The case for aligning narrowly superhuman models — LessWrong