Why I expect successful (narrow) alignment — LessWrong