AI alignment open problem — LessWrong