Global AI governance: Realistic difficulties, mode selection