Infernal Corrigibility, Fiendishly Difficult — AI Alignment Forum