In defense of probably wrong mechanistic models — AI Alignment Forum