Why responsible AI development needs cooperation on safety