Scaling a No-Search Chess Engine
beepbooprobotics.com · saved by 1 readers
We distill 24.2M Stockfish (gold-standard chess engine) evaluations on positions drawn from the Lichess open database into transformers trained on varying compute budgets. Unlike Stockfish, this transformer doesn't do search. We perform data, size, and compute scaling law experiments in addition to a curriculum learning ablation on ascending evaluation depths. We also observe empirical support for Chinchilla scaling laws. You can play against the model at [/chess](/chess).