nn-chess

Play against a neural-network opponent that learns solely from the win/loss signal of games it plays against you — no opening books, no piece tables, no hand-coded evaluation. The opponent starts out terrible and (hopefully) improves as it accumulates games. Use the Opponent dropdown below to switch to the player-imitator (interpretability), Stockfish (calibrated weak), or the human-style Maia-2 network — your moves train the imitator in every mode; the opponent NN's replay buffer is only touched when you play against it directly. See the nn-chess README for the full design: goals, architecture, roadmap, and the deliberate constraints that keep the learning signal honest. There's also a technical write-up of how MCTS picks each move with an interactive mate-in-2 walkthrough driven by your own networks.

Stockfish sparring → — watch the opponent learn from automated games against Stockfish.