A summary of aligning narrowly superhuman models — LessWrong