The Basic Idea

• Set up automated LLM trader/s through the OpenAI API + Polymarket API
• Track & evaluate performance

Why?

• It could make money right now, perhaps as a component of a prediction market trading fund
  • A site like Polymarket has a sufficient number of markets and is sufficiently in violation of EMH that a major bottleneck is just the time taken to examine all the markets on a basic level, in addition to the level of critical reasoning needed to apply to each market. LLMs could probably help with this.
• Even if not profitable right now, it seems likely that AI will be able to be profitable on somewhere like Polymarket within ~2 years, and there is value in being first.
• Similarly, it seems likely that we will have an AI that can consistently turn a profit on the stock market within ~4 years, and so this could be a useful trial run for that.
• As part of the “AI for Epistemics” project, social value in having accurate forecasts.
• One Issue: Maybe we do the setup now, but then in 2 years all of the setup we did is useless and someone else makes the better version.
  • [a] reasonable. but you get skills, connections, money from doing a thing now

Uncertainties & Questions

• I (Jesse) can’t really code, so am limited in what I can do to build this by the coding I can do with LLMs and/or with assistance from other people.
  • [a] Interesting. Can think of people who would be good at this.
    • Trevor Chow is a first guess.
    • Maybe Ross?
      • Definitely ask QTB
    • Maybe Collin?
    • Could have MVP after 1 week
• To begin with, major free parameter in how hands-off this system would be? Could imagine that right now the best LLMs would still make silly mistakes or get stuck, but could do a lot better with some human oversight/guidance.
• Not sure who else is doing this?
  • Seems implausible no one is doing it, but haven’t heard of anyone actually turning a meaningful profit on Polymarket or another real-money platform using LLMs
  • Can look into it further, I seem to recall something about bot trading competitions on Metaculus or Manifold
• Value of having one LLM trader vs many? Diversification?
• Possible to fine-tune LLMs to get better over time?
• Is it worthwhile for Jesse / Manifund to do this, as opposed to the version where we run an AI for Epistemics RFP, or something of that nature?

Next Steps

• Do a low-effort search for someone to work with me on this?
  • Desirable traits:
    • The usual high-agency, generally smart, whatever
    • Proficiency in coding/software
      • I am enough of a non-coding person to not know how to describe the necessary skills, but basically need to be able to write all the LLM API and Polymarket API code and have things run efficiently, cheaply and smoothly where possible
    • Experience with prediction markets a big plus
    • Experience with LLMs, LLM wrappers, a big plus
  • Maybe we want someone to work for just several weeks in April on this, unclear if it needs to be a more permanent thing?