Wherein we explore interpretable world models in Connect 4. Repository.

Current understanding

A while ago I read the Othello GPT paper and decided that it would be interesting to reproduce it. But I thought that reusing Othello would be a bit boring, so I decided to try the game of Connect 4 instead. Although Connect 4 is much simpler, I think we'll still be able to find something gripping (I hope).

Open questions

  • Is the design of the game inviting enough for the model to build a world model at all?
  • What is "championship" dataset for Connect 4?
  • I also didn't read the Neel Nanda's follow-up paper where they suggest using "empty/mine/yours" perspective. I deliberately want to postpone the reading, so that I'll try figure that out as needed. But now I'm certainly intrigued to understand why this approach helped!