Supervising strong learners by amplifying weak experts — AI Alignment Forum