7+ tractable directions in AI control — LessWrong