How to think about slowing AI — LessWrong