Reliability and controllability can emerge at scale too
Ilya SutskeverIlya Sutskever (OpenAI Chief Scientist) — Why next-token prediction could surpass human intelligenceat 34:00
From the conversation
How many people have the ability to SSH into the machine with the weights? The security people have done a really good job so I'm really not worried about the weights being leaked. What kinds of emergent properties are you expecting from these models at this scale? Is there something that just comes about de novo? I'm sure really new surprising properties will come up, I would not be surprised. The thing which I'm really excited about, the things which I’d like to see is — reliability and controllability. I think that this will be a very, very important class of emergent properties. If you have reliability and controllability that helps you solve a lot of problems. Reliability means you can trust the model's output, controllability means you can control it. And we'll see but it will be very cool if those emergent properties did exist. Is there some way you can predict that in advance? What will happen in this parameter count, what will happen in that parameter count? I think it's possible to make some predictions about specific capabilities though it's definitely not simple and you can’t do it in a super fine-grained way, at least today.…
Summary
Not all emergence is dangerous — Sutskever argues that reliability and controllability are themselves emergent properties we want to appear at scale, and that understanding the conditions under which beneficial properties emerge is a core research challenge.
Watch the clip on YouTubeStarts at 34:00