Grandmaster-level chess without search

Original Article ↗ Hacker News Discussion ↗

A new DeepMind project trains a 270M-parameter transformer to play chess at near–grandmaster blitz strength using only a single forward pass, after distilling evaluations from the traditional search-based engine Stockfish. Commenters debate how meaningful “without search” really is when the training data itself comes from heavy search, contrast the model’s performance with modern engines like Stockfish NNUE and Leela, and explore whether such systems can be adapted to play more human-like chess or to generalize beyond narrowly defined rules and tokenization schemes.

Core idea & method

  • Paper is seen as a knowledge-distillation result: distilling Stockfish’s complex search-based evaluation into a single transformer “intuition” function over board states.
  • Model: ~270M parameters, trained on ~10M games, with ~15B positions annotated by Stockfish 16 (action-values).
  • Input is FEN plus UCI moves via a custom tokenizer; no explicit search or domain heuristics at inference.

Performance & evaluation

  • Reported Lichess blitz rating vs humans is ~2895; against bots it is ~700 Elo lower.
  • It outperforms AlphaZero’s policy/value networks without MCTS and GPT‑3.5‑turbo‑instruct on puzzles/chess strength.
  • Some commenters argue the blitz human results are inflated: the bot never flags on time, humans do, and Stockfish fallback in tactical crises may convert lost/drawn positions to wins.
  • Requests for additional baselines: limited-depth Stockfish, positions that require deep search, “anti-bot” openings.

Relation to existing engines

  • Debate over whether this advances practical engine strength: top search-based engines (Stockfish NNUE, Leela) are said to remain clearly stronger under standard conditions.
  • Some note Leela-style transformer work predates this paper and achieves better strength with smaller models.
  • Others see value as a fast GPU-friendly evaluator that could be combined with search for efficient parallel exploration.

Human-like play & non-GM engines

  • Many participants want engines that play like humans at specific ratings, not just “weakened” super-engines.
  • Suggestions include:
    • Sampling among near-best moves instead of strict best.
    • Conditioning on Elo during training.
    • Training to predict human moves (as in models like Maia).
    • Modeling typical human blunders, attention focus, and style.
  • Several existing projects and ideas are mentioned, but consensus is that truly human-like adjustable engines remain an open problem.

Critiques & conceptual issues

  • Some argue the “without search” title is misleading because generating labels required massive Stockfish search; the model is essentially a compressed approximation of Stockfish’s search tree.
  • Others frame this as a standard and valuable form of knowledge distillation (“compiling search into a single forward pass”).
  • There is debate over whether custom tokenization and problem-specific setups reduce the generality or significance of the result.

Broader chess-AI context

  • Ongoing discussion about whether chess is “solved” (consensus: only small endgames are; full-game optimal play remains far beyond reach).
  • The work is taken as evidence that high-level play can, in principle, be approximated by a pure evaluation heuristic if trained at sufficient scale.