Show HN: Chess bot based on the transformer architecture
Category: ai-ml
Tags: chess, transformer, mcts, ai-engine, deep-learning
Score: 8.0/10 (Innovation: 7, Technical: 9, Documentation: 9, Utility: 7)
This project is a transformer-based chess engine trained solely on human games, achieving ~2100 Elo without reinforcement learning or self-play. It combines a custom 11.7M-parameter transformer with AlphaZero-style MCTS search, optimized inference via torch.compile and CUDA graphs, and offers both a web demo and live Lichess bot. The novel use of chess-geometric relative position biases and the impressive scaling of strength with search simulations make it a standout contribution to AI game-playing.
Target audience: AI/ML researchers, game developers, chess enthusiasts
Repository: https://github.com/tchauffi/ChessTransformer · Jupyter Notebook · MIT
View on Hacker News