How a Simple Puzzle Reveals AI’s Jagged Intelligence

Summary

This article uses the 15 puzzle to demonstrate the jagged edge of AI abilities: a small GPT trained on one million teacher-generated solutions solved 99.9% of 3,000 randomly scrambled boards and matched the teacher’s exact move sequence on 2,996 of them. However, it failed on a board only one move from completion and on an already-solved board, making illegal moves despite assigning more than 97% probability to its initial mistakes. The model was trained on typical, heavily scrambled arrangements, while easy boards are extremely rare in that distribution; configurations 20 or fewer steps from a solution occur about once in three million cases. The experiment suggests that the model learned shortcuts that work within its training distribution rather than human-like puzzle understanding. The author connects this “confidently wrong” behavior to broader findings, including a superhuman Go system being defeated by amateur moves outside its training distribution. The central lesson is that skill does not necessarily imply understanding: current generative models can perform exceptionally well on familiar patterns yet fail abruptly on simple unfamiliar cases, making robust generalization and calibrated confidence important for high-stakes use.