Contra "Grandmaster-Level Chess Without Search" (cosmo.tardis.ac)

Google DeepMind has introduced a new transformer-based model for chess called "Amortized Planning with Large-Scale Transformers," which was trained using the strong chess engine Stockfish 16. This approach allows the model to analyze game states by predicting the state value, action value, and probability distribution over possible moves in a way that mimics the output of a rapid 50ms Stockfish search. The authors assert that their model can play at a human Grandmaster level, evidenced by achieving a Lichess Blitz rating of 2895. However, critics highlight that the model's strength may diminish in longer time controls, indicating it might be less effective in standard chess formats. The significance of this development lies in its implications for AI strategy in game-playing contexts, building on concepts introduced by DeepMind's earlier work with AlphaZero. While the new model's architecture is similar, skeptics argue that it does not represent a groundbreaking advancement, pointing to the success of open-source projects like Leela Chess Zero, which have surpassed AlphaZero's capabilities. Furthermore, the authors' claims about the model potentially surpassing Stockfish's performance raise questions about the validity of their comparison methodologies. Overall, while the paper claims substantial progress, it appears to overlook critical developments in the chess AI landscape, suggesting the need for careful scrutiny and validation of their results.