A first look at the hard problem of corrigibility — LessWrong