Failure Modes of Teaching AI Safety — LessWrong