search

adamkarvonen / chess_llm_interpretability

Visualizing the internal board state of a GPT trained on chess PGN strings, and performing interventions on its internal board state and representation of player Elo.
190 stars 13 forks source link

Can you try to incorporate self-play technology similar to AlphaGo Zero? #1

Open win10ogod opened 8 months ago

win10ogod commented 8 months ago

Can you try to incorporate self-play technology similar to AlphaGo Zero?

adamkarvonen commented 8 months ago

For now, my goal is more focused on the interpretability of an LLM trained to play chess, rather than making a better chess LLM. When making a better chess LLM I would first explore using a larger model trained on more data with more compute, and identifying which one of those factors is the current Elo bottleneck. This would probably be more compute efficient than a self-play approach. We have an effectively unlimited amount of human chess games data, and an existence proof in GPT-3.5-turbo-instruct that an LLM can play at 1800 Elo off of human chess games.

However, self play with LLMs could also be interesting as well. I currently have no plans of exploring this, but anyone is welcome to try.