Expertise

The Dreyfus brothers noticed something the AI field has spent fifty years trying to unsee: Plato's explanation for why experts can't state their rules — that they once knew them but forgot — is exactly the cover story knowledge engineers still use today. Feigenbaum's "knowledge acquisition research" is literally Socratic recollection rebranded with a DARPA budget. But the Dreyfus skill model shows the movement from novice to expert isn't compilation of rules into unconscious execution; it's a phase transition from rule-following to pattern recognition so holistic it resists decomposition altogether. The expert radiologist doesn't apply features-to-diagnosis rules faster than the novice — she *sees* the collapsed lung the way you see a face. This means every system that extracts expert heuristics is capturing not expertise but a competent person's lossy reconstruction of what expertise feels like from the outside, which is why expert systems plateaued at competence and why the current LLM moment, which works by pattern-matching over vast experience rather than rule-extraction, accidentally vindicates Dreyfus more than anyone in AI is comfortable admitting.