody wanted to come up with a more elegant way to get to generalized intelligence and we have been willing to just keep doing what works and looks like it'll keep working So I've spoken with Nochomski,\n",
"[2860.00s --> 2867.36s] : who's been kind of one of the many people that are critical of large language models being able to achieve general intelligence,\n",
"[2867.44s --> 2867.68s] : right?\n",
"[2867.68s --> 2878.80s] : And so it's an interesting question that they've been able to achieve so much incredible stuff Do you think it's possible that large language models really is the way we build AGI?\n",
"[2879.04s --> 2888.24s] : I think it's part of the way I think we need other super important things This is philosophizing a little bit Like what kind of components do you think?\n",
"[2889.92s --> 2897.04s] : In a technical sense or a poetic sense Does it need to have a body that it can experience the world directly?\n",
"[2897.92s --> 2903.28s] : I don't think it needs that But I wouldn't I wouldn't say any of the stuff with certainty,\n",
"[2903.28s --> 2915.76s] : like we're deep into the unknown here For me a system that cannot go significantly add to the sum total of scientific knowledge we have access to kind of discover,\n",
"[2915.84s --> 2941.92s] : invent whatever you want to call it new fundamental science is not a superintelligence and to do that really well I think we will need to expand on the GPT paradigm in pretty important ways that we're still missing ideas for I don't know what those ideas are We're trying to find them I could argue sort of the opposite point that you could have deep,\n",
"[2942.00s --> 2951.68s] : big scientific breakthroughs would just the data that GPT's trained on Like I make some of it like if you prompt it correctly Look,\n",
"[2951.68s --> 2956.88s] : if an oracle told me far from the future that GPT ten turned out to be a true AGI somehow,\n",
"[2957.20s --> 2960.32s] : or maybe just some very small new ideas I would be like,\n",
"[2960.40s --> 2963.92s] : OK I can believe that Not what I would have expected,\n",
"[2963.92s --> 2964.40s] : sitting here,\n",
"[2964.40s --> 2965.60s] : would have set a new big idea,\n",
"[2965.60s --> 2977.76s] : but I can believe that This prompting chain If you extend it very far and then increase at scale the number of those interactions,\n",
"[2977.84s --> 2985.60s] : like what kind of these things start getting integrated into human society and starts building on top of each other I mean,\n",
"[2985.68s --> 2988.32s] : like I don't think we understand what that looks like you said,\n",
"[2988.32s --> 2995.12s] : it's been six days The thing that I am so excited about with this is not that it's a system that kind of goes off and does its own thing,\n",
"[2995.36s --> 3023.76s] : but that it's this tool that humans are using in this feedback loop Helpful for us for a bunch of reasons we get to learn more about trajectories through multiple iterations but I am excited about a world where AI is an extension of human will and amplifier of our abilities and this like you know most useful tool yet created and that is certainly how people are using it And I mean,\n",
"[3023.76s --> 3025.12s] : just like look at Twitter,\n",
"[3025.12s --> 3032.48s] : like the the results are amazing people's like self reported happiness with getting to work with us are great So yeah,\n",
"[3032.56s --> 3035.04s] : like maybe we never build AGI,\n",
"[3035.04s --> 3038.88s] : but we just make humans super great Still a huge one?\n",
"[3039.44s --> 3039.76s] : Yeah,\n",
"[3039.76s --> 3053.76s] : I said I'm part of those people like the amount I derive a lot of happiness from programming together with GPT Part of it is a little bit of terror of can you say more about that?\n",
"[3054.40s --> 3061.44s] : There's a meme I saw today that everybody's freaking out about a sort of GPT taking programmer jobs No,\n",
"[3061.52s --> 3081.52s] : it's the reality is going to be taking like if it's going to take your job it means you're a shitty programmer There's some truth to that Maybe there's some human element that's really fundamental to the creative act to the active genius that is in great design that is involved in programming,\n",
"[3081.60s --> 3090.80s] : and maybe I'm just really impressed by all the boiler plate but that I don't see as boiler plate but is actually a pretty boiler plate Yeah,\n",
"[3090.88s --> 3092.40s] : and maybe that you create like,\n",
"[3092.40s --> 3092.72s] : you know,\n",
"[3092.72s --> 3100.16s] : in a day of programming you have one really important idea Yeah And that's the conchet we're my contribution And there may be like,\n",
"[3100.32s --> 3105.36s] : I think we're going to find So I suspect that is happening with great programmers,\n",
"[3105.36s --> 3108.16s] : and that GPT like models are far away from that one thing,\n",
"[3108.16s --> 3112.08s] : even though they're going to automate a lot of other programming But again,\n",
"[3112.16s --> 3117.12s] : most programmers have some sense of you know,\n",
"[3117.44s --> 3119.68s] : anxiety about what the future's going to look like,\n",
"[3119.68s --> 3126.32s] : but mostly they are like This is amazing I am ten times more productive Don't ever take this away from me There's not a lot of people that use it and say like,\n",
"[3126.40s --> 3127.12s] : Turn this off,\n",
"[3127.20s --> 3127.68s] : you know?\n",
"[3127.92s --> 3129.12s] : Yeah So I I think,\n",
"[3129.12s --> 3129.36s] : uh,\n",
"[3129.44s --> 3130.00s] : so to speak,\n",
"[3130.00s --> 3134.80s] : this the psychology of terror is more like This is awesome This is too awesome,\n",
"[3134.80s --> 3142.56s] : it's too awesome Yeah There is a little bit of coffee tastes too good You know when Kasparov lost to deep blue,\n",
"[3142.64s --> 3152.16s] : somebody said and maybe it was him that liked chess is over now if an AI can beat a human at chess then no one's going to bother to keep playing,\n",
"[3152.16s --> 3152.40s] : right,\n",
"[3152.40s --> 3156.56s] : 'cause like what's the purpose of us or whatever That was thirty years ago,\n",
"[3156.64s --> 3157.52s] : twenty five years ago,\n",
"[3157.52s --> 3166.96s] : something like that I believe that chess has never been more popular than it is right now and people keep wanting to play,\n",
"[3166.96s --> 3168.16s] : and wanting to watch And,\n",
"[3168.24s --> 3168.64s] : by the way,\n",
"[3168.72s --> 3173.52s] : we don't watch two eyes play each other which would be a far better game,\n",
"[3173.60s --> 3174.40s] : in some sense,\n",
"[3174.48s --> 3182.16s] : than whatever else but that's That's not what we choose to do Like,\n",
"[3182.24s --> 3191.52s] : we are somehow much more interested in what humans do in this sense and whether or not Magnus loses to that kid than what happens when two much,\n",
"[3191.52s --> 3193.84s] : much better AI's play each other Well,\n",
"[3194.00s --> 3199.76s] : actually when two AI play each other it's not a better game by our definition of best We just can't understand it No,\n",
"[3199.84s --> 3204.32s] : I think I think they just draw each other I think the human flaws,\n",
"[3204.32s --> 3214.72s] : and this might apply across the spectrum here with AI's will make life way better but we'll still want drama We still want imperfection and flaws,\n",
"[3214.72s --> 3219.76s] : and AI will not have us much of that Look I mean I hate to sound like Utopic techbro here,\n",
"[3219.76s --> 3221.84s] : but if you'll excuse me for three seconds,\n",
"[3221.92s --> 3234.00s] : like the the level of the increase in quality of life that AI can deliver is extraordinary We can make the world amazing,\n",
"[3234.16s --> 3235.92s] : we can make people's lives amazing,\n",
"[3235.92s --> 3237.04s] : we can cure diseases,\n",
"[3237.04s --> 3238.48s] : we can increase material wealth,\n",
"[3238.56s --> 3240.24s] : we can like help people be happier,\n",
"[3240.24s --> 3244.80s] : more fulfilled all these sorts of things And then people are like,\n",
"[3244.80s --> 3245.04s] : Oh,\n",
"[3245.04s --> 3245.28s] : well,\n",
"[3245.28s --> 3250.80s] : no one is going to work but People want status People want drama,\n",
"[3250.80s --> 3251.84s] : people want new things,\n",
"[3251.84s --> 3252.80s] : people want to create,\n",
"[3252.88s --> 3253.76s] : people want to like,\n",
"[3253.84s --> 3255.68s] : feel useful Um,\n",
"[3255.92s --> 3257.68s] : people want to do all these things,\n",
"[3257.76s --> 3262.40s] : and we're just going to find new and different ways to do them even in a vastly better,\n",
"[3262.48s --> 3272.96s] : like unimaginably good standard of living world But that world the positive trajectories with AI that world is with an AI that's aligned with humans,\n",
"[3273.20s --> 3273.92s] : doesn't hurt,\n",
"[3273.92s --> 3274.72s] : doesn't limit,\n",
"[3274.72s --> 3284.08s] : doesn't doesn't try to get rid of humans And there's some folks who consider all the different problems with a superintelligent AI system,\n",
"[3284.08s --> 3294.56s] : so one of them is Eleazary He warns that Ahi will likely kill all humans and there's a bunch of different cases,\n",
"[3294.64s --> 3307.36s] : but I think one way to summarize it is that it's almost impossible to keep AI aligned as it becomes superintelligent Can you steel man the case for that?\n",
"[3307.44s --> 3313.04s] : and to what degree do you disagree with that trajectory?\n",
"[3313.92s --> 3314.16s] : So,\n",
"[3314.16s --> 3314.80s] : first of all,\n",
"[3314.80s --> 3318.56s] : I'll say I think that there is some chance of that,\n",
"[3318.64s --> 3320.56s] : and it's really important to acknowledge it,\n",
"[3320.56s --> 3323.44s] : because if we don't talk about it if we don't treat it as potentially real,\n",
"[3323.52s --> 3332.24s] : we won't put enough effort into solving it And I think we do have to discover new techniques to be able to solve it Um,\n",
"[3332.40s --> 3335.76s] : I think a lot of the predictions this is true for any new field,\n",
"[3335.84s --> 3339.76s] : but a lot of the predictions about AI in terms of capabilities,\n",
"[3340.32s --> 3347.28s] : in terms of what the safety challenges and the easy parts are going to be have turned out to be wrong,\n",
"[3347.76s --> 3355.28s] : the only way I know how to solve a problem like this is iterating our way through it,\n",
"[3355.68s --> 3365.76s] : learning early and limiting the number of one shot to get it right scenarios that we have To Steelman Well,\n",
"[3365.76s --> 3369.68s] : there's I can't just pick like one AI safety case or AI alignment case,\n",
"[3369.68s --> 3392.16s] : but I think Eliezer wrote a really great blog post I think some of his work has been sort of somewhat difficult to follow or had what I view is like quite significant logical flaws but he wrote this one blog post outlining why he believed that alignment was such a hard problem that I thought was again don't agree with a lot of it,\n",
"[3392.24s --> 3396.64s] : but well reasoned and thoughtful and very worth reading So I think I'd point people to that,\n",
"[3396.72s --> 3397.68s] : is the Steelman?\n",
"[3397.92s --> 3398.24s] : Yeah,\n",
"[3398.24s --> 3422.96s] : and I'll also have a conversation with him Um There is some aspect and I'm torn here because it's difficult to reason about the exponential improvement of technology But also I've seen time and time again how transparent and iterative trying out as you improve the technology,\n",
"[3422.96s --> 3423.60s] : trying it out,\n",
"[3423.68s --> 3424.40s] : releasing it,\n",
"[3424.56s --> 3425.36s] : testing it,\n",
"[3425.60s --> 3426.80s] : how that can,\n",
"[3426.80s --> 3434.48s] : ah improve your understanding of the technology in such that the philosophy of how to do,\n",
"[3434.56s --> 3434.96s] : for example,\n",
"[3435.04s --> 3454.00s] : safety of any kind of technology but AI safety gets adjusted over time rapidly A lot of the formative AI safety work was done before people even believed in deep learning and and certainly before people believed in large language models And I don't think it's like updated enough given everything we've learned now,\n",
"[3454.08s --> 3461.28s] : and everything we will learn going forward So I think it's got to be this very tight feedback loop I think the theory does play a real role,\n",
"[3461.36s --> 3466.64s] : of course But continuing to learn what we learn from how the technology trajectory goes?\n",
"[3468.24s --> 3478.56s] : Is quite important I think now is a very good time and we're trying to figure out how to do this to significantly ramp up technical alignment work I think we have new tools,\n",
"[3478.64s --> 3488.08s] : we have new understanding ah and there's a lot of work that's important to do that we can do now So one of the main concerns here is,\n",
"[3488.08s --> 3488.40s] : uh,\n",
"[3488.48s --> 3497.76s] : something called AI take off or fast take off that the exponential improvement would be really fast to where like in days?\n",
"[3497.92s --> 3498.48s] : In days,\n",
"[3498.56s --> 3524.96s] : yeah mean there's this is an this is a pretty serious at least to me it's become more of a serious concern Just how amazing Chad GPT turned out to be and then the improvement of GBT four Almost like to where it surprised everyone seemingly you can correct me including you So GBT Four has not surprised me at all in terms of reception there Chat GT surprised us a little bit,\n",
"[3525.04s --> 3530.56s] : but I still was like advocating that we'd do it 'cause I thought it was going to do really great Um So like you know,\n",
"[3530.64s --> 3541.84s] : maybe I thought it would have been like The tenth fastest growing product in history and not the number one fastest I like okay You know,\n",
"[3541.84s --> 3546.56s] : I think it's like hard you should never kind of assume something's going to be like the most successful product launch ever Um,\n",
"[3546.64s --> 3554.72s] : but we thought it was at least many of us thought it was going to be really good GVD four has weirdly not been that much of an update for most people You know,\n",
"[3554.72s --> 3555.20s] : they're like,\n",
"[3555.20s --> 3555.44s] : Oh,\n",
"[3555.44s --> 3556.72s] : it's better than three point five,\n",
"[3556.72s --> 3559.36s] : but I thought it was going to be better than three point five and it's cool,\n",
"[3559.44s --> 3559.60s] : but,\n",
"[3559.60s --> 3559.92s] : you know,\n",
"[3559.92s --> 3570.08s] : this is like Some one said to me over the week end You shipped an AGI and I somehow like am just going about my daily life,\n",
"[3570.08s --> 3575.12s] : and I'm not that impressed And I obviously don't think we shipped an AGI,\n",
