An AI crash is our best bet for restricting AI — LessWrong