Grandmaster-Level Chess Without Search
A new chess AI claims “grandmaster-level” play using a transformer model that chooses moves without running a traditional search during games, instead learning from millions of positions evaluated by the Stockfish engine. Commenters debate whether this still counts as “without search,” given its heavy reliance on Stockfish for training data and some endgame assistance, and compare it to systems like AlphaZero and Leela that integrate search directly. The exchange highlights broader questions about what it means for neural networks to “replace” classical algorithms, how much of their strength comes from precomputed expert knowledge, and whether such models can match engine-level precision or mainly excel against humans at fast time controls.
Use of Stockfish and the “without search” claim
- Major debate over the title: critics argue it’s misleading because Stockfish search is used to label 15B training positions and to pick moves in some winning positions.
- Defenders distinguish between using search during training vs not calling a search procedure at inference time; they claim “no search at play time” is still meaningful.
- Some call this “precompiled search” or “neuro‑symbolic,” arguing the system’s capabilities fundamentally depend on a classical search engine.
Endgames and handling of won positions
- The model often fails to reliably convert overwhelming advantages or short mates because all strong moves are labeled “100% win.”
- This leads to random-looking play in clearly winning positions; Stockfish is invoked when the model’s top moves all exceed a 99% predicted win rate.
- Critics say this Stockfish “patch” is non-trivial and would not be acceptable in real competition; others frame it as a bug workaround that doesn’t change the core result.
Performance and evaluation
- The system achieves grandmaster-level ratings against humans in online blitz but performs much worse against engines.
- Some see this “aggressive, human-beatable but engine-weak” style as more entertaining and human-like; others stress that strong endgame technique and longer-time-control strength are essential for true GM equivalence.
- Concerns raised about evaluation conditions, especially reliance on fast time controls.
Search, memorization, and implicit computation
- Ongoing debate whether the transformer is doing any “implicit search” or mainly learning a lossy approximation / interpolation of Stockfish evaluations.
- Some argue pure memorization is impossible given the enormous state space; others note large nets routinely memorize complex datasets and still behave fuzzily.
- Depth vs width scaling behavior is mentioned as possible evidence of learned multi-step algorithms.
Practical significance and limitations
- Enthusiasm for compressing search into a feedforward model that runs efficiently on GPUs/TPUs and can batch thousands of position evaluations.
- Skeptics note Stockfish is already tiny and fast on CPUs; the large model may be more of an academic exercise than a practical engine.
- Broader reflections connect this to bootstrapped AGI, data dependence on existing engines/humans, and comparisons to AlphaZero-style self-play.