Corrigibility thoughts III: manipulating versus deceiving — LessWrong