Why Stockfish Falls Short for Human Chess Review and How Maia 2 Helps

Chess players often find that engine analysis flags natural-feeling moves as blunders, since engines like Stockfish seek only the objectively strongest move. The author, a developer and chess player, argues that reviewing mistakes also requires knowing what a human at a given rating would likely play. Maia 2, a model from University of Toronto researchers with collaborators at Cornell and Microsoft Research, is trained on millions of human Lichess games and predicts typical human moves rather than optimal ones. The author describes running Maia 2 client-side in the browser using ONNX Runtime Web and WebAssembly, without an inference server. The integration was chosen for a puzzle review workflow in which the author logs and analyzes failed puzzles.