In September, multiple AI CEOs signed a voluntary safety pact to self-pace development if models present undue risk