Otter AI Predicts Human Chess Moves with Time and History

A new human chess AI named Otter has been developed, boasting 15.3 million parameters. Unlike previous models that analyze chess positions in isolation, Otter models human play as a dynamic, time-aware sequential process. This approach allows it to predict human move selection more accurately by incorporating two crucial conditioning signals. Otter's predictions are influenced by a move history encoder, which considers the last 20 moves to capture opening preferences, positional drift, and intra-game behavioral tendencies. Additionally, a time control module modulates predictions based on the pressure of the game clock. The model was trained on an extensive dataset of 6.1 billion positions from 117 million Lichess rapid games over 30 days using a single T4 GPU. The results demonstrate Otter's superior performance, achieving 55.23% top-1 and 90.95% top-5 move-prediction accuracy. This significantly surpasses Maia 2, the prior state-of-the-art human chess model, despite Otter using substantially fewer parameters and less training data. Its accuracy peaks at 57.38% in the 1900-1999 Elo bracket, confirming that modeling chess as a time-aware, sequential activity yields more human-accurate predictions. The code and trained models are publicly available.