veekaybee

Anti-hype LLM reading list

Goals: Add links that are reasonable and good explanations of how stuff works. No hype and no vendor content if possible. Practical first-hand accounts of models in prod eagerly sought.

Foundational Concepts

Pre-Transformer Models

yoavg

Reinforcement Learning for Language Models

Yoav Goldberg, April 2023.

Why RL?

With the release of the ChatGPT model and followup large language models (LLMs), there was a lot of discussion of the importance of "RLHF training", that is, "reinforcement learning from human feedback". I was puzzled for a while as to why RL (Reinforcement Learning) is better than learning from demonstrations (a.k.a supervised learning) for training language models. Shouldn't learning from demonstrations (or, in language model terminology "instruction fine tuning", learning to immitate human written answers) be sufficient? I came up with a theoretical argument that was somewhat convincing. But I came to realize there is an additional argumment which not only supports the case of RL training, but also requires it, in particular for models like ChatGPT. This additional argument is spelled out in (the first half of) a talk by John Schulman from OpenAI. This post pretty much

mvidaldp

I wrote this short tutorial because extending my internal storage using my new micro SD card on my Retroid Pocket 2+ failed all the time. Only setting it up as portable/external worked. However, this instructions should work in any Android 5.0+ device.

So, in case you have problems setting up your SD card on your Android device via graphical interface (setting up storage as extended internal memory or portable), and you get a corrupted SD card or any other error, follow these steps to fix it via adb shell:

1. Make sure you have adb access to your Android device: Settings > System > About, touch/click on Build number until Developer options are enabled:
2. Go to Settings > System > Developer options and enable USB debugging.
3. Assuming you have adb installed on your remote terminal run the following:

adb shell

cassiozen

Pixelbook Setup

Change your channel

Some of the features mentioned in this document only work on the beta or Dev channel. To change your channel:

1. chrome://help in a browser window
2. Click Detailed Build Information
3. Change Channel
4. Select Beta (Or Dev, if you're feeling adventurous)

halhen

# data from http://ec.europa.eu/eurostat/web/gisco/geodata/reference-data/population-distribution-demography/geostat
# Originally seen at http://spatial.ly/2014/08/population-lines/

# So, this blew up on both Reddit and Twitter. Two bugs fixed (southern Spain was a mess,
# and some countries where missing -- measure twice, submit once, damnit), and two silly superflous lines removed after
# @hadleywickham pointed that out. Also, switched from geom_segment to geom_line.

# The result of the code below can be seen at http://imgur.com/ob8c8ph

library(tidyverse)

aparrish

inchoate

The Problem

The documentation for how to deploy a pipeline with extra, non-PyPi, pure Python packages on GCP is missing some detail. This gist shows how to package and deploy an external pure-Python, non-PyPi dependency to a managed dataflow pipeline on GCP.

TL;DR: You external package needs to be a python (source/binary) distro properly packaged and shipped alongside your pipeline. It is not enough to only specify a tar file with a setup.py.

Preparing the External Package

Your external package must have a proper setup.py. What follow is an example setup.py for our ETL package. This is used to package version 1.1.1 of the etl library. The library requires 3 native PyPi packages to run. These are specified in the install_requires field. This package also ships with custom external JSON data, declared in the package_data section. Last, the setuptools.find_packages function searches for all available packages and returns that

afeld

Disclaimer: The majority of this list was created pre-COVID. Many other organizations are likely hiring remote now.

---

• A1M Solutions (distributed)
• Ad Hoc (distributed)
• ACLU Product & Technology Department
• Beeck Center (at Georgetown)
• BetaNYC
• Blue Ridge Labs

chrishwiggins

FAQ:

where are some fun datasets to play with?

1. CMU:
http://lib.stat.cmu.edu/datasets/

2. UCI:

a) MLR@UCI (machine learning repository / machine learning archive )

chrishwiggins

Q: what book should i use to learn ML?

A: use several, and find the one that speaks to you.

the list below assumes you know a bit of math but
are not very mathematical, and are interested in learning
enough to be practical. that is, it is not at the
mathematical level of MIJ's alleged list
(cf. https://news.ycombinator.com/item?id=1055389 )