From New York Times' opinion, this is the Ezra Klein Show.

So I think you can date this era in artificial intelligence back to the launch of ChatGPT.

And what is weird if you talk to artificial intelligence people about that is they'll tell you ChatGPT, it was just a wrapper, an interface system.

The underlying system, GPD3 had been around for a while. I mean, I'd had access to GPD3 for quite a while before ChatGPT came around.

What ChatGPT did was it allowed you to talk to GPD3 like you were a human and it was a human.

So it made AI more human. It made it more able to communicate back and forth with us by doing a better job mimicking us and understanding us.

Which is amazing. I don't mean to take anything away from it, but it created this huge land rush for AI's that functionally mimic human beings.

AI's that relate as if they are human beings and try to fool us into thinking that they're human.

But I've always been more interested in more inhuman AI systems. When you ask somebody who's working on artificial intelligence, including people who believe it could do terrible harm to the world, why are you doing it? What's a point of this?

They don't say, oh, we should risk these terrible consequences because it's fun to chat with ChatGPT.

They say, oh, AI, it's going to solve all these terrible scientific problems we have, clean energy and drug discovery.

And it's going to create an era of innovation like nothing humanity's ever experienced.

There aren't that many examples though of AI doing that yet.

But there is one which you may have heard me mention before. And that's alpha fold. The system built by DeepMind that solve