ChessLab
A chess learner that starts from random weights and is never told anything about chess.
The learner gets the rules of chess and nothing else: no openings, no piece values, no engine scores as targets. It plays itself, trains a candidate on its own games, and the candidate only becomes the new champion if it beats the old one over enough games to prove it.
Everything is measured against something the learner cannot author. Fixed opponents that never drift, frozen mate positions, and its own past generations, archived so they can be replayed.
The interesting findings are the ones that contradicted the plan. A network sixteen times bigger learned nothing more. Half the time went to Python’s garbage collector. And the search was the ceiling, not the model: the same network solves a third more mates when it is allowed to think longer.
- Proven losses made unplayable on every path
- Champion grows in place instead of restarting
- Bot opens to human players on Lichess
- Bot begins seeking Lichess opponents, one at a time
- Per-node board cost cut 2.2x, search 14% faster
- Generation 21 reaches 87 on the forced-mate probe
- Chess-gpt gap closes from 382 to 108 Elo
- Gumbel AlphaZero added as an optional search mode
- Four times the thinking, exactly the same progress per hour
The learner searched 48 positions per move before storing what it learned from. I raised it to 200 and let it run a day. Per cycle it…