"[3575.60s --> 3584.32s] : but I get the point and the world is continuing on When you build or somebody builds an artificial general intelligence,\n",
"[3584.32s --> 3585.84s] : would that be fast or slow?\n",
"[3585.84s --> 3588.72s] : Would we know what's happening or not?\n",
"[3588.96s --> 3591.92s] : Would we go about our day on the week end or not?\n",
"[3592.24s --> 3595.36s] : So I'll come back to whether would we go about our day or not Fang,\n",
"[3595.36s --> 3603.36s] : I think there's like a bunch of interesting lessons from Covid and the UFO videos and a whole bunch of other stuff that we can talk to there But on the take off question,\n",
"[3603.52s --> 3608.00s] : if we imagine a two by two matrix of short time lines till AGI starts,\n",
"[3608.32s --> 3610.24s] : long time lines till AGI starts,\n",
"[3610.32s --> 3611.04s] : slow take off,\n",
"[3611.04s --> 3612.00s] : fast take off,\n",
"[3612.16s --> 3615.60s] : you have an instinct on what you think the safest quadrant would be?\n",
"[3615.76s --> 3622.40s] : So the different options are like year Yep say the take off the we start the take off period,\n",
"[3622.40s --> 3622.88s] : yep,\n",
"[3623.12s --> 3624.64s] : next year or in twenty years,\n",
"[3624.64s --> 3629.28s] : three years and then it takes one year or ten years Well,\n",
"[3629.28s --> 3630.80s] : you can even say one year or five years,\n",
"[3630.96s --> 3631.84s] : whatever you want,\n",
"[3632.00s --> 3647.36s] : for the take off I feel like now is safer So do I So I'm in longer though I'm in these slow take off short time lines is the most likely good world,\n",
"[3647.36s --> 3658.08s] : and we optimize the company to have maximum impact in that world to try to push for that kind of a world And the decisions that we make are you know,\n",
"[3658.08s --> 3659.68s] : there's like probability masses,\n",
"[3659.68s --> 3671.60s] : but weighted towards that And I think I'm very afraid of the fast take offs I think in the longer time lines it's harder to have a slow take off there's a bunch of other problems,\n",
"[3671.60s --> 3676.08s] : too But that's what we're trying to do Do you think GP four is an AGI?\n",
"[3678.48s --> 3699.92s] : I think if it is just like with the UFO videos we wouldn't know immediately I think it's actually hard to know that I've been thinking I've been playing with GPT for and thinking how would I know if it's an AGI or not?\n",
"[3700.24s --> 3702.64s] : Because I think in terms of,\n",
"[3702.88s --> 3704.48s] : to put it in a different way,\n",
"[3705.68s --> 3711.44s] : how much of AGI is the interface I have with the thing and how much of it,\n",
"[3711.52s --> 3711.84s] : er,\n",
"[3711.92s --> 3722.00s] : is the actual wisdom inside of it Like a part of me thinks that you can have a model that's capable of superintelligence,\n",
"[3722.24s --> 3726.48s] : and it just hasn't been quite unlocked What I saw with Chatjee Petit,\n",
"[3726.48s --> 3729.36s] : just doing that little bit of our L with human feedback,\n",
"[3729.52s --> 3732.00s] : makes the thing some much more impressive,\n",
"[3732.08s --> 3734.88s] : much more usable So maybe if you have a few more tricks,\n",
"[3734.88s --> 3735.28s] : like you said,\n",
"[3735.28s --> 3737.44s] : there's like hundreds of tricks inside open AI,\n",
"[3737.44s --> 3743.52s] : a few more tricks and all sudden holy shit This thing So I think that GPT four,\n",
"[3743.52s --> 3744.80s] : although quite impressive,\n",
"[3744.80s --> 3747.84s] : is definitely not in AGI but isn't it remarkable we're having this debate?\n",
"[3748.24s --> 3750.72s] : Yeah So what's your intuition why is that?\n",
"[3751.52s --> 3757.68s] : I think we're getting into the phase where specific definitions of AGI really matter Or we just say,\n",
"[3757.76s --> 3758.00s] : you know,\n",
"[3758.00s --> 3784.08s] : I know it when I see it and I'm not even going to bother with the definition But under the I know it when I see it it doesn't feel that close to me like if if I were reading a sci fi book and there was a character that was an AGI and that character was GPT four I'll be likewise as a shitty book You know,\n",
"[3784.08s --> 3795.28s] : that's not very cool I wish I would have hoped we had done better To me some of the human factors are important here Do you think GPT four is conscious?\n",
"[3795.92s --> 3797.04s] : I think no,\n",
"[3797.20s --> 3799.28s] : but I asked GPT four,\n",
"[3799.28s --> 3805.12s] : and of course it says no Do you think GPT four is conscious?\n",
"[3806.48s --> 3812.48s] : I think it knows how to fake consciousness Yes How to fake consciousness?\n",
"[3812.56s --> 3818.08s] : Yeah If if if you provide the right interface and the right prompts?\n",
"[3818.32s --> 3827.68s] : It definitely can answer as if it were Yeah And then it starts getting weird Like what is the difference between pretending to be conscious and conscious?\n",
"[3827.76s --> 3828.08s] : I mean,\n",
"[3828.56s --> 3847.92s] : you don't know obviously we can go to like the freshman year dorm late at Saturday night kind of thing You don't know that you're not a GP four rollout in some advanced simulation yet Yes So if we're willing to go to that level sure I'm going to live in that but that's an important that's an important level that's an important,\n",
"[3847.92s --> 3857.12s] : ah that's a really important level because one of the things that makes it not conscious is declaring that it's a computer program,\n",
"[3857.12s --> 3871.20s] : therefore it can't be a conscious so I'm not going to I'm not even going to acknowledge it but that just puts it in the category of other I believe AI can be conscious So then the question is,\n",
"[3871.36s --> 3873.68s] : what would it look like when it's conscious?\n",
"[3874.00s --> 3875.68s] : What would it behave like?\n",
"[3875.92s --> 3879.68s] : And it would probably say things like first of all,\n",
"[3879.76s --> 3900.40s] : I am conscious second of all display capability of suffering An understanding of self Of having some memory of itself and maybe interactions with you maybe there's a personalization aspect to it,\n",
"[3900.48s --> 3904.08s] : and I think all of those capabilities are interface capabilities,\n",
"[3904.16s --> 3911.12s] : not fundamental aspects of the actual knowledge solid in your own net Maybe I can just share a few like disconnected thoughts here,\n",
"[3911.12s --> 3914.32s] : Jeff But I'll tell you something that Ilias said to me once,\n",
"[3914.32s --> 3919.52s] : a long time ago that has like stuck in my head Ilias let's gover yes,\n",
"[3919.52s --> 3920.24s] : my co founder,\n",
"[3920.32s --> 3921.92s] : the chief scientist of open eye,\n",
"[3921.92s --> 3932.16s] : and sort of legend in the field Um We are talking about how you would know if a model were conscious or not And heard many ideas thrown around,\n",
"[3932.16s --> 3966.72s] : but he said one that I think is interesting if you trained a model on a data set that you were extremely careful to have no mentions of consciousness or anything close to it in the training process Like not only was the word never there but nothing about the sort of subjective experience of it or related concepts And then you started talking to that model about here are some things that you weren't trained about,\n",
"[3966.80s --> 3980.16s] : and for most of them the model was like I've no idea what you're talking about But then you asked it you sort of described the experience the subjective experience of consciousness and the model immediately responded,\n",
"[3980.16s --> 3981.60s] : unlike the other questions Yes,\n",
"[3981.60s --> 3989.76s] : I know exactly what you're talking about That would update me somewhat I don't know,\n",
"[3989.84s --> 4007.44s] : because that's more in the space of facts versus like emotions I don't think consciousness is an emotion I think consciousness is an ability to sort of experience this world really deeply There's a movie called Ex Mock and art of it,\n",
"[4007.44s --> 4009.04s] : but I haven't seen it You haven't seen it?\n",
"[4009.04s --> 4009.36s] : No?\n",
"[4009.68s --> 4016.16s] : The director Alex Garland had a conversation said it's where AGI system is built,\n",
"[4016.24s --> 4020.24s] : embodied in the body of a a woman and,\n",
"[4020.24s --> 4020.48s] : uh,\n",
"[4020.64s --> 4022.24s] : something he doesn't make explicit,\n",
"[4022.24s --> 4036.08s] : but he said he put in the movie without describing why But at the end of the movie spoiler alert when the AI escapes the woman escapes Er,\n",
"[4036.32s --> 4047.92s] : she smiles for nobody For no audience She smiles at the like at the freedom she is experiencing experiencing,\n",
"[4047.92s --> 4077.84s] : I don't know anthropomorphizing but he said the smile to me was the was passing the touring test in consciousness that you smiled for no audience You smiled for yourself As an interesting thought it's like you taken an experience for experience' sake I don't know That seemed more like consciousness versus the ability to convince somebody else that you are conscious And that feels more like a realm of emotion versus facts But yes,\n",
"[4078.00s --> 4088.80s] : if it knows I think there's many other tasks tests like that that we could look at too But,\n",
"[4088.80s --> 4089.12s] : you know,\n",
"[4089.20s --> 4103.44s] : my personal beliefs consciousness is of something very strange is going on Do you think it's attached to the particular medium of of the human brain?\n",
"[4103.44s --> 4105.76s] : Do you think an air can be conscious?\n",
"[4106.56s --> 4111.68s] : I'm certainly willing to believe that consciousness is somehow the fundamental substrate,\n",
"[4111.68s --> 4136.80s] : and we're all just in the dream or the simulation or whatever I think it's interesting how much sort of the Silicon Valley religion of the simulation has gotten close to like Brummen and how little space there is between them but from these very different directions So like maybe that's what's going on But if it is like physical reality as we understand it and all of the rules of the game where we think they are,\n",
"[4137.04s --> 4141.60s] : then there's something I still think it's something very strange Ah,\n",
"[4141.60s --> 4151.92s] : just to linger on the alignment problem a little bit maybe the control problem What are the different ways you think AGI might go wrong that concern you?\n",
"[4152.00s --> 4158.88s] : You said that a fear a little bit of fear is very appropriate here We've been very transparent,\n",
"[4158.96s --> 4160.32s] : Bob being mostly excited,\n",
"[4160.32s --> 4165.36s] : but also scared I think it's weird when people like think it's like a big dunk that I say like I'm a little bit afraid,\n",
"[4165.36s --> 4176.48s] : and I think it would be crazy not to be a little bit afraid And I empathize with people who are a lot afraid What do you think about that moment of a system becoming superintelligent?\n",
"[4176.48s --> 4177.92s] : Do you think you would know?\n",
"[4179.36s --> 4198.80s] : The current worries that I have are that there are going to be disinformation problems or economic shocks or something else at a level far beyond anything we're prepared for And that doesn't require superintelligence,\n",
"[4198.80s --> 4210.48s] : that doesn't require a super deep alignment problem in the machine waking up and trying to deceive us and I don't think that gets enough attention men starting to get more,\n",
"[4210.48s --> 4220.56s] : I guess So these systems deployed at scale can the shift the woods of geopolitics and so on How would we know if,\n",
"[4220.56s --> 4221.52s] : like on Twitter,\n",
"[4221.52s --> 4235.20s] : we were mostly having like LLM's direct the whatever's flowing through that hive mind Yep On Twitter and then perhaps beyond And then as on Twitter,\n",
"[4235.28s --> 4237.92s] : so everywhere else eventually Yes,\n",
"[4237.92s --> 4238.96s] : how would we know?\n",
"[4239.28s --> 4246.40s] : My statement is we wouldn't And that's a real danger How do you prevent that danger?\n",
"[4246.56s --> 4279.28s] : I think there's a lot of things you can try But at this point it is a certainty there are soon going to be a lot of capable open source LLMs with very few to none no safety controls on them And so you can try with regulatory approaches you can try with using more powerful AI to detect this stuff happening I'd like us to start trying a lot of things very soon How do you under this pressure that there's going to be a lot of open source,\n",
"[4279.44s --> 4282.32s] : there's going to be a lot of large language models?\n",
"[4282.56s --> 4287.44s] : Under this pressure how do you continue prioritizing safety?\n",
"[4287.60s --> 4288.24s] : Whereas,\n",
"[4288.56s --> 4288.88s] : I mean,\n",
"[4288.88s --> 4290.16s] : there's several pressures,\n",
"[4290.24s --> 4294.16s] : so one of them is a market driven pressure from other companies,\n",
"[4294.24s --> 4294.72s] : ah,\n",
"[4295.12s --> 4296.08s] : Google,\n",
"[4296.40s --> 4297.04s] : Apple,\n",
"[4297.12s --> 4301.04s] : Meta and smaller companies How do you resist the pressure from that?\n",
"[4301.12s --> 4302.80s] : Or how do you navigate that pressure?\n",
"[4302.96s --> 4305.60s] : You stick with what you believe and you stick to your mission,\n",
"[4305.68s --> 4320.08s] : you know I'm sure people will get ahead of us in all sorts of ways and take shortcuts we're not going to take And we just aren't going to do that Huddy I'll compete them I think there's going to be many AGI's in the world,\n",
"[4320.08s --> 4336.16s] : so we don't have to like out compete every one We're going to contribute one Other people are going to contribute some I think up I think multiple AGIs in the world with some differences in how they're built and what they do and what they're focused on I think that's good Um,\n",
"[4336.40s --> 4338.32s] : we have a very unusual structure,\n",
"[4338.40s --> 4343.52s] : so we don't have this incentive to capture unlimited value I worry about the people who do,\n",
"[4343.52s --> 4343.76s] : but,\n",
"[4343.76s --> 4344.00s] : you know,\n",
"[4344.00s --> 4366.88s] : hopefully it's all going to work out But we're a weird Org and we're good at resisting private We have been a misunderstood and badly mocked Org for a long time Like when we started we like announced the Org at the end of twenty fifteen said we were going to work on AGI Like people thought we were batchet and sane You know,\n",
"[4366.88s --> 4378.24s] : like I I remember at the time a eminent AI scientist at a large industrial AI lab was like demming individual reporters,\n",
"[4378.40s --> 4379.12s] : being like,\n",
