How a Simple Puzzle Reveals the Jagged Edge of AI’s Abilities
An interactive investigation that trains a GPT on one million generated 15-puzzle solutions, then compares performance on typical, easy, and already-solved boards; the author reports 99.9% success on randomly scrambled test boards and failures on easy edge cases.
Published: · Author’s article archive shows Sept. 23, 2026.
Tools and source notes
Tools named in the source: GPT, 15-puzzle.
Tool names are source attributions. An exact model or complete production workflow may not be disclosed.
Mobile compatibility has not been confirmed.