SpinGPT - Training a Poker LLM
Date:
At the Advances in Computer Games conference (21–23 October 2025), I presented the main results of my work on SpinGPT, a poker AI built on top of a large language model.
SpinGPT is trained in two stages. First, it is fine-tuned on 320,000 hands from my former professional poker history. Second, it is trained with reinforcement learning on 320,000 hands generated by a solver*. After these two phases, SpinGPT selects the game-theoretically best action roughly four times out of five and is able to defeat Slumbot, the reigning poker AI world champion.
*A poker solver is a program that computes strategies close to the game-theoretic optimum in specific game situations.
The slides of the talk are available here, along with the video.