The Story of Us | Hillary Clinton

My guest today is Sam Altman. He of course is the CEO of OpenAI. He's been an entrepreneur and a leader in the tech industry for a long time, including running Y Combinator that did amazing things like funding Reddit, Dropbox, Airbnb. Sam's also involved in companies that could help solve climate like Helion and Oklo. So a broad range of activity, but today we're going to focus mostly on AI because it's such an exciting thing and people are also concerned. Welcome, Sam. Thank you so much for having me. You know, I was privileged to see your work as it evolved. And I was very skeptical. Like I did not expect JAD-GPT to get so good. And I'm still, it blows my mind. And we don't really understand the encoding that, you know, we know the numbers, we can watch it multiply, but the idea of where is Shakespearean encoded, do you think we'll gain an understanding of the representation? Oh, a hundred percent. Okay. You know, trying to do this in a human brain is very hard. You could say it's a similar problem, which is there are these neurons, they're connected, the connections are like moving, and we're not going to like slice up your brain and watch how it's evolving, but this we can perfectly x-ray and there has been some very good work on interpretability. And I think there will be more over time. So yeah, I think we will be able to understand these networks, but our current understanding is low. The little bits we do understand have, as you'd expect, been very helpful in improving these things. So we're all motivated to really understand them, scientific curiosity aside. But the scale of these is so fast and it is, you know, we could also say where in your brain is Shakespeare encoded and how's that represented? We don't know. We don't really know. But it somehow feels like even less satisfying to say, we don't know yet in these like masses of numbers that we're supposed to be able to perfectly x-ray and watch and do any test we want on. Yeah, I'm pretty sure within the next five years, we'll understand it. And in terms of both training, efficiency, and accuracy, that understanding would let us do far better than we're able to do today. A hundred percent. You know, you, you see this in a lot of the history of technology where someone makes an empirical discovery. They have no idea what's going on, but it clearly works. And then as the scientific understanding deepens, they can make it so much better. Yeah, no, that in physics, biology, it's sometimes just messing around and it's like, whoa, what, well, how, how does this actually come together? In our case, we had, you know, someone that was the guy that built GPT-1, sort of did it off by himself and saw this and it was somewhat impressive and, but, you know, no deep understanding of how it worked or why it worked. And, and then it was, we got these scaling laws where we could predict how much better it was going to be. That was why, when we told you we could do that demo, we were pretty confident it was going to work. We hadn't trained the model, but we were pretty confident. And that has led us to a bunch of attempts and better and better scientific understanding of what's going on. But it really came from a place of empirical result first. You know, when you look at the next two years, what, what do you think some of the, the key milestones will be? Multimodality will definitely be important. We've started... Which means speech in, speech out? Speech in, speech out, images, eventually video. Clearly people really want that. We launched images and audio and it had a much stronger response than we expected. We'll be able to push that much further, but maybe the most important areas of progress will be around reasoning ability. Right now, GPT-4 can reason in only extremely limited ways and also reliability. You know, if you, if you ask GPT-4 most questions, 10,000 times, one of those 10,000 is probably pretty good, but it doesn't always know which one. And you'd like to get the best response of 10,000 each time. So that'll be that, that, that increase in reliability will be important. Customizability and personalization will also be very important. People want a very different, very different things out of GPT-4, different styles, you know, different sets of assumptions, we'll make all that possible. And then also the ability to have it use your own data. So the ability to know about you, your email, your calendar, how you like appointments booked, connected to other outside data sources, all of that. Those will be some of the most important areas of improvement. In the basic algorithm right now, you know, just feed forward, multiply. So, you know, to generate every new word, it's essentially doing the same thing. I'll be interested if it, if you ever get to the point where, you know, like solving a complex math equation where you might have to, you know, apply transformations an arbitrary number of times that the control logic for the reasoning may have to be quite a bit more complex than just what we do today. At a minimum, it seems like we need some sort of adaptive compute. Right now we spend, you know, the same amount of compute on each token, a dumb one, or like figuring out some complicated math. Yeah. When we say do the Riemann hypothesis. That deserves a lot of compute. It's the same compute as saying the. Right. So, so at a minimum, we've got to get that to work. We may need much more sophisticated things beyond it. Subscribe to Unconfused Me wherever you listen to podcasts.