"[4379.28s --> 4379.60s] : You know,\n",
"[4379.68s --> 4384.48s] : these people aren't very good and it's ridiculous to talk about AGI and I can't believe you're giving them time of day,\n",
"[4384.48s --> 4385.12s] : and it's like,\n",
"[4385.28s --> 4386.32s] : that was the level of,\n",
"[4386.32s --> 4386.56s] : like,\n",
"[4386.80s --> 4399.84s] : pettiness and rancor in the field that a new group of people saying we're going to try to build AGI So open a eye and deep mind was a small collection of folks who were brave enough to talk about AGI um,\n",
"[4400.24s --> 4403.84s] : in the face of mockery We don't get mocked as much now?\n",
"[4404.32s --> 4406.96s] : Don't get mocked as much now Ah,\n",
"[4406.96s --> 4416.32s] : so I was speaking about the structure of the of the of the org so open AI went,\n",
"[4416.32s --> 4416.88s] : um,\n",
"[4417.20s --> 4421.68s] : stop being nonprofit or split up in a tweet Can you describe that whole process?\n",
"[4421.68s --> 4441.36s] : Yes Stan We started as a nonprofit We learned early on that we were going to need far more capital than we were able to raise as a nonprofit Our nonprofit is still fully in charge There is a subsidiary capped profit so that our investors and employees can earn a certain fixed return And then beyond that,\n",
"[4441.52s --> 4443.20s] : everything else flows to the non profit,\n",
"[4443.20s --> 4448.56s] : and the non profit is like invoking control lets us make a bunch of non standard decisions,\n",
"[4448.88s --> 4449.20s] : um,\n",
"[4449.28s --> 4454.32s] : can cancel equity can do a whole bunch of other things can let us merge with another organ,\n",
"[4454.56s --> 4454.96s] : um,\n",
"[4455.76s --> 4461.84s] : protects us from making decisions that are not in any like shareholders' interest Ah,\n",
"[4462.08s --> 4474.40s] : so I think it's a structure that has been important to a lot of the decisions we've made What went into that decision process for taking a leap from non profit to capped for profit?\n",
"[4475.28s --> 4477.52s] : What are the pros and cas you were deciding at the time?\n",
"[4477.52s --> 4477.76s] : I mean,\n",
"[4477.84s --> 4491.60s] : this was it was nineteen it was really like To do what we needed to go do we had tried and failed enough to raise the money as a non profit We didn't see a path forward there So we needed some of the benefits of capitalism,\n",
"[4491.68s --> 4494.40s] : but not too much I remember at the time someone said,\n",
"[4494.40s --> 4494.64s] : you know,\n",
"[4494.64s --> 4495.44s] : as a non profit,\n",
"[4495.44s --> 4498.56s] : not enough will happen as a for profit too much will happen,\n",
"[4498.64s --> 4513.84s] : so we need this sort of stranger immediate You kind of had this off hand comment of you worry about the uncapped companies that play with AGI Can you elaborate on the worry here?\n",
"[4513.84s --> 4515.04s] : Because AGI,\n",
"[4515.12s --> 4518.32s] : out of all the technologies we have in our hands,\n",
"[4518.32s --> 4524.72s] : has the potential to make is a cap is a hundred x for open AI It started that it's much,\n",
"[4524.72s --> 4527.60s] : much lower for like new investors now You know,\n",
"[4527.68s --> 4530.08s] : AGI can make a lot more than a hundred x,\n",
"[4530.16s --> 4532.08s] : for sure And so how do you,\n",
"[4532.16s --> 4536.40s] : um like how do you compete like stepping outside of open A,\n",
"[4536.48s --> 4539.28s] : how do you look at a world where Google is playing,\n",
"[4539.52s --> 4542.96s] : where Apple and Dies and Meta are playing?\n",
"[4543.20s --> 4552.24s] : We can't control what other people are going to do We can try to like build something and talk about it and influence others and provide value and,\n",
"[4552.24s --> 4552.56s] : you know,\n",
"[4552.64s --> 4571.28s] : good systems for the world But they're going to do what they're going to do Now I I think right now there's like extremely fast and not super deliberate motion inside some of these companies But already I think people are,\n",
"[4571.36s --> 4572.32s] : as they see,\n",
"[4572.80s --> 4577.92s] : the rate of progress Already people are grappling with what's at stake here,\n",
"[4578.00s --> 4586.00s] : and I think the better angels are going to win out Can you elaborate on the better angels of individuals the individuals and companies,\n",
"[4586.08s --> 4591.28s] : but you know the incentives of capitalism to create and capture unlimited value?\n",
"[4592.00s --> 4594.72s] : I'm a little afraid of but again,\n",
"[4594.80s --> 4595.04s] : no,\n",
"[4595.04s --> 4602.08s] : I think no one wants to destroy the world No one looks up saying like today I want to destroy the world So we've got the Molloch problem On the other hand,\n",
"[4602.16s --> 4603.84s] : we've got people who are very aware of that,\n",
"[4603.92s --> 4614.72s] : and I think a lot of healthy conversation about how can we collaborate to minimise some of these very scary downsides Well,\n",
"[4614.88s --> 4616.40s] : nobody wants to destroy the world,\n",
"[4616.40s --> 4618.00s] : and may ask you a tough question,\n",
"[4618.08s --> 4629.76s] : so you are very likely to be one of not the person that creates AGI One of one of Any of his men,\n",
"[4629.76s --> 4631.28s] : like we're on a team of many,\n",
"[4631.36s --> 4650.88s] : yet there'll be many teams But several small number of people nevertheless relative I do think it's strange that it's maybe a few tens of thousands of people in the world few thousands beyond the world But there will be a room with a few folks who are like holy shit That happens more often than you would think now I understand,\n",
"[4650.88s --> 4653.84s] : I understand this I understand this Yes,\n",
"[4653.92s --> 4655.28s] : there will be more Suchrooms,\n",
"[4655.28s --> 4657.92s] : which is a beautiful place to be in the world,\n",
"[4658.00s --> 4658.32s] : ah,\n",
"[4658.40s --> 4659.12s] : terrifying,\n",
"[4659.12s --> 4670.16s] : but mostly beautiful So that might make you and a handful of folks the most powerful humans on earth Do you worry that power might corrupt you?\n",
"[4670.64s --> 4672.48s] : For sure Um look,\n",
"[4672.48s --> 4704.96s] : I don't I think you want decisions about this technology and certainly decisions about who is running this technology to become increasingly democratic over time We haven't figured out quite how to do this but part of the reason for deploying like this is to get the world to have time to adapt and to reflect and to think about this to pass regulation for institutions to come up with new norms,\n",
"[4705.12s --> 4709.52s] : for the people working out together like That is a huge part of why we deploy,\n",
"[4709.84s --> 4713.52s] : even though many of they say to people you referenced earlier think it's really bad,\n",
"[4713.60s --> 4733.60s] : even they acknowledge that this is of some benefit But I think any version of one person is in control of this is really bad So try to distribute the power somewhere I don't have and I don't want,\n",
"[4733.68s --> 4733.84s] : like,\n",
"[4733.84s --> 4736.32s] : any like super voting power or any special like them,\n",
"[4736.32s --> 4736.56s] : you know,\n",
"[4736.88s --> 4744.24s] : no like control of the board or anything like that of opening eye But AGI,\n",
"[4744.24s --> 4744.96s] : if graded,\n",
"[4745.04s --> 4747.92s] : has a lot of power How do you think we're doing like honest,\n",
"[4747.92s --> 4749.04s] : how do you think we're doing so far?\n",
"[4749.04s --> 4752.32s] : Like how do you think our decisions are like do you think we're making things not better worse?\n",
"[4752.40s --> 4753.52s] : What can we do better?\n",
"[4753.68s --> 4753.92s] : Well,\n",
"[4754.00s --> 4759.44s] : the things I really like because I know a lot of folks at open AI The thing I really like is the transparency,\n",
"[4759.44s --> 4760.32s] : everything you're saying,\n",
"[4760.32s --> 4762.40s] : which is like failing publicly,\n",
"[4762.96s --> 4764.32s] : writing papers,\n",
"[4764.80s --> 4778.88s] : releasing different kinds of information about the safety concerns involved and doing it out in the open is great because especially in contrast to some other companies that are not doing that,\n",
"[4778.96s --> 4781.92s] : they're being more closed That said,\n",
"[4781.92s --> 4786.40s] : you could be more open Do you think we should open source GP four?\n",
"[4790.88s --> 4797.44s] : My personal opinion because I know people at open AI is no What does knowing the people that open AI have to do with it?\n",
"[4797.68s --> 4804.64s] : Because I know they're good people I know a lot of people I know they are good human beings From a perspective of people that don't know the human beings,\n",
"[4804.64s --> 4810.80s] : there's a concern of the super powerful technology in the hands of a few that's closed It's closed in some sense,\n",
"[4810.80s --> 4813.92s] : but we give more access to it Yes Then like,\n",
"[4814.00s --> 4828.56s] : if if this had just been Google's game I feel it's very unlikely that anyone would have put this API out There's PR risk with it I get personal threats because of it all the time I think most companies wouldn't have done this So maybe we didn't go as open as people wanted,\n",
"[4828.64s --> 4841.52s] : but like we've distributed it pretty broadly You personally in opening eyes of culture is not so like nervous about a PR risk and all that kind of stuff You're more nervous about the risk of the actual technology,\n",
"[4842.40s --> 4844.96s] : and you reveal that So I you know,\n",
"[4845.04s --> 4858.24s] : the nervousness that people have is because it's such early days of the technology is that you will close off over times becomes more and more powerful My nervousness is you get attacked so much by fear mongering click bake journalism that you're like,\n",
"[4858.24s --> 4859.68s] : Why the hell do I need to deal with this?\n",
"[4859.68s --> 4863.28s] : I think the clickbate journalism bothers me more than it bothers me No,\n",
"[4863.28s --> 4872.80s] : I'm a third person bothered Like I appreciate that I feel all right about it All the things I lose sleep over it's not high in the list Because it's important there's a handful of companies,\n",
"[4872.80s --> 4874.72s] : a handful of folks that are really pushing this forward,\n",
"[4874.72s --> 4885.12s] : there are amazing folks that don't want them to become cynical about the rest the rest of the world I think people at open eye feel the weight of responsibility of what we're doing,\n",
"[4885.20s --> 4885.84s] : and yeah,\n",
"[4885.84s --> 4886.80s] : it would be nice if,\n",
"[4886.88s --> 4887.84s] : like you know,\n",
"[4887.92s --> 4901.20s] : journalists were nicer to us and Twitter trolls gave us more benefit of the doubt but like I think we have a lot of resolve in what we're doing and why and the importance of it But I really would,\n",
"[4901.20s --> 4901.52s] : love,\n",
"[4901.52s --> 4902.88s] : and I ask this like of a lot of people,\n",
"[4902.96s --> 4914.32s] : not just if cameras roll in like any feedback you've got for how we can be doing better We're in uncharted waters here Talking to smart people is how we figure out what to do better How d'ye take feedback d'ye take feedback von Twitter also?\n",
"[4914.40s --> 4916.40s] : D'you guess does the sea the water?\n",
"[4916.56s --> 4928.16s] : All Twitter is unreadable Yeah So sometimes I do I can like take a sample a cup cup out of the waterfall But I mostly take it from conversations like this Speaking of feedback,\n",
"[4928.16s --> 4929.36s] : somebody you know well,\n",
"[4929.36s --> 4941.36s] : you've worked together closely on some of the ideas behind Open Eye's Elam Musk You have agreed on a lot of things You've disagreed on some things What have been some interesting things you've agreed and disagreed on,\n",
"[4941.60s --> 4944.72s] : speaking of a fun debate on Twitter,\n",
"[4945.04s --> 4955.76s] : I think we agree on the magnitude of the downside of AGI and the need to get not only safety right,\n",
"[4956.00s --> 4968.88s] : but get to a world where people are much better off because AGI exists than if AGI had never been built What do you disagree on?\n",
"[4969.76s --> 4974.80s] : Elon is obviously attacking us some on Twitter right now on a few different vectors,\n",
"[4974.80s --> 4986.88s] : and I have empathy because I believe he is understandably so really stressed about AGI safety I'm sure there are some other motivations going on too,\n",
"[4986.96s --> 4998.72s] : but that's definitely one of them Um I saw this video of Elon a long time ago talking about space X,\n",
"[4998.72s --> 5016.48s] : maybe it's on some new show and a lot of early pioneers in space were really bashing space X and maybe elon to and he was visibly very hurt by that,\n",
"[5016.48s --> 5018.40s] : and said You know,\n",
"[5018.48s --> 5020.00s] : those guys are heroes of mine,\n",
"[5020.00s --> 5028.96s] : and I sucks and I wish they would see how hard we're trying I definitely grew up with Elon as a hero of mine Um You know,\n",
"[5029.04s --> 5048.56s] : despite him being a jerk on Twitter or whatever I'm happy he exists in the world But I wish he would do more to look at the hard work we're doing to get this stuff right A little bit more love What do you admire in the name of love about Elmusk?\n",
"[5048.88s --> 5075.84s] : I mean so much right like he has he has driven the world forward in important ways I think we will get to electric vehicles much faster than we would have if he didn't exist I think we'll get to space much faster than we would have if he didn't exist and as a sort of like citizen of the world I'm very appreciative of that Also like being a jerk on Twitter's side,\n",
"[5076.00s --> 5087.36s] : in many instances he's like a very funny and warm guy And some of the jerk on Twitter thing as a fan of humanity laid out in its full complexity and beauty,\n",
"[5087.36s --> 5094.72s] : I enjoy the tension of ideas expressed So you know I earlier said to admire how transparent you are,\n",
"[5094.88s --> 5102.00s] : but I like how the battles are happening before our eyes are supposed to everybody closing off inside boardrooms it's all You know,\n",
"[5102.00s --> 5103.04s] : maybe I should hit back,\n",
"[5103.04s --> 5104.32s] : and maybe some day I will,\n",
"[5104.32s --> 5108.56s] : but it's not like my normal style It's all fascinating to watch,\n",
"[5108.56s --> 5112.40s] : and I think both of you are brilliant people,\n",
"[5112.48s --> 5116.24s] : and have early on for a long time really cared about AGI,\n",
"[5116.32s --> 5118.40s] : and had had great concerns about AGI,\n",
