Chess engines do weird stuff
girl.surgery
February 17, 2026
4 min read
🔥🔥🔥🔥🔥
55/100
Summary
Chess engines like AlphaZero and lc0 use reinforcement learning by having the engine play itself multiple times to train the model on game outcomes. A combination of a weaker model and strong search capabilities can outperform a stronger model alone, as the search can significantly enhance performance.
Key Takeaways
- lc0-style chess engines can distill a good model from a bad model combined with search, eliminating the need for extensive self-play training after the initial model is developed.
- A new runtime technique allows chess engines to adapt their evaluations based on real-time search results, improving accuracy in position assessments.
- The SPSA technique used by lc0 involves random weight perturbations to enhance model performance, achieving significant elo gains despite being computationally expensive.
- lc0's transformer architecture outperforms previous convolution-based models, with a specific system called "smolgen" providing substantial accuracy improvements.
Community Sentiment
Positives
- The use of SPSA for tuning in chess engines like lc0 and Stockfish highlights a unique approach, suggesting that traditional methods may not always be optimal in specialized AI applications.
Concerns
- The article's complexity may alienate readers, indicating a potential disconnect between the author's expertise and the audience's understanding of chess AI.
- Concerns about the choice of SPSA over more popular optimization algorithms suggest that chess engine development might lag behind advancements in AI research.
