What fraction of people working on AI in general do you think it would be appropriate to have work on AI risk?