Teaching an AI not to cheat? — LessWrong