To predict a person's next move, the AI had to take the time of day into account

Predicting a human player’s next move in chess turned out to be more important than simply evaluating the position well. Otter takes into account the last 20 moves and the remaining time, so it tries to understand not the ideal move, but what a specific player is most likely to do right now.

This is the key shift. The traditional approach views the position almost in isolation: the pieces are arranged this way—the model suggests a probable move. Otter adds the context of the game: opening preferences, gradual changes in the position, and the player’s individual habits. A separate module accounts for time control and time pressure.

The result is particularly interesting given the model’s size. Otter contains 15.3 million parameters, was trained on 6.1 billion positions from 117 million rapid games on Lichess, and outperformed the previous Maia 2 model while using significantly fewer parameters and data.

The accuracy of predicting the best move was 55.23%, and one of the five most likely moves matched the human move in 90.95% of cases. The model achieved its highest accuracy—57.38%—among players with ratings between 1900 and 1999. In other words, context helps not only beginners: the decisions made by strong players also differ significantly from the abstract “best” move.

But this is not a new chess champion. Otter predicts human choices; it does not prove that it plays chess better than humans. Furthermore, the model was trained on rapid games on Lichess, so one cannot conclude from the description that it performs just as well in blitz, classical chess, or outside this gaming environment. The code, trained models, and training logs have been published, but practical testing on other datasets is even more important than impressive percentages.

If an AI learns to predict your moves, which is more important to you: your history of moves or time pressure? Source: cs.AI updates on arXiv.org