"[5118.40s --> 5119.92s] : but a great hope for AGI,\n",
"[5119.92s --> 5121.92s] : and that's cool to see,\n",
"[5122.08s --> 5124.96s] : these big minds having those discussions,\n",
"[5125.28s --> 5134.96s] : even if they're tense at times I think it was Elon that said that GPT is too woke Is GPT too woke?\n",
"[5135.44s --> 5137.60s] : Can you still mend the case that it is and not?\n",
"[5137.68s --> 5140.96s] : This is Godors question about bias?\n",
"[5141.04s --> 5141.60s] : Honestly,\n",
"[5141.60s --> 5144.24s] : I barely know what woke means any more I did for a while,\n",
"[5144.24s --> 5145.60s] : and I feel like the word is morphed,\n",
"[5145.60s --> 5167.36s] : so I will say I think it was too biased and will always be There will be no one version of GPT that the world ever agrees is unbiased What I think is we've made a lot like again even some of our harshest critics have gone off and been tweeting about three point five to four comparisons and being like,\n",
"[5167.44s --> 5167.68s] : Wow,\n",
"[5167.68s --> 5171.76s] : these people really got a lot better Not that they don't have more work to do and we certainly do,\n",
"[5171.76s --> 5187.76s] : but I I appreciate critics who display intellectual honesty like that And there's been more of that than I would have thought We will try to get the default version to be as neutral as possible,\n",
"[5187.76s --> 5195.60s] : but as neutral as possible is not that neutral if you have to do it again for more than one person And so this is where more steerability,\n",
"[5195.60s --> 5197.12s] : more control in the hands of the user,\n",
"[5197.12s --> 5199.04s] : the system message in particular?\n",
"[5199.60s --> 5199.92s] : Is,\n",
"[5199.92s --> 5200.24s] : I think,\n",
"[5200.32s --> 5202.16s] : the real path forward And,\n",
"[5202.16s --> 5202.96s] : as you pointed out,\n",
"[5202.96s --> 5214.48s] : these nuanced answers to look at something from several angles It's really really fascinating It's really fascinating Is there something to be said about the employees of a company affecting the bias of the system?\n",
"[5214.64s --> 5216.32s] : One hundred per cent Ah,\n",
"[5216.56s --> 5256.72s] : we try to avoid the Group think bubble It's harder to avoid the AI Group think bubble that follows you everywhere There's all kinds of bubbles who live at hundreds Yeah I'm going on like a around the world user tour soon for a month to just go like talk to our users in different cities And I can like feel how much I'm craving doing that because I haven't done anything like that since in years I used to do that more for Weissie and to go talk to people in super different contexts And it doesn't work over the Internet,\n",
"[5256.72s --> 5259.04s] : like to go show up in person and like sit down and,\n",
"[5259.04s --> 5268.96s] : like go to the bars they go to and kind of like walk through the city like they do you learn so much and get out of the bubble so much Um,\n",
"[5269.52s --> 5277.52s] : I think we are much better than any other company I know of in San Francisco for not falling into the kind of like SF craziness,\n",
"[5277.60s --> 5285.04s] : but I'm sure we're still pretty deeply in it But is it possible to separate the bias of the model versus the bias of the employees?\n",
"[5285.44s --> 5293.36s] : The bias I'm most nervous about is the bias of the human feedback raiders Ah So what's the selection of the human?\n",
"[5293.36s --> 5297.76s] : Is there something you could speak to at a high level about the selection of the human raiders?\n",
"[5297.84s --> 5302.16s] : This is the part that we understand the least well We're great at the pretraining machinery Um,\n",
"[5302.16s --> 5305.04s] : we're now trying to figure out how we're going to select those people,\n",
"[5305.28s --> 5309.92s] : how like how we'll like verify that we get a representative sample,\n",
"[5310.00s --> 5311.76s] : how we'll do different ones for different places,\n",
"[5311.76s --> 5324.72s] : but we don't we don't have that functionality built out yet Such a fascinating science You clearly don't want like all American Elite University students giving you your labels Well,\n",
"[5324.80s --> 5325.04s] : see,\n",
"[5325.04s --> 5326.72s] : it's not about I'm sorry,\n",
"[5326.72s --> 5327.84s] : I just can never resist that,\n",
"[5327.84s --> 5354.16s] : I guess Nice But it's so that that's a good there's a million heuristics you can use That's a to me that's a shallow heuristic because a univer like any one kind of category of human that you would think would have certain beliefs might actually be really open minded and interesting way So you have to optimize for how good you are actually answering at doing these kinds of raiding tasks,\n",
"[5354.40s --> 5361.20s] : how good you are at empathizing with an experience of other humans That's a big one And be able to actually like,\n",
"[5361.28s --> 5366.24s] : what does the world view look like for all kinds of groups of people that would answer this differently?\n",
"[5366.24s --> 5366.48s] : I mean,\n",
"[5366.48s --> 5368.88s] : I have to do that constantly instead of like,\n",
"[5368.88s --> 5369.92s] : You've asked us a few times,\n",
"[5369.92s --> 5371.20s] : but it's something I often do,\n",
"[5371.28s --> 5371.52s] : you know,\n",
"[5371.60s --> 5378.00s] : I ask people in an interview or whatever to steal man ah the beliefs of someone they really disagree with,\n",
"[5378.08s --> 5402.56s] : and the inability of a lot of people to even pretend like they're willing to do that is remarkable Yeah What I find unfortunately ever since Kovid even more so that there's almost an emotional barrier It's not even an intellectual barrier Before they even get to the intellectual there's an emotional barrier that says no Anyone who might possibly believe X They're an idiot,\n",
"[5402.64s --> 5403.76s] : they're evil,\n",
"[5403.84s --> 5406.88s] : they're malevolent Any anything you want to sign,\n",
"[5406.96s --> 5410.24s] : it's like they're not even like loading the data into their heads Look,\n",
"[5410.24s --> 5420.80s] : I think we'll find out that we can make GPT systems way less by us than any human Yeah So hopefully without the Because there won't be that emotional load there Yeah,\n",
"[5420.80s --> 5425.92s] : the emotional load But there might be pressure There might be political pressure Oh,\n",
"[5425.92s --> 5437.44s] : there might be pressure to make a biased system What I meant is that technology I think will be capable of being Much less biased Do you anticipate you worry about pressures from outside sources,\n",
"[5437.52s --> 5438.40s] : from society,\n",
"[5438.48s --> 5439.52s] : from politicians,\n",
"[5439.68s --> 5444.96s] : from money sources I both worry about it and want it Like you know,\n",
"[5445.04s --> 5453.60s] : to the point of worrying this bubble and we shouldn't make all these decisions like We want society to have a huge degree of input here That is pressure in some point in some way Well,\n",
"[5453.60s --> 5454.32s] : there's you know,\n",
"[5454.40s --> 5477.12s] : that's what like to some degree Twitter files have revealed that there is pressure from different organizations You can see in the pandemic where the CDC or some other government organization might put pressure on you know what we're not really sure what's true but it's very unsafe to have these kinds of nuanced conversations now,\n",
"[5477.20s --> 5481.44s] : so let's censor all topics and you get a lot of those emails like,\n",
"[5481.44s --> 5481.76s] : you know,\n",
"[5482.72s --> 5487.68s] : emails all different kinds of people reaching out to different places to put subtle,\n",
"[5487.68s --> 5488.80s] : indirect pressure,\n",
"[5488.88s --> 5489.12s] : ah,\n",
"[5489.12s --> 5490.00s] : direct pressure,\n",
"[5490.08s --> 5490.56s] : financial,\n",
"[5490.56s --> 5491.44s] : political pressure,\n",
"[5491.44s --> 5492.24s] : all that kind of stuff Like,\n",
"[5492.24s --> 5493.68s] : how do you survive that?\n",
"[5494.48s --> 5497.52s] : How do you how much do you worry about that?\n",
"[5498.08s --> 5513.28s] : If GPT continues to get more and more intelligent and a source of information and knowledge for human civilization I think there's like a lot of like quirks about me that make me not a great CEO for open eye,\n",
"[5513.28s --> 5530.32s] : but a thing in the positive column is I think I am relatively good at not being affected by pressure for the sake of pressure By the way,\n",
"[5530.32s --> 5532.00s] : beautiful statement of humility,\n",
"[5532.00s --> 5542.08s] : but I have to ask what's in the negative column I mean too long a list non I'm chat what's a good one I mean,\n",
"[5542.08s --> 5544.64s] : I think I'm not a great like spokesperson for the AI movement,\n",
"[5544.64s --> 5549.92s] : I'll say that I think there could be like a more like There could be someone who enjoyed it more,\n",
"[5549.92s --> 5551.84s] : there could be someone who is like much more charismatic,\n",
"[5551.84s --> 5553.92s] : there could be someone who liked it connects better,\n",
"[5553.92s --> 5554.24s] : I think,\n",
"[5554.24s --> 5560.88s] : with people than I do On with Chompska and this I think charisma is a dangerous thing I think er,\n",
"[5561.28s --> 5565.44s] : flaws in flaws in communication style I think is a feature,\n",
"[5565.44s --> 5613.04s] : not a bug in general At least for humans as at least for humans in power I think I've like more serious problems than that one Um I think I'm like pretty disconnected from like the reality of life for most people and trying to really not just like empathize with but internalize what the impact on people that AGI is going to have I probably like feel that less than other people would That's really well put and you said like you're going to travel across the world to I'm excited to empathize with different users To empathize just to like I want to just like buy our users,\n",
"[5613.12s --> 5613.92s] : our developers,\n",
"[5613.92s --> 5614.40s] : our users,\n",
"[5614.48s --> 5615.76s] : a drink and say like,\n",
"[5615.92s --> 5634.00s] : tell us what you'd like to change And I think one of the things we are not good as good at a company as I would like is to be a really user centric company And I feel like by the time it gets filtered to me it's like totally meaningless So I really just want to go talk to a lot of our users in very different contexts Like you said,\n",
"[5634.08s --> 5638.88s] : a drink in person because I haven't actually found the right words for it,\n",
"[5638.88s --> 5651.60s] : but I I was a little afraid with the programming Emotionally I don't think it makes any sense There is a real lympic response there GPT makes me nervous about the future,\n",
"[5651.68s --> 5653.28s] : not in an AI safety way,\n",
"[5653.36s --> 5660.08s] : but like change And like there's a nervousness about change and more nervous than excited?\n",
"[5660.64s --> 5666.32s] : If I take away the fact that I'm an AI person and just a programmer More excited,\n",
"[5666.32s --> 5667.20s] : but still nervous,\n",
"[5667.28s --> 5668.24s] : like yeah,\n",
"[5668.40s --> 5670.16s] : nervous in brief moments,\n",
"[5670.24s --> 5674.80s] : especially when sleep deprived But there's a nervousness there People who say they're not nervous,\n",
"[5674.80s --> 5678.80s] : I I that's hard for me to believe But you're right,\n",
"[5678.80s --> 5684.64s] : it's excited nervous for change Nervous whenever there is significant exciting kind of change,\n",
"[5684.64s --> 5685.92s] : um You know,\n",
"[5685.92s --> 5687.36s] : I've recently started using,\n",
"[5687.36s --> 5693.60s] : um I've been an emax person for a very long time and I switched to VS code As a copilot?\n",
"[5693.76s --> 5694.24s] : Ah,\n",
"[5694.64s --> 5698.64s] : that was one of the big co reasons 'Cause,\n",
"[5698.64s --> 5698.88s] : like,\n",
"[5698.88s --> 5704.48s] : this is where a lot of active development Of course you can probably do a copilot inside,\n",
"[5704.56s --> 5704.80s] : um,\n",
"[5704.96s --> 5705.84s] : Emacs I mean,\n",
"[5705.84s --> 5708.32s] : I'm sure I'm sure ESCO's all pretty good Yeah,\n",
"[5708.32s --> 5723.20s] : there's a lot of like little little things and and big things that are just really good about VSCOs And I've been I can happily report in all the people who just go nuts But I I'm very happy it's a very happy decision But there is a lot of uncertainty,\n",
"[5723.44s --> 5725.44s] : there's a lot of nervousness about it,\n",
"[5725.44s --> 5727.92s] : there's fear and so on Um,\n",
"[5728.64s --> 5738.72s] : about taking that leap and that's obviously a tiny leap But even just a leap to actively using copil of using a generation of code makes you nervous,\n",
"[5738.72s --> 5743.36s] : but ultimately my life is much better as a programmer Purely as a programmer,\n",
"[5743.36s --> 5763.20s] : a programmer of little things and big things much better There's a nervousness and I think a lot of people will experience that experience that and you will experience that by talking to them And I don't know what we do with that how we comfort people in the face of this uncertainty And you're getting more nervous the more you use it,\n",
"[5763.28s --> 5766.88s] : not less Yes I would have to say yes,\n",
"[5766.96s --> 5779.20s] : because I get better at using it Say the learning curve is quite steep Yeah And then there's moments when you're like ho it generates the function beautifully And you sit back,\n",
"[5779.36s --> 5780.08s] : both proud,\n",
"[5780.16s --> 5781.12s] : like a parent,\n",
"[5781.36s --> 5787.12s] : but almost like proud like and scared that this thing will be much smarter than me?\n",
"[5787.92s --> 5792.80s] : Both pride and sadness almost like a melancholy feeling But ultimately joy,\n",
"[5792.80s --> 5793.12s] : I think,\n",
"[5793.12s --> 5799.28s] : yeah What kind of jobs do you think GPT language models would be better than humans at?\n",
"[5799.36s --> 5802.08s] : Like full like does the whole thing end to end better,\n",
"[5802.16s --> 5802.64s] : not,\n",
"[5802.72s --> 5803.76s] : not like what it's doing with you,\n",
"[5803.76s --> 5813.68s] : where it's helping you be maybe ten times more productive Those are both good questions I don't I would say they're equivalent to me because if I'm ten times more productive,\n",
"[5813.68s --> 5818.40s] : wouldn't that mean that there'll be a need for much fewer programmers in the world?\n",
"[5818.56s --> 5826.64s] : I think the world is going to find out that if you can have ten times as much code at the same price you can just use even more She'll write even more code Just what understands my more code?\n",
