OPEN LINK ↗
// 66d agoPRODUCT UPDATE
Autochess V3 hits ~2700 Elo
Autochess is a browser-playable neural chess engine that combines a residual CNN, transformer layers, and learned thought tokens with a training pipeline spanning Lichess pretraining, Syzygy fine-tuning, and self-play RL. The project also ships an inspectable web UI with play, editor, PGN replay, puzzles, and move analysis, so it reads like a real research artifact rather than just a benchmark claim.
// ANALYSIS
This feels like a legitimately ambitious systems project, not a thin AI wrapper. The architecture and UX are compelling, but the Elo story will only land if the evaluation methodology is pinned down much more tightly.
- –The browser experience is a real strength: play, board editor, replay, puzzles, and move-probability visualizations make the model easy to inspect.
- –The rating claim is interesting but fragile unless the opponent pool, time control, opening rules, and exact ladder setup are documented clearly.
- –Thought tokens, DAB, and “Temporal Look-Ahead” are promising hypotheses, but they need ablations against a simpler CNN-plus-search baseline before anyone can tell if they add real signal.
- –For classical-engine comparisons, fixed test suites, fixed time or node budgets, and repeated runs will be more convincing than casual human-vs-bot outcomes.
// TAGS
researchbenchmarkgpuinferenceautochess
DISCOVERED
66d ago
2026-03-21
PUBLISHED
67d ago
2026-03-21
RELEVANCE
8/ 10
AUTHOR
Adam_Jesion
