Debunking the myth of safe AI — LessWrong