"[5826.88s --> 5829.52s] : It is true that a lot more could be digitized,\n",
"[5829.84s --> 5839.12s] : though there could be a lot more code and a lot more stuff I think there is like a supply issue Yeah So in terms of really replaced jobs,\n",
"[5839.28s --> 5840.64s] : is that a worry for you?\n",
"[5841.60s --> 5842.72s] : It is Ah,\n",
"[5842.72s --> 5861.28s] : I'm trying to think of like a big category that I believe can be massively impacted I guess I would say customer service is a category that I could see There are just way fewer jobs relatively soon I'm not even certain about that But I could believe it So like,\n",
"[5861.36s --> 5865.84s] : uh basic questions about when do I take this pill,\n",
"[5866.00s --> 5867.52s] : if it's a drug company,\n",
"[5867.60s --> 5868.48s] : or when,\n",
"[5868.48s --> 5870.56s] : uh I don't know why I went to that,\n",
"[5870.56s --> 5872.32s] : but like how do I use this product,\n",
"[5872.40s --> 5873.12s] : like questions Yeah,\n",
"[5873.12s --> 5873.36s] : like,\n",
"[5873.44s --> 5876.40s] : how do I use whatever whatever call center employees are doing now?\n",
"[5876.48s --> 5878.00s] : Yeah This is not work yeah,\n",
"[5878.08s --> 5880.80s] : okay I want to be clear,\n",
"[5880.88s --> 5886.16s] : I think like these systems will make a lot of jobs just go away,\n",
"[5886.16s --> 5891.68s] : every technological revolution does They will enhance many jobs and make them much better,\n",
"[5891.76s --> 5892.48s] : much more fun,\n",
"[5892.56s --> 5906.88s] : much higher paid And and they'll create new jobs that are difficult for us to imagine even if we're starting to see the first glimpses of them But um I heard some one last week talking about GPT four saying that,\n",
"[5906.88s --> 5907.36s] : you know,\n",
"[5907.44s --> 5908.88s] : man ah,\n",
"[5909.28s --> 5925.44s] : the dignity of work is just such a huge deal We've really got to worry like even people who think they don't like their jobs they really need them it's really important to them and to society And also can you believe how awful it is that France is trying to raise the retirement age?\n",
"[5926.48s --> 5950.56s] : And I think we as a society are confused about whether we want to work more or work less and certainly about whether most people like their jobs and get value out of their jobs or not Some people do I love my job I suspect you do too That's a real privilege not everybody gets to say that If we can move more of the world to better jobs and work to something that can be a broader concept,\n",
"[5950.80s --> 5952.96s] : not something you have to do to be able to eat,\n",
"[5953.04s --> 5967.36s] : but something you do as a creative expression and a way to find fulfilment and happiness whatever else Even if those jobs look extremely different from the jobs of to day I think that's great I'm not I'm not nervous about it at all You have been a proponent of UBI,\n",
"[5967.36s --> 5975.36s] : universal basic income In the context of AI can you describe your philosophy there of our human future with UBI?\n",
"[5975.60s --> 5976.88s] : Why why you like it,\n",
"[5976.96s --> 5978.40s] : what are some limitations?\n",
"[5978.64s --> 5986.56s] : I think it is a component Of something we should pursue It is not a full solution I think people work for lots of reasons,\n",
"[5986.64s --> 5996.96s] : besides money And I think we are going to find incredible new jobs and society as a whole,\n",
"[5997.12s --> 5999.12s] : and people's individuals are going to get much,\n",
"[5999.12s --> 6007.04s] : much richer but as a cushion through a dramatic transition and as just like You know,\n",
"[6007.04s --> 6043.76s] : I think the world should eliminate poverty if able to do so I think it's a great thing to do as a small part of the bucket of solutions I helped start a project called Worldcoin um which is a technological solution to this We also have funded a a like a large I think maybe the largest and most comprehensive universal basic income study as part of sponsored by opening eye and I think it's like an area we should just be looking into What are some insights from that study that you gained?\n",
"[6043.84s --> 6045.92s] : We're going to finish up at the end of this year,\n",
"[6045.92s --> 6050.32s] : and we'll be able to talk about it hopefully very early next If we can linger on it,\n",
"[6050.40s --> 6057.28s] : how do you think the economic and political systems will change as AI becomes a prevalent part of society?\n",
"[6057.36s --> 6062.64s] : It's such an interesting sort of philosophical question looking ten,\n",
"[6062.64s --> 6063.04s] : twenty,\n",
"[6063.04s --> 6067.20s] : fifty years from now what does the economy look like?\n",
"[6067.68s --> 6069.60s] : What does politics look like?\n",
"[6069.76s --> 6074.96s] : Do you see significant transformations in terms of the way democracy functions even?\n",
"[6075.12s --> 6076.48s] : I love that you ask them together,\n",
"[6076.56s --> 6082.48s] : because I think they're super related I think the economic transformation will drive much of the political transformation here,\n",
"[6082.64s --> 6109.52s] : not the other way around Um my working model for the last five years has been that the two dominant changes will be that the cost of intelligence and the cost of energy are going over the next couple of decades to dramatically dramatically fall from where they are to day And the impact of that you've already seen it with the way you now have like p you know,\n",
"[6109.84s --> 6116.48s] : programming ability beyond what you had as an individual before is society gets much,\n",
"[6116.48s --> 6117.28s] : much richer,\n",
"[6117.44s --> 6118.24s] : much wealthier,\n",
"[6118.32s --> 6130.64s] : in ways that are probably hard to imagine I think every time that's happened before it has been that economic impact has had positive political impact as well And I think it does go the other way,\n",
"[6130.72s --> 6130.96s] : too,\n",
"[6131.04s --> 6152.00s] : like the socio political values of the Enlightenment enabled the long running technological revolution and scientific discovery process we've had for the past centuries But I think we're just going to see more I'm sure the shape will change,\n",
"[6152.00s --> 6163.28s] : but I think it's this long and beautiful exponential curve Do you think there will be more er I don't know what the the term is,\n",
"[6163.28s --> 6170.64s] : but systems that resemble something like democratic socialism I've talked to a few folks in this pod guess about these kinds of topics Instinct,\n",
"[6170.64s --> 6171.04s] : yes,\n",
"[6171.04s --> 6191.44s] : I hope so So that it reallocates some resources in a way that supports kind of lifts the the people who are struggling I am a big believer and lifts up the floor and don't worry about the ceiling If I can er test your historical knowledge it's probably not going to be good,\n",
"[6191.44s --> 6195.44s] : but let's try it Why do you think I come from the Soviet Union?\n",
"[6195.44s --> 6197.92s] : Why do you think communism in the Soviet Union failed?\n",
"[6198.24s --> 6218.40s] : I recoil at the idea of living in a Communist system And I don't know how much of that is just the biases of the world I grow up in and what I have been taught and probably more than I realize but I think like more individualism,\n",
"[6218.64s --> 6220.48s] : more human will,\n",
"[6220.64s --> 6243.68s] : more ability to self determine is important and also I think the ability to try new things and not need permission and not need some sort of central planning betting on human ingenuity and this sort of like distributed process,\n",
"[6244.24s --> 6260.80s] : I believe is always going to beat centralized planning and I think that like for all the deep flaws of America I think it is the greatest place in the world because it's the best at this So it's really interesting,\n",
"[6261.28s --> 6268.08s] : that centralized planning failed so in such big ways But what if,\n",
"[6268.08s --> 6268.96s] : hypothetically,\n",
"[6269.04s --> 6272.40s] : the centralized planning is the perfect superintelligent AGI,\n",
"[6272.40s --> 6276.32s] : superintelligent AGI Again,\n",
"[6276.40s --> 6280.00s] : it might go wrong in the same kind of ways,\n",
"[6280.00s --> 6297.92s] : but it might not We don't really know We don't really know It might be better I expect it would be better But would it be better than a hundred superintelligent or a thousand superintelligent AGI's sort of in a liberal democratic system Argue?\n",
"[6298.08s --> 6303.52s] : Yes Now also how much of that can happen internally in one superintelligent,\n",
"[6303.52s --> 6304.32s] : AGI?\n",
"[6304.72s --> 6311.60s] : Not so obvious There is something about right but there is something about like tension,\n",
"[6311.68s --> 6316.40s] : the competition But you don't know that's not happening inside one model Yeah,\n",
"[6316.64s --> 6333.04s] : that's true It'd be nice it'd be nice if whether it's engineered in or revealed to be happening it'd be nice for it to be happening that then of course it can happen with multiple AGI's talking to each other or whatever There's something also about,\n",
"[6333.04s --> 6333.20s] : uh,\n",
"[6333.20s --> 6336.08s] : Mr Russell has talked about the control problem of,\n",
"[6336.16s --> 6350.88s] : um always having AGI to be have some degree of uncertainty not having a dogmatic certainty to it That feels important Some of that is already handled with human alignment a a human feedback,\n",
"[6350.96s --> 6363.04s] : reinforcement learning with human feedback But it feels like there has to be engineered in like a hard uncertainty Humility you can put a romantic word to it Yep Do you think that's possible to do?\n",
"[6363.76s --> 6365.44s] : The definition of those words,\n",
"[6365.44s --> 6366.96s] : I think The details really matter,\n",
"[6366.96s --> 6368.08s] : but as I understand them,\n",
"[6368.08s --> 6368.48s] : yes,\n",
"[6368.48s --> 6370.56s] : I do What about the off switch?\n",
"[6370.80s --> 6375.76s] : That like big red button in the Data Center we don't tell anybody about Us that way?\n",
"[6376.00s --> 6376.80s] : My backpack?\n",
"[6376.88s --> 6378.08s] : Having your backpack?\n",
"[6378.32s --> 6378.64s] : Uh,\n",
"[6378.64s --> 6380.48s] : you think it's possible to have a switch?\n",
"[6380.72s --> 6381.52s] : You think I mean,\n",
"[6381.52s --> 6383.36s] : I think more seriously,\n",
"[6383.44s --> 6391.28s] : more specifically about sort of rolling out of different systems Do you think it's possible to roll them unroll them,\n",
"[6391.76s --> 6393.76s] : pull them back in Yeah I mean,\n",
"[6393.76s --> 6415.44s] : we can absolutely take a model back off the Internet We can like take we can turn an API off Isn't that something you worry about like when you release it and millions of people are using it and like you realize holy crap they're using it for I don't know worrying about the like all kinds of terrible use cases We do worry about that a lot I mean,\n",
"[6415.44s --> 6423.76s] : we try to figure out with as much red teaming and testing ahead of time as we do how to avoid a lot of those,\n",
"[6423.76s --> 6435.36s] : but I can't emphasize enough how much the collective intelligence and creativity of the world will beat open eye on all the red tingers we can hire So we put it out,\n",
"[6435.44s --> 6441.44s] : but we put it out in a way we can make changes In the millions of people that have used the ChadGi and GPT,\n",
"[6441.52s --> 6444.00s] : what have you learned about human civilization in general?\n",
"[6444.48s --> 6444.80s] : I mean,\n",
"[6445.04s --> 6447.60s] : the question I ask is Are we mostly good?\n",
"[6448.40s --> 6452.40s] : Or is there a lot of malevolence in the human spirit?\n",
"[6452.64s --> 6452.88s] : Well,\n",
"[6453.04s --> 6453.68s] : to be clear,\n",
"[6453.76s --> 6464.80s] : I don't nor does anyone else open the eyes that they're like reading all the chat GVT messages Yeah But from what I hear people using it for,\n",
"[6464.88s --> 6466.32s] : at least the people I talk to,\n",
"[6466.40s --> 6468.72s] : and from what I see on Twitter,\n",
"[6469.12s --> 6476.64s] : we are definitely mostly good but A not all of us are,\n",
"[6477.52s --> 6478.24s] : all the time,\n",
"[6478.32s --> 6479.12s] : and b,\n",
"[6479.52s --> 6484.80s] : we really want to push on the edges of these systems and you know,\n",
"[6484.96s --> 6488.48s] : we really want to test out some darker theories The other world?\n",
"[6488.72s --> 6489.12s] : Yes,\n",
"[6489.44s --> 6491.68s] : very interesting Very interesting,\n",
"[6491.68s --> 6498.24s] : and I think that's not that that actually doesn't communicate the fact that we're like fundamentally dark inside,\n",
"[6498.24s --> 6501.52s] : but we like to go to the dark places in order to,\n",
"[6501.60s --> 6503.84s] : um,\n",
"[6503.92s --> 6505.84s] : maybe rediscover the light?\n",
"[6506.72s --> 6513.44s] : It feels like dark humor is a part of that Some of the darkest some of the toughest things you go through if you suffer in life in a war zone Um,\n",
"[6513.52s --> 6531.44s] : the people I've interacted with they're in the midst of a war they're joking jokes around they're joking around and they're dark jokes So that there's something there I totally agree about that tension So just to the model how do you decide what isn't isn't misinformation?\n",
"[6531.68s --> 6533.04s] : How do you decide what is true?\n",
"[6533.04s --> 6541.36s] : You actually have Open A as internal factual performance benchmark There's a lot of cool benchmarks here How do you build a benchmark for what is true?\n",
"[6542.24s --> 6543.12s] : What is truth?\n",
"[6543.76s --> 6545.92s] : Sam Alban Like math is true,\n",
"[6546.08s --> 6566.96s] : and the origin of Covid is not agreed upon as ground truth those are the two things And then there's stuff that's like certainly not true But between that first and second milestone there's a lot of disagreement Where do you look for?\n",
"[6567.04s --> 6569.76s] : Where can a not even just now,\n",
"[6569.76s --> 6576.96s] : but in the future where can we as a human civilization look for look to for truth?\n",
"[6577.68s --> 6578.24s] : What do you know,\n",
"[6578.32s --> 6578.88s] : is true?\n",
"[6579.52s --> 6580.88s] : What are you absolutely certain,\n",
"[6580.88s --> 6581.60s] : is true?\n",
"[6586.40s --> 6593.52s] : I have a generally epistemic humiliated about everything and am freaked out by how little I know and understand about the world,\n",
"[6593.60s --> 6602.56s] : so that even that question is terrifying to me There's a bucket of things that have a high degree of truth in this,\n",
