Contra "Grandmaster-level chess without search" (2024)

2 days ago

Google DeepMind developed a transformer-based model for chess, trained on Stockfish 16, predicting state value, action value, and action probabilities.
The model's performance is compared to AlphaZero, with claims of Grandmaster-level play, evidenced by a Lichess Blitz rating of 2895.
Criticism includes the model's inability to improve with more time and potential redundancy given existing open-source projects like Leela Chess Zero (Lc0).
The paper's methodology and conclusions are questioned, especially the comparison to human masters instead of stronger engines like Stockfish.
An addendum compares searchless puzzle-solving abilities across AlphaZero, the paper's model, and Lc0.

Hasty Briefsbeta