Humanity needs a Ulysses Pact for AI — LessWrong