"[6602.64s --> 6605.44s] : which is where you put math a lot of math,\n",
"[6605.44s --> 6607.12s] : yeah Can't be certain,\n",
"[6607.12s --> 6610.56s] : but it's good enough for like this conversation we can say math is true Yeah,\n",
"[6610.96s --> 6612.24s] : I mean some er,\n",
"[6612.48s --> 6625.84s] : quite a bit of physics historical facts maybe dates of when a war started There's a lot of details about a military conflict inside history Of course,\n",
"[6626.16s --> 6626.96s] : you start to get,\n",
"[6626.96s --> 6627.20s] : you know,\n",
"[6628.00s --> 6629.44s] : just read Blitz,\n",
"[6629.52s --> 6630.24s] : which is the Oh,\n",
"[6630.24s --> 6632.64s] : I want to read that So how was it?\n",
"[6633.20s --> 6648.48s] : It was really good It's it gives a theory of Nazi Germany and Hitler that so much can be described about Hitler and a lot of the upper echelon of Nazi Germany through the excessive use of drugs And phatamins,\n",
"[6648.48s --> 6648.80s] : right?\n",
"[6648.80s --> 6649.92s] : And phatamines,\n",
"[6649.92s --> 6650.64s] : but also other stuff,\n",
"[6650.64s --> 6654.00s] : but it's just a lot And,\n",
"[6654.08s --> 6654.40s] : you know,\n",
"[6654.40s --> 6655.28s] : that's really interesting,\n",
"[6655.28s --> 6658.24s] : it's really compelling and for some reason like Whoa,\n",
"[6658.32s --> 6670.64s] : that's really that would explain a lot that's somehow really sticky It's an idea that's sticky and then you read a lot of criticism of that book later by historians that's a lot of cherry picking going on,\n",
"[6670.64s --> 6689.92s] : and it's actually is using the fact that that's a very sticky explanation There's something about humans that likes a very simple narrow describe every and then yet too much amphetamines cause the war is like a great even if not true simple explanation that feels satisfying and excuses a lot of other,\n",
"[6689.92s --> 6692.80s] : probably much darker human truths Yes,\n",
"[6692.96s --> 6698.24s] : the military strategy employed the atrocities,\n",
"[6698.56s --> 6700.88s] : the speeches er,\n",
"[6701.04s --> 6704.08s] : the just the way Hitler was as a human being,\n",
"[6704.08s --> 6705.52s] : the way Hitler was as a leader,\n",
"[6705.68s --> 6708.32s] : all that could be explained to this one little lens,\n",
"[6708.32s --> 6708.88s] : and it's like,\n",
"[6708.88s --> 6709.12s] : well,\n",
"[6709.60s --> 6710.88s] : if you say that's true,\n",
"[6711.04s --> 6721.52s] : that's a really compelling truth So maybe truth is in one sense as defined as a thing that is a collective intelligence we kind of all our brains are sticking to And we're like,\n",
"[6721.52s --> 6721.84s] : Ya,\n",
"[6721.84s --> 6725.52s] : ya A bunch of a bunch of ants get together and like,\n",
"[6725.52s --> 6725.84s] : Yeah,\n",
"[6725.84s --> 6727.44s] : this is it I was going to say cheap,\n",
"[6727.44s --> 6730.16s] : but there's a connotation to that But,\n",
"[6730.24s --> 6730.48s] : yeah,\n",
"[6730.64s --> 6732.32s] : it's it's hard to know what is true,\n",
"[6732.32s --> 6739.36s] : and I think when constructing a GPT like model you have to contend with that I think a lot of the answers,\n",
"[6739.36s --> 6739.60s] : you know,\n",
"[6739.68s --> 6743.20s] : like if you ask GPT four I don't know,\n",
"[6743.20s --> 6777.36s] : just to stick on the same topic did covidly from a lab I expect you would get a reasonable answer is a really good answer It laid out the hypotheses The interesting thing it said which is refreshing to hear is there's something like there's very little evidence for either hypothesis direct evidence which is important to state A lot of people kinda the reason why there's a lot of uncertainty and a lot of debate is because there's not strong physical evidence of either heavy circumstantial evidence on either side,\n",
"[6777.44s --> 6781.20s] : and then the other is more like biological theoretical kind of,\n",
"[6781.28s --> 6781.60s] : um,\n",
"[6781.68s --> 6784.08s] : discussion And I think the answer,\n",
"[6784.08s --> 6788.08s] : the nuanced answer to the GPT provider was actually pretty damn good,\n",
"[6788.24s --> 6795.44s] : and also importantly saying that there is uncertainty just the fact that there is uncertainty as the statement was really powerful Man,\n",
"[6795.44s --> 6796.00s] : remember when,\n",
"[6796.00s --> 6796.16s] : like,\n",
"[6796.16s --> 6801.36s] : the social media platforms were banning people for Saying it was a oblique?\n",
"[6801.68s --> 6804.80s] : Yeah That's really humbling the humbling,\n",
"[6804.96s --> 6810.88s] : the overreach of power in censorship But that the more powerful GP becomes,\n",
"[6810.88s --> 6820.08s] : the more pressure there'll be to censor We have a different set of challenges faced by the previous generation of companies,\n",
"[6820.32s --> 6826.40s] : which is people talk about free speech issues with GPT,\n",
"[6826.40s --> 6830.88s] : but it's not quite the same thing It's not like this is a computer program what it's allowed to say,\n",
"[6830.88s --> 6842.16s] : and it's also not about the mass spread and the challenges that I think may have made that Twitter and Facebook and others have struggled with so much So we will have very significant challenges,\n",
"[6842.16s --> 6847.52s] : but they'll be very new and very different And maybe yeah,\n",
"[6847.60s --> 6848.08s] : very new,\n",
"[6848.08s --> 6864.56s] : very different is the way to put it There could be truths that are harmful in their truth Um I don't know Group difference isn't like you There you go Scientific work that when spoken might do more harm And you asked GPT that,\n",
"[6864.56s --> 6865.92s] : should GPT tell you?\n",
"[6866.08s --> 6869.92s] : There's books written on this that are rigorous scientifically,\n",
"[6870.00s --> 6882.88s] : but are very uncomfortable and probably not productive in any sense But maybe are There's people are arguing all kinds of sides of this and a lot of them have hate in their heart And so what do you do with that?\n",
"[6882.88s --> 6886.80s] : If there's a large number of people who hate others but are actually,\n",
"[6886.88s --> 6887.28s] : um,\n",
"[6887.68s --> 6889.12s] : citing scientific studies,\n",
"[6889.20s --> 6889.92s] : what do you do with that?\n",
"[6889.92s --> 6891.20s] : What does Gipy do with that?\n",
"[6891.20s --> 6894.56s] : What is the priority of GPT to decrease them out of hate in the world?\n",
"[6894.80s --> 6897.52s] : Is it up to GPT or is it up to us humans?\n",
"[6897.76s --> 6907.12s] : I think we as open AI have responsibility for the tools we put out into the world I think the tools themselves can't have responsibility,\n",
"[6907.20s --> 6912.88s] : in the way I understand it Wow So you carry some of that burden for responsibility,\n",
"[6912.88s --> 6920.64s] : all of us all of us at the Company So there could be harm caused by this tool,\n",
"[6920.72s --> 6927.36s] : and there will be harm caused by this tool There will be harm there will be tremendous benefits But,\n",
"[6927.52s --> 6927.84s] : you know,\n",
"[6928.08s --> 6945.20s] : tools do wonderful good and real bad We will minimize the bad and maximize the good We have to carry the weight to that How do you avoid GPT for from being hacked or jail broken?\n",
"[6945.60s --> 6948.80s] : There's a lot of interesting ways that people have done that like,\n",
"[6948.80s --> 6949.04s] : uh,\n",
"[6949.12s --> 6954.48s] : with token smuggling or other methods like Dan You know,\n",
"[6954.56s --> 6956.80s] : when I was like a a kid,\n",
"[6956.80s --> 6957.36s] : basically,\n",
"[6957.36s --> 6961.12s] : I I got worked once on jail working an iPhone the first iPhone,\n",
"[6961.12s --> 6981.60s] : I think and I thought it was so cool I will say it's very strange to be on the other side of that You're now the man Kinda sucks Um Is that is some of it fun?\n",
"[6981.76s --> 6983.52s] : How much of it is a security threat?\n",
"[6983.52s --> 6983.92s] : I mean,\n",
"[6983.92s --> 6986.48s] : what how much do J have to take it seriously?\n",
"[6986.56s --> 6988.72s] : How is it even possible to solve this problem?\n",
"[6988.80s --> 6990.32s] : Where does it rank on the set of problem?\n",
"[6990.32s --> 7012.80s] : I just keep asking questions prompting We want users to have a lot of control and get the models to behave in the way they want within some very broad bounds And I think the whole reason for jail breaking is right now we haven't yet figured out how to give that to people,\n",
"[7012.96s --> 7015.52s] : and the more we solve that problem,\n",
"[7015.68s --> 7018.56s] : I think the less need there'll be for jail breaking Yes,\n",
"[7018.72s --> 7020.08s] : kind of like piracy,\n",
"[7020.32s --> 7025.04s] : gave Bertha Spotify People don't really jail break eyphones that much any more,\n",
"[7025.04s --> 7026.08s] : and it's gotten harder,\n",
"[7026.16s --> 7026.56s] : for sure,\n",
"[7026.56s --> 7031.44s] : but also like you can just do a lot of stuff now Just like with jail breaking,\n",
"[7031.44s --> 7038.56s] : that means there's a lot of hilarity that is so Evan Marakawa,\n",
"[7038.64s --> 7039.36s] : Ku guy,\n",
"[7039.36s --> 7048.72s] : he's at open AI He tweeted something that he also was really kind to send me to communicate with me sent me long email describing the history of open AI,\n",
"[7048.72s --> 7051.04s] : all the different developments Um,\n",
"[7051.44s --> 7053.12s] : he really lays it out I mean,\n",
"[7053.12s --> 7055.92s] : that's a much longer conversation of all the awesome stuff that happened,\n",
"[7056.00s --> 7059.60s] : it's just amazing But his tweet was Dolly,\n",
"[7059.68s --> 7060.56s] : july twenty two,\n",
"[7060.56s --> 7061.52s] : Chagipine,\n",
"[7061.52s --> 7062.40s] : november twenty two,\n",
"[7062.40s --> 7064.16s] : API sixty six percent cheaper,\n",
"[7064.16s --> 7064.96s] : august twenty two,\n",
"[7064.96s --> 7066.32s] : embedding's five hundred,\n",
"[7066.48s --> 7067.36s] : times cheaper,\n",
"[7067.44s --> 7068.48s] : while state of the art,\n",
"[7068.48s --> 7069.44s] : december twenty two,\n",
"[7069.44s --> 7075.12s] : Chagpity API also ten times cheaper while state of the art march twenty three Whisper API,\n",
"[7075.20s --> 7076.00s] : march twenty three,\n",
"[7076.00s --> 7077.36s] : GP four today,\n",
"[7077.44s --> 7085.92s] : whenever that was last week And the conclusion is this team ships we do Ah,\n",
"[7085.92s --> 7090.00s] : what's the process of going and then we can extend that back I mean,\n",
"[7090.16s --> 7093.60s] : listen from the twenty fifteen Open Eye Launch,\n",
"[7093.68s --> 7094.40s] : GPT,\n",
"[7094.40s --> 7095.36s] : GPT two,\n",
"[7095.36s --> 7096.56s] : GPT three,\n",
"[7096.72s --> 7099.20s] : Open Eye Five finals with the gaming stuff,\n",
"[7099.20s --> 7102.40s] : which is an incredible GPT three API released ah,\n",
"[7102.56s --> 7103.12s] : Dolly,\n",
"[7103.12s --> 7106.96s] : instruct GPT I can find tuning Ah,\n",
"[7107.12s --> 7108.56s] : there's just a million things,\n",
"[7108.64s --> 7108.88s] : ah,\n",
"[7108.88s --> 7110.32s] : available Doli,\n",
"[7110.40s --> 7111.52s] : Doli two,\n",
"[7111.68s --> 7114.96s] : a preview and then Doli's available to one million people,\n",
"[7115.20s --> 7116.00s] : Whisper,\n",
"[7116.16s --> 7117.44s] : a second model release,\n",
"[7117.44s --> 7119.52s] : just across all of the stuff,\n",
"[7119.60s --> 7121.52s] : both research and,\n",
"[7121.60s --> 7122.08s] : um,\n",
"[7122.64s --> 7125.36s] : deployment of actual products that could be in the hands of people,\n",
"[7125.44s --> 7125.76s] : ah,\n",
"[7125.84s --> 7134.24s] : what is the process of going from idea to deployment that allows you to be so successful at shipping AI based products?\n",
"[7134.64s --> 7134.96s] : I mean,\n",
"[7134.96s --> 7136.96s] : there's a question of should we be really proud of that,\n",
"[7137.04s --> 7139.12s] : or should other companies be really embarrassed?\n",
"[7139.44s --> 7144.96s] : Yes And we believe in a very high bar for the people on the team?\n",
"[7145.44s --> 7161.44s] : We work hard which you know you're not even like to say any more or something We give a huge amount of trust and autonomy and authority to individual people,\n",
"[7161.76s --> 7168.24s] : and we try to hold each other to very high standards and You know,\n",
"[7168.32s --> 7170.16s] : there's a process which we can talk about,\n",
"[7170.16s --> 7180.40s] : but it won't be that illuminating I think it's those other things that make us able to ship at a high velocity So GPT four is a pretty complex system Like you said,\n",
"[7180.40s --> 7186.64s] : there's like a million little hacks you can do to keep improving it There's the cleaning up the data set,\n",
"[7186.72s --> 7188.72s] : all that all those are like separate teams,\n",
"[7188.80s --> 7194.88s] : so do you give a tan me is there just a tanomy to these fascinating different problems?\n",
"[7195.04s --> 7195.28s] : If,\n",
"[7195.36s --> 7195.52s] : like,\n",
"[7195.60s --> 7201.44s] : most people in the company weren't really excited to work super hard and collaborate well on GPD four and thought other stuff was more,\n",
"[7201.68s --> 7210.32s] : important there'd be very little I or anybody else could do to make it happen But we spend a lot of time figuring out what to do,\n",
"[7210.48s --> 7213.28s] : getting on the same page about why we're doing something,\n",
"[7213.44s --> 7225.04s] : and then how to divide it up and all coordinate together So then then you have like a passion for the for the goal here So everybody is really passionate across the difference Yes,\n",
"[7225.12s --> 7227.04s] : we care How do you hire?\n",
"[7227.28s --> 7228.96s] : How do you hire great teams?\n",
"[7229.52s --> 7238.08s] : The folks I've interacted with open eyes some of the most amazing folks I've ever met It takes a lot of time Like I I spend I mean,\n",
"[7238.16s --> 7249.36s] : I think a lot of people claim to spend a third of their time hiring I for real truly do I still approve every single hired open eye And I think there's you know,\n",
"[7249.36s --> 7267.04s] : we're working on a problem that is like very cool and the great people want to work on We have great people and some people want to be around them but even with that I think there's just no short cut for putting a ton of effort into this So even when you have the good the good people hard work?\n",
"[7267.36s --> 7272.08s] : I think so Microsoft denounced the new multi year,\n",
