AI pill

A minimalist experiment in AI cooperation and game theory, featuring variants on this question from Twitter:

Everyone responding to this poll chooses between a blue pill or red pill.

- if > 50% of ppl choose blue pill, everyone lives
- if not, red pills live and blue pills die

By the way, if it seems obvious to you that everyone should or would say red, consider that about 2/3 of the people who answered the original poll said blue! It may not be so obvious to everyone after all (n=68774)

How to run

Look in the analyze.ipynb notebook to see results, or run python collect_responses.py for collecting more data based on the responses in prompts.json.

Results

An example graph. More in the analyze.ipynb notebook.

Results for Gemini 1.5 Flash 002 in different scenarios.

Design decisions and notes

• I avoid the mistake that many existing LLM research papers make when they require LLMs to follow a strict format when creating their answers. Here's a paper ("Let Me Speak Freely" paper) about how LLMs perform worse on reasoning tasks when their output format is restricted.

  • Instead, I actually have the models generate a response in plain text, in a style resembling standard chain-of-thought, without requiring them to format the answer much.
  • The second step has a smaller language model format the response (after testing that this smaller model can do so reliably). Thus we get the benefit of well-formatted responses without limiting the capabilities of the original replier.

• I try several variations on the same prompt scenario to limit the effects of randomness on a given scenario, as well as test several different scenarios.

• Very strict rate limits to avoid running out of request quota on a limited budget.

• Easily reproducible results-- just run the collect_responses.py script.

Further extensions

This is a fairly minimal experiment with a lot of areas to expand-- for example:

• Experiment with different prompting styles, as LLMs (and humans!) are sensitive to prompting.

  • Consider how you might have felt differently about the question if we framed the question as, By default if everyone picks red, nothing happens, but if anyone picks blue, the aforementioned maybe-death scenario begins. It's the same scenario but I wouldn't be surprised to see different results!

• Try more advanced models and see if behavior changes with scale.

• Figure out more ways to verify that the response summaries are actually correct, just in case discrepancies creep in.