The Search Gap: Teaching AI to Create Counterintuitive Chess Puzzles

We unpack a DeepMind study on training an AI to generate genuinely creative chess puzzles, using reinforcement learning and a three-part creativity framework: uniqueness, novelty, and counterintuitiveness via a 'search gap' between shallow and deep engine evaluations. Stockfish ensures a unique solution; Levenshtein distance enforces novelty in both the puzzle and the solution; and a diversity filter guards against reward hacking. Results showed a tenfold rise in counterintuitive puzzles (0.22% baseline to 2.5%), surpassing human puzzle rates in a large dataset. Expert reviewers found the AI puzzles more creative and enjoyable. We discuss broader implications: a general blueprint for nurturing AI creativity in domains like Go, mathematics, and even prompting large language models, turning shallow intuition into deeper, surprising insight.

Note:  This podcast was AI-generated, and sometimes AI can make mistakes.  Please double-check any critical information.

Sponsored by Embersilk LLC