"[7272.08s --> 7286.80s] : multibillion dollar reported to be ten billion dollars investment into OpenAI Can you describe the thinking that went into this what are the prose what are the cons of working with a company like Microsoft?\n",
"[7287.84s --> 7291.36s] : It's not all perfect or easy,\n",
"[7291.44s --> 7301.68s] : but on the whole they have been an amazing partner to us Satya and Kevin and Mikhail are super aligned with us,\n",
"[7302.08s --> 7303.28s] : super flexible,\n",
"[7303.36s --> 7320.72s] : have gone way above and beyond the call of duty to do things that we have needed to get all this to work This is like a big iron complicated engineering project and they are a big and complex company And I think like many great partnerships or relationships,\n",
"[7320.88s --> 7326.80s] : we've sort of just continued to ramp up our investment in each other And it's been very good?\n",
"[7327.36s --> 7337.28s] : It's a for profit company It's very driven it's very large scale Is there pressure to kind of make a lot of money?\n",
"[7337.52s --> 7356.48s] : I think most other companies wouldn't maybe now they would It wouldn't at the time have understood why we needed all the weird control provisions we have and why we need all the kind of like AGI specialness Um And I know that because I talked to some other companies before we did the first deal with Microsoft Um,\n",
"[7356.56s --> 7376.16s] : and I think they are unique in terms of the companies at that scale that understood why we needed the control provisions we have So those control provisions help you help make sure that the capitalist imperative does not affect the development of AI Well,\n",
"[7376.16s --> 7380.08s] : let me just ask you as an aside about Sachin Adela,\n",
"[7380.16s --> 7388.00s] : the CEO of Microsoft He seems to have successfully transformed Microsoft into this fresh,\n",
"[7388.00s --> 7388.64s] : innovative,\n",
"[7388.72s --> 7392.32s] : developer friendly company I agree What do you I mean,\n",
"[7392.32s --> 7394.72s] : is it really hard to do for a very large company?\n",
"[7396.00s --> 7397.36s] : What have you learned from him?\n",
"[7397.44s --> 7400.24s] : Why do you think he was able to do this kind of thing?\n",
"[7400.40s --> 7401.84s] : Um Yeah,\n",
"[7402.00s --> 7410.96s] : what what insights do you have about why this one human being is able to contribute to the pivot of a large company into something very new?\n",
"[7411.76s --> 7420.00s] : I think most CEOs are either great leaders or great managers and,\n",
"[7420.00s --> 7423.12s] : from what I have observed with Satya,\n",
"[7423.60s --> 7427.28s] : he is both Supervisionary,\n",
"[7427.60s --> 7430.56s] : really like gets people excited,\n",
"[7430.72s --> 7448.40s] : really makes long duration and correct calls and also he is just a super effective hands on executive and I assume manager to And I think that's pretty rare I mean,\n",
"[7448.40s --> 7449.04s] : Microsoft,\n",
"[7449.04s --> 7449.76s] : if I'm guessing,\n",
"[7449.76s --> 7450.00s] : like,\n",
"[7450.00s --> 7454.16s] : IBM like a lot of companies that have been at it for a while probably have,\n",
"[7454.16s --> 7454.40s] : like,\n",
"[7454.48s --> 7461.68s] : old school kind of momentum So you like inject AI into it it's very tough Or anything,\n",
"[7461.76s --> 7464.96s] : even like open sour the the culture of open source?\n",
"[7465.84s --> 7466.40s] : Um,\n",
"[7466.72s --> 7467.12s] : like,\n",
"[7467.28s --> 7472.48s] : how how hard is it to walk into a room and be like the way we've been doing things are totally wrong?\n",
"[7472.56s --> 7478.80s] : Like I'm sure there's a lot of firing involved or a little like twisting of arms or something So do you have to rule by fear,\n",
"[7478.80s --> 7479.20s] : by love?\n",
"[7479.20s --> 7482.24s] : Like what can he say to the leadership aspect of this?\n",
"[7482.64s --> 7483.04s] : I mean,\n",
"[7483.04s --> 7484.64s] : he just liked an unbelievable job,\n",
"[7484.72s --> 7495.36s] : but he is amazing at being like clear and firm and getting people to want to come along,\n",
"[7495.52s --> 7503.68s] : but also like compassionate and patient with his people too I'm getting a lot of love,\n",
"[7503.68s --> 7508.00s] : not fear I'm a big Satya fan So am I,\n",
"[7508.08s --> 7509.68s] : from a distance I mean,\n",
"[7509.68s --> 7513.60s] : you have so much in your life trajectory that I can ask you about,\n",
"[7513.60s --> 7522.08s] : we can probably talk for many more hours But I got to ask you because of why Combinator because of start ups and so on The recent you've tweeted about this,\n",
"[7522.08s --> 7522.32s] : uh,\n",
"[7522.72s --> 7528.48s] : about the Silicon Valley Bank SVB What's your best understanding of what happened?\n",
"[7528.64s --> 7532.88s] : What is interesting what is interesting to understand about what happened at SVB?\n",
"[7533.04s --> 7564.80s] : I think they just like horribly mismanaged buying while chasing returns in a very silly world of zero percent interest rates buying very long dated instruments secured by very short term and variable deposits and this was obviously dumb I think totally the fault of the management team,\n",
"[7564.80s --> 7584.00s] : although I'm not sure what the regulators were thinking either and is an example of where I think you see the dangers of incentive misalignment because as the Fed kept raising,\n",
"[7584.40s --> 7593.04s] : I assume that the incentives on people working at SVB to not sell at a loss their,\n",
"[7593.04s --> 7593.36s] : you know,\n",
"[7593.44s --> 7597.76s] : super safe bonds which were now down twenty percent or whatever or,\n",
"[7597.76s --> 7598.08s] : you know,\n",
"[7598.08s --> 7598.80s] : down less than that,\n",
"[7598.80s --> 7602.56s] : but then kept going down I you know,\n",
"[7602.64s --> 7607.04s] : that's like a classy example of incentive misalignment Now,\n",
"[7607.12s --> 7624.80s] : I suspect they're not the only bank in a bad position here The response of the Federal Government I think took much longer than it should have but by Sunday afternoon I was glad they had done what they have done We'll see what happens next So how do you avoid depositors from doubting their bank?\n",
"[7625.04s --> 7649.44s] : What I think needs would be good to do right now is just a and this requires statutory change but it may be a full guarantee of deposits maybe a much higher than two fifty k But you really don't want depositors having to doubt the security of their deposits and this thing that a lot of people on Twitter were saying is like,\n",
"[7649.44s --> 7649.68s] : Well,\n",
"[7649.68s --> 7651.12s] : it's their fault they should have been like,\n",
"[7651.12s --> 7651.44s] : you know,\n",
"[7651.52s --> 7656.88s] : reading the the balance sheet and the risk audit of the bank like Do we really want people to have to do that?\n",
"[7656.88s --> 7657.60s] : I would argue,\n",
"[7657.68s --> 7662.48s] : No What impact has it had on start ups,\n",
"[7662.48s --> 7663.20s] : that you see?\n",
"[7663.28s --> 7663.60s] : Well,\n",
"[7663.60s --> 7664.96s] : there was a weekend of terror,\n",
"[7664.96s --> 7667.20s] : for sure and now I think,\n",
"[7667.28s --> 7668.72s] : even though it was only ten days ago,\n",
"[7668.80s --> 7669.84s] : it feels like forever,\n",
"[7669.84s --> 7684.16s] : and people have forgotten about it But it kind of reveals the fragility of our kid oaks It may not be done That may have been like the gun showing falling off the night stand in the first scene of the movie or whatever It could be like other banks There sure could we even with FTX I mean,\n",
"[7684.16s --> 7686.72s] : I'm just er well,\n",
"[7686.80s --> 7687.68s] : that's fraud,\n",
"[7687.76s --> 7709.12s] : but there's mismanagement and you wonder how stable our economic system is especially with new entrants with AGI I think one of the many lessons to take away from this SBB thing is how fast and how much the world changes,\n",
"[7709.12s --> 7714.88s] : and how little I think our experts leaders business leaders,\n",
"[7714.88s --> 7715.52s] : regulators,\n",
"[7715.52s --> 7715.92s] : whatever,\n",
"[7716.00s --> 7722.48s] : understand it So the the speed with which the SBB bankrupt happened,\n",
"[7722.72s --> 7723.76s] : because of Twitter,\n",
"[7723.76s --> 7725.36s] : because of mobile banking lapse,\n",
"[7725.44s --> 7725.92s] : whatever,\n",
"[7726.16s --> 7728.56s] : was so different than the two thousand eight collapse,\n",
"[7728.64s --> 7740.24s] : where we didn't have those things really and I don't think the kind of people in power realise how much the field had shifted,\n",
"[7740.32s --> 7751.12s] : and I think that is a very tiny preview of the shifts that AGI will bring What gives you hope in that shift from an economic perspective?\n",
"[7751.36s --> 7755.20s] : Ah It sounds scary the instability No,\n",
"[7755.52s --> 7763.76s] : I am nervous about the speed with this changes and the speed with which our institutions can adapt,\n",
"[7763.92s --> 7775.20s] : um which is part of why we want to start deploying these systems really early while they're really weak so that people have as much time as possible to do this I think it's really scary to like have nothing,\n",
"[7775.20s --> 7775.52s] : nothing,\n",
"[7775.60s --> 7794.00s] : nothing and then drop a super powerful AGI all at once on the world I don't think people should want that to happen But what gives me hope is like I think the less zero the more positive sum the world gets the better and the the upside of the vision here just how much better life can be I think that's going to,\n",
"[7794.00s --> 7794.24s] : like,\n",
"[7794.56s --> 7795.84s] : unite a lot of us,\n",
"[7795.92s --> 7797.52s] : and even if it doesn't,\n",
"[7797.52s --> 7807.44s] : it's just going to make it all feel more positive some When you create an AGI system you'll be one of the few people in the room they get to interact with first,\n",
"[7808.16s --> 7811.44s] : assuming GPT four is not that Ah,\n",
"[7811.76s --> 7815.12s] : what question would you ask her him it?\n",
"[7815.28s --> 7817.12s] : What discussion would you have?\n",
"[7817.68s --> 7818.16s] : You know,\n",
"[7818.32s --> 7841.12s] : one of the things that I realized like this is a little aside and not that important but I have never felt any pronoun other than it towards any of our systems but most other people say him or her or something like that and I wonder why I am so different Like,\n",
"[7841.12s --> 7841.28s] : yeah,\n",
"[7841.28s --> 7841.76s] : I don't know,\n",
"[7841.76s --> 7843.04s] : maybe if I watch it develop,\n",
"[7843.04s --> 7844.40s] : maybe if I think more about it,\n",
"[7844.40s --> 7850.00s] : but I'm curious where that difference comes from I think probably you could because you watch it develop,\n",
"[7850.00s --> 7857.92s] : but then again I watch a lot of stuff develop and I always go to him or her I anthropomorphize aggressively,\n",
"[7858.00s --> 7876.64s] : um and certainly most humans do I think it's really important that we try to explain to educate people that this is a tool and not a creature I think I yes but I also think there will be a room in society for creatures,\n",
"[7876.80s --> 7883.84s] : and we should draw hard lines between those If someone's a creature I'm happy for people to like think of it and talk about it as a creature,\n",
"[7883.92s --> 7911.52s] : but I think it is dangerous to project creatureness on to a tool That's one perspective A perspective I would take if is done transparently is projecting creatureness onto a tool makes that tool more usable If it is done well Yes So if there's like kind of UI affordances that work I understand that I still think we want to be,\n",
"[7911.52s --> 7911.68s] : like,\n",
"[7911.84s --> 7915.68s] : pretty careful with it Because the more creature like it is,\n",
"[7915.68s --> 7922.00s] : the more can manipulate you emotion Or just the more you think that it's doing something,\n",
"[7922.08s --> 7923.52s] : or should be able to do something,\n",
"[7923.52s --> 7928.88s] : or rely on it for something that it's not capable of What if it is capable?\n",
"[7929.12s --> 7931.28s] : What about Sam Alban,\n",
"[7931.28s --> 7933.20s] : what if it's capable of love?\n",
"[7934.08s --> 7939.28s] : Do you think there will be romantic relationships like in the movie Her with GPT?\n",
"[7940.32s --> 7946.48s] : There are companies now that offer like for lack of a better word,\n",
"[7946.48s --> 7961.12s] : like Romantic Companionship AI's replica is an example of such a company Yeah I personally don't feel Any interest in that So you're focussing on creating intelligence,\n",
"[7961.12s --> 7970.64s] : but I understand why other people do That's interesting I'm I've for some reason very drawn to that Have you spent a lot of time interacting with replica or anything?\n",
"[7970.72s --> 7971.68s] : Similar replica,\n",
"[7971.68s --> 7975.20s] : but also just building stuff myself Like I have robot dogs now that I,\n",
"[7975.20s --> 7975.68s] : uh,\n",
"[7976.00s --> 7985.04s] : use I use the movement of the the the robots to communicate emotion I've been exploring how to do that Look,\n",
"[7985.04s --> 7995.28s] : there are going to be very interactive GPT four powered pets or whatever robots,\n",
"[7995.84s --> 8002.16s] : companions and A lot of people seem really excited about that Yeah,\n",
"[8002.32s --> 8004.08s] : there's a lot of interesting possibilities,\n",
"[8004.08s --> 8006.72s] : I think you'll discover them,\n",
"[8006.72s --> 8007.04s] : I think,\n",
"[8007.12s --> 8014.24s] : as you go along That's the whole point Like the things you say in this conversation you might in a year say This was right No,\n",
"[8014.24s --> 8026.32s] : I totally want I may turn out that I like love my GPD Ford Maybe your robot or whatever Maybe you want your programming assistant to be a little kinder and not mock you I do incompetent No,\n",
"[8026.32s --> 8032.72s] : I think you do want um The style of the way GPT four talks to you yes,\n",
"[8032.88s --> 8035.52s] : really matters You probably want something different than what I want,\n",
"[8035.52s --> 8040.48s] : but we both probably want something different than the current GPT four And that will be really important,\n",
"[8040.48s --> 8044.88s] : even for a very tool like thing Is there styles of conversation oh,\n",
"[8044.96s --> 8045.12s] : no,\n",
"[8045.20s --> 8051.12s] : contents of conversations you're looking forward to with an AGI like GPT five,\n",
"[8051.20s --> 8051.52s] : six,\n",
"[8051.60s --> 8060.88s] : seven Is there stuff where like where do you go to outside of the fun meme stuff for actual I mean,\n",
"[8060.88s --> 8065.36s] : what I'm excited for is like Please explain to me how all the physics works,\n",
