Like every week I’d have these calls with Ilya Sutskever at OpenAI and I’d tell him about my progress on watermarking, and he would say, “Well, that’s great, Scott, and you should keep working on that. But what we really want to know is how do you formalize what it means for the AI to love humanity? And what’s the complexity theoretic definition of goodness?” And I’m like, “Yeah Ilya, I’m going to keep thinking about that. Those are really tough questions, but I don’t have a lot of progress to report there.”
That was suprising to me. Sounds like OpenAI care about alignment enough to headhunt Scott and have the CTO refocus on it weekly
That was suprising to me. Sounds like OpenAI care about alignment enough to headhunt Scott and have the CTO refocus on it weekly