Untrusted smart models and trusted dumb models — LessWrong