Google DeepMind's AI Learns to Create Original Chess Puzzles, Reviewed by GMs (www.chess.com)

DeepMind released a 75‑page study showing an AI that generates original, often surprising chess puzzles and that attracted positive, nuanced praise from grandmasters. The system was trained on four million Lichess puzzles using large neural nets and then fine‑tuned with reinforcement learning. Its reward function explicitly favored puzzles that are unique (one clear winning move) and “counterintuitive” — i.e., positions that are solvable by strong engines but tend to fool weaker ones. A curated set was evaluated by GMs Matthew Sadler, Jonathan Levitt and FM Amatzia Avni, whose 39‑page review called several positions beautiful and original while flagging some trivial or unrealistic outputs. Technically, the work is significant because it moves beyond pattern‑mining to optimize for aesthetic and surprise qualities, using behavioral/engineic gaps as proxies for creativity. Concrete examples include a puzzle where the winning idea begins with sacrificing both rooks to enable a slow queen infiltration — a theme the experts labeled “unorthodox” and hard to spot. DeepMind frames this as a milestone toward human–AI collaboration in composition and design: not yet prize‑winning, but a proof of concept that RL can encode subjective notions like elegance, and a potential tool for composers, training, and studying computational models of creativity.