Can an LLM solve this combinatorics problem?
Summary
The blog post investigates using an LLM to solve a combinatorics puzzle about identifying a permutation with binary queries on cards. It compares human-derived and AI-assisted strategies across 4-, 8-, and 16-card variants, discusses information-theoretic lower bounds, and presents an open question about the exact worst-case minimum D(16). The author reflects on the challenges of guiding AI research, the effectiveness of a meta-harness, and the gap between provable guarantees and empirical performance.