The self-unalignment problem — AI Alignment Forum