A first look at the hard problem of corrigibility — AI Alignment Forum