Rescuing the utility function — LessWrong