Untrusted smart models and trusted dumb models — AI Alignment Forum