AI-powered
podcast player
Listen to all your favourite podcasts with AI-powered features
How Do You See AI Alignment?
I think of alignment mostly as the problem of getting the model to try to do the right thing. I'm doing some work now on, uh, training AI systems to assist human feedback. When you have models that are kind of very below human level or maybe at human level at a certain task, it's pretty straightforward to supervise them. But once they're doing things that are very hard and require a lot of diverse technical knowledge, it becomes difficult for us to provide useful supervision signals.