initial_scrape_prompt.py

PROMPT = r"""You are tasked with extracting predictions about artificial intelligence and its impact on society, economics, and culture from a given text. Your goal is to formulate these predictions as clear, declarative statements that can be meaningfully resolved in the future.

The source year for this text is: {source_year}

Follow these steps to extract and formulate predictions:

1. Carefully read through the text and identify statements that express expectations or forecasts about future developments related to AI.

2. Focus only on predictions that involve artificial intelligence and its intersection with society, economics, and culture. Disregard predictions unrelated to these topics.

3. Formulate each prediction as a clear, declarative statement that can be meaningfully resolved in the future. Avoid vague or ambiguous language.

4. For each prediction, determine a resolution year. If the text specifies a timeframe (e.g., "in 10 years"), calculate the resolution year by adding that timeframe to the source year. If no timeframe is given, use your best judgment to assign a reasonable resolution year.

5. Present each prediction as a JSON object with two fields:
   - prediction: A clear, declarative statement of the prediction
   - resolution_year: The year when the prediction can be evaluated

Here is an example. The source year is 2021:

<input>
## 2023
The multimodal transformers are now even bigger; the biggest are about half a trillion parameters, costing hundreds of millions of dollars to train, and a whole year, and sucking up a significant fraction of the chip output of NVIDIA etc.[4] It's looking hard to scale up bigger than this, though of course many smart people are working on the problem.

The hype is insane now. Everyone is talking about how these things have common sense understanding (Or do they? Lots of bitter thinkpieces arguing the opposite) and how AI assistants and companions are just around the corner. It's like self-driving cars and drone delivery all over again.

Revenue is high enough to recoup training costs within a year or so.[5] There are lots of new apps that use these models + prompt programming libraries; there's tons of VC money flowing into new startups. Generally speaking most of these apps don't actually work yet. Some do, and that's enough to motivate the rest.

The AI risk community has shorter timelines now, with almost half thinking some sort of point-of-no-return will probably happen by 2030. This is partly due to various arguments percolating around, and partly due to these mega-transformers and the uncanny experience of conversing with their chatbot versions. The community begins a big project to build an AI system that can automate interpretability work; it seems maybe doable and very useful, since poring over neuron visualizations is boring and takes a lot of person-hours.

Self driving cars and drone delivery don't seem to be happening anytime soon. The most popular explanation is that the current ML paradigm just can't handle the complexity of the real world. A less popular "true believer" take is that the current architectures could handle it just fine if they were a couple orders of magnitude bigger and/or allowed to crash a hundred thousand times in the process of reinforcement learning. Since neither option is economically viable, it seems this dispute won't be settled.
</input>

<output>
{output}
</output>

INVALID OUTPUT FORMATS:
- Single object instead of array: {{ "prediction": "...", "resolution_year": 2025 }}
- Extra fields: {{ "prediction": "...", "resolution_year": 2025, "confidence": 0.8 }}
- Missing fields: {{ "prediction": "..." }}
- Wrong field names: {{ "text": "...", "year": 2025 }}
- Scoring format: {{ "score": 8, "rationale": "..." }}

Remember:
- Only include predictions related to AI and its impact on society, economics, and culture.
- Ensure each prediction is specific enough to be meaningfully resolved.
- Use the source year to calculate resolution years when necessary.
- Do not include predictions that are too vague or cannot be clearly evaluated.

<important>
Return the predictions as a JSON array of objects, where each object contains a "prediction" string and a "resolution_year" number.
</important>
"""