The Superalignment team led by OpenAI chief scientist Ilya Sutskever has devised a way to guide the behavior of AI models as they get ever smarter.