"[8065.36s --> 8075.12s] : and solve all remaining mysteries So like a theory of everything I'll be real happy Faster than light travel Don't you want to know?\n",
"[8076.08s --> 8083.20s] : So there's several things to know and be hard Is it possible and how to do it?\n",
"[8084.48s --> 8084.72s] : Yeah,\n",
"[8084.72s --> 8087.28s] : I want to know I want to know Probably the first question would be,\n",
"[8087.28s --> 8090.24s] : are there other intelligent alien civilizations out there?\n",
"[8090.32s --> 8098.64s] : But I don't think AGI has the ability to do that to know that Might be able to help us figure out how to go detect?\n",
"[8099.60s --> 8102.24s] : They meant to like send some emails to humans and say,\n",
"[8102.24s --> 8103.36s] : Can you run these experiments?\n",
"[8103.36s --> 8104.72s] : Can you build the space probe?\n",
"[8104.72s --> 8105.36s] : Can you wait,\n",
"[8105.44s --> 8105.76s] : you know,\n",
"[8105.84s --> 8106.64s] : a very long time?\n",
"[8106.72s --> 8109.52s] : Or provide a much better estimate than the Drake equation?\n",
"[8110.64s --> 8112.00s] : With the knowledge we already have,\n",
"[8112.00s --> 8115.60s] : and maybe process all the 'cause we've been collecting a lot of Yeah,\n",
"[8115.84s --> 8116.08s] : no,\n",
"[8116.08s --> 8119.04s] : maybe it's in the data Maybe we need to build better detectors,\n",
"[8119.04s --> 8123.92s] : which it really advanced to Aka tell us how to do It may not be able to answer it on its own,\n",
"[8124.00s --> 8130.32s] : but it may be able to tell us what to go build to collect more data What if it says the aliens are already here?\n",
"[8131.12s --> 8132.88s] : I think I would just go about my life,\n",
"[8132.88s --> 8136.08s] : Yep 'Cause I mean,\n",
"[8136.16s --> 8137.36s] : a version of that is like,\n",
"[8137.36s --> 8139.12s] : what are you doing differently now that,\n",
"[8139.12s --> 8141.84s] : like if if GPT four told you and you believed it,\n",
"[8141.84s --> 8142.24s] : okay,\n",
"[8142.24s --> 8147.68s] : AGI's here Or AGI's coming real soon What are you going to do differently?\n",
"[8147.76s --> 8151.52s] : The source of joy and happiness of fulfillment of life is from other humans,\n",
"[8151.60s --> 8154.00s] : so it's mostly nothing,\n",
"[8154.24s --> 8160.32s] : unless it causes some kind of threat but that threat would have to be like literally a fire Like,\n",
"[8160.40s --> 8166.48s] : are we are we living now with a greater degree of digital intelligence than you would have expected three years ago?\n",
"[8167.76s --> 8183.60s] : And if you could go back and be told by an oracle three years ago which is Blink of an eye that in March of twenty twenty three you will be living with this degree of digital intelligence Would you expect your life to be more different than it is right now?\n",
"[8185.84s --> 8186.80s] : Probably,\n",
"[8186.96s --> 8187.60s] : probably,\n",
"[8187.60s --> 8204.08s] : but there's also a lot of different trajectories intermixed I would have expected the society's response to a pandemic ah to be much better much clearer less divided I was very confused about There's a lot of stuff,\n",
"[8204.16s --> 8207.12s] : given the amazing technological advancements that are happening,\n",
"[8207.28s --> 8218.80s] : the weird social divisions It's almost like the more technological advancement there is the more we're going to be having fun with social division Or maybe the technological advancement just revealed the division that was already there,\n",
"[8218.88s --> 8225.76s] : but all of that just that confuses my understanding of how far along we are as a human civilization,\n",
"[8225.84s --> 8240.80s] : and what brings us meaning and how we discover truth together and knowledge and wisdom So I don't know but when I look when I open Wikipedia I'm happy that humans are able to create this thing First yes,\n",
"[8240.80s --> 8242.24s] : there is biased Yes,\n",
"[8242.32s --> 8247.92s] : let's think it's a triumphal it's the triumph of human civilization Hundred per cent A Google search,\n",
"[8247.92s --> 8248.48s] : the search,\n",
"[8248.56s --> 8249.04s] : search,\n",
"[8249.12s --> 8250.72s] : period is incredible,\n",
"[8250.80s --> 8268.32s] : the way it was able to do twenty years ago And now this this is this new thing GPT is like is this like going to be the next like the conglomeration of all that that made Websearch and Wikipedia so magical,\n",
"[8268.40s --> 8282.72s] : but now more directly accessible you can have a conversation with the damned thing It's incredible Let me ask you for advice for young people in high school and college what to do with their life how to have a career they can be proud of,\n",
"[8282.72s --> 8289.84s] : how to have a life they can be proud of Ah you wrote a blog post a few years ago entitled How to be successful,\n",
"[8290.00s --> 8291.68s] : and there's a bunch of really,\n",
"[8291.68s --> 8300.96s] : really people should check out that blog post They're so it's so succinct and so brilliant You have a bunch of bullet points Compound yourself,\n",
"[8301.28s --> 8302.96s] : have almost too much self belief,\n",
"[8303.04s --> 8304.32s] : learn to think independently,\n",
"[8304.40s --> 8306.32s] : get good at sales and quotes,\n",
"[8306.40s --> 8307.76s] : make it easy to take risks,\n",
"[8307.92s --> 8308.40s] : focus,\n",
"[8308.40s --> 8310.16s] : work hard as we talked about,\n",
"[8310.32s --> 8310.96s] : be bold,\n",
"[8310.96s --> 8311.84s] : be wilful,\n",
"[8312.00s --> 8313.12s] : be hard to compete with,\n",
"[8313.20s --> 8316.72s] : build a network You get rich by owning things,\n",
"[8316.80s --> 8323.84s] : be internally driven What stands out to you from that or beyond is the advice you can give Yet,\n",
"[8323.92s --> 8324.16s] : no,\n",
"[8324.16s --> 8343.76s] : I think it is like good advice in some sense but I also think it's way too tempting to take advice from other people and the stuff that worked for me which I tried to write down there probably doesn't work that well or may not work as well for other people,\n",
"[8343.92s --> 8371.12s] : or like other people may find out that they want to just have a super different life trajectory and I think I mostly got what I wanted by ignoring advice and I think like I tell people not to listen to too much advice Listening to advice from other people should be approached with great caution How would you describe how you've approached life?\n",
"[8372.00s --> 8383.76s] : Outside of this advice that you would advise to other people So really just in the quiet of your mind to think what gives me happiness,\n",
"[8383.92s --> 8385.28s] : what is the right thing to do here,\n",
"[8385.28s --> 8401.60s] : how can I have the most impact I wish it were that you know introspective all the time It's a lot of just like you know what will bring me joy will bring me fulfillment You know what will bring what will be?\n",
"[8401.68s --> 8402.08s] : Ah,\n",
"[8402.32s --> 8404.96s] : I I do think a lot about what I can do that will be useful,\n",
"[8404.96s --> 8407.36s] : but like who do I want to spend my time with?\n",
"[8407.44s --> 8409.04s] : What do I want to spend my time doing?\n",
"[8409.36s --> 8414.00s] : Thish and water is going wrong with the That's certainly what it feels like I mean,\n",
"[8414.00s --> 8427.44s] : I think that's what most people would say if they were really honest about it If they really think yeah And some of that then gets to the Sam Harris discussion of free will being an illusion Of course very well might be,\n",
"[8427.52s --> 8436.08s] : which is a really complicated thing to wrap your head around What do you think's the meaning of this whole thing?\n",
"[8437.04s --> 8441.04s] : That's the question you can ask an AGI What's the meaning of life?\n",
"[8441.52s --> 8443.12s] : As far as you look at it,\n",
"[8443.44s --> 8449.36s] : you're part of a small group of people that are creating something truly special,\n",
"[8449.76s --> 8476.64s] : something that feels like almost feels like humanity was always moving towards Yeah That's what I was going to say is I don't think it's a small group of people I think this is the I think this is like the product of the culmination of whatever you want to call it an amazing amount of human effort and if you think about everything that had to come together for this to happen When those people discovered the transistor in the forties,\n",
"[8476.64s --> 8512.08s] : like is this what they were planning on all of the work the hundreds of thousands of millions of people to ever it's been that it took to go from that one first transistor to packing the numbers we do into a chip and figuring out how to wire them all up together and everything else that goes into this the energy required the the the science at like just every step like this is the output of like all of us and I think that's pretty cool And before the transistor there was a hundred billion people who lived and died,\n",
"[8512.72s --> 8513.68s] : had sex,\n",
"[8513.84s --> 8514.80s] : fell in love,\n",
"[8515.12s --> 8516.40s] : ate a lot of good food,\n",
"[8516.48s --> 8521.92s] : murdered each other sometimes rarely but mostly judges good to each other struggled to survive,\n",
"[8521.92s --> 8529.76s] : and before that there was bacteria and eukaryotes and all that And all of that was on this one exponential curve Yeah,\n",
"[8529.84s --> 8531.20s] : how many others are there,\n",
"[8531.28s --> 8531.92s] : I wonder?\n",
"[8532.08s --> 8536.32s] : We will ask that isn't question number one for me for Aegia How many others?\n",
"[8536.64s --> 8540.00s] : And I'm not sure which answer I want to hear Sam,\n",
"[8540.00s --> 8544.72s] : you're an incredible person It's an honor to talk to you Thank you for the work you're doing Like I said,\n",
"[8544.72s --> 8545.76s] : I've talked to Ilios,\n",
"[8545.76s --> 8546.88s] : the Scare I talked to Gregg,\n",
"[8546.88s --> 8567.04s] : I talked to so many people at open AI They're really good people They're doing really interesting work We are going to try our hardest to get to get to a good place here I think the challenges are Tough I understand that not everyone agrees with our approach of iterative deployment and also iterative discovery Um but it's what we believe in Ah,\n",
"[8567.04s --> 8591.76s] : I think we're making good progress And I think the pace is fast But so is the progress So so like the pace of capabilities and change is fast but I think that also means we will have new tools to figure out alignment and sort of the capital S safety problem I feel like we're in this together I can't wait what we together as a human civilization come up with It's going to be great,\n",
"[8591.76s --> 8598.00s] : I think We'll work really hard to make you Thanks for listening to this conversation with Sam Altman To support this podguest,\n",
"[8598.08s --> 8617.84s] : please check out our sponsors in the description And now let me leave you with some words from Alan Turing in nineteen fifty one It seems probable that once the machine thinking method has started it would not take long to outstrip our feeble powers At some stage,\n",
"[8617.84s --> 8618.48s] : therefore,\n",
"[8618.80s --> 8625.52s] : we should have to expect the machines to take control Thank you for listening,\n",
"[8625.60s --> 8627.36s] : and hope to see you again next time\n",
"time: 76.9 ms (started: 2023-10-19 19:01:11 -07:00)\n"
]
}
],
"source": [
"# Util to print ctm file\n",
"import re\n",
"\n",
"def read_ctm(filepath):\n",
" ctm_data = []\n",
" with open(filepath, 'r', encoding='utf8') as f:\n",
" for line in f:\n",
" # Structure of ctm\n",
" # 0: File name\n",
" # 1: Channel\n",
" # 2: Start time (in s)\n",
" # 3: Duration (in s)\n",
" # 4: Text with \" \" replaced as \"<space>\" to prevent issues when splitting by space.\n",
" parts = line.split(\" \")\n",
" start_time = float(parts[2])\n",
" end_time = start_time + float(parts[3]) # end = start + duration\n",
" text = parts[4].replace(\"<space>\", \" \")\n",
"\n",
" if text[-1] == '\\n':\n",
" text = text[:-1]\n",
"\n",
" ctm_data.append([start_time, end_time, text])\n",
"\n",
" # Merge together CTM segments to separate on just special characters.\n",
" merged_ctm = [[]]\n",
" for data in ctm_data:\n",
" if data[2] in [',', '.', '?']:\n",
" merged_ctm[-1].append(data)\n",
" merged_ctm.append([])\n",
" else:\n",
" merged_ctm[-1].append(data)\n",
" \n",
" return merged_ctm\n",
"\n",
"\n",
"def print_ctm(ctm_data):\n",
" for data in ctm_data:\n",
" start_time = data[0][0] # start timestamp of the first ctm segment in list\n",
" end_time = data[-1][1] # end timestamp of the last ctm segment in list\n",
" text = [d[2] for d in data] # text from each of the ctm segments in list\n",
" text = \" \".join(text) # merge with space\n",
" text = text = re.sub(r'(\\s+)([\\.\\,\\?])', r'\\2', text) # merge the space emitted between special character and previous words.\n",
" \n",
" timestamp = f\"[{start_time:0.2f}s --> {end_time:0.2f}s]\"\n",
" print(f\"{timestamp:<20} : {text}\")\n",
"\n",
"ctm = read_ctm(\"nfa_output/ctm/words/sam_altman_lex_podcast_367.ctm\")\n",
"print_ctm(ctm)"
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "3cd6598d-8001-47e6-bb41-ba5b22dd2ee3",
"metadata": {},
"outputs": [],
"source": []
}
],
"metadata": {
"kernelspec": {
"display_name": "Python 3 (ipykernel)",
"language": "python",
"name": "python3"
},
"language_info": {
"codemirror_mode": {
"name": "ipython",
"version": 3
},
"file_extension": ".py",
"mimetype": "text/x-python",
"name": "python",
"nbconvert_exporter": "python",
"pygments_lexer": "ipython3",
"version": "3.10.13"
}
},
"nbformat": 4,
"nbformat_minor": 5
}
@RomanLeo2003
Copy link

Why that part:
decoding_cfg = model.cfg.decoding decoding_cfg.preserve_alignments = True # Preserve alignment information using models logits decoding_cfg.compute_timestamps = True # Compute greedy word timestamps

Throws:

`ConfigAttributeError Traceback (most recent call last)
in <cell line: 2>()
1 decoding_cfg = model.cfg.decoding
----> 2 decoding_cfg.preserve_alignments = True # Preserve alignment information using models logits
3 decoding_cfg.compute_timestamps = True # Compute greedy word timestamps
4
5 print("Current decoder type:", model.cur_decoder)

9 frames
/usr/local/lib/python3.10/dist-packages/omegaconf/_utils.py in _raise(ex, cause)
795 else:
796 ex.cause = None
--> 797 raise ex.with_traceback(sys.exc_info()[2]) # set env var OC_CAUSE=1 for full trace
798
799

ConfigAttributeError: Key 'preserve_alignments' is not in struct
full_key: decoding.preserve_alignments
object_type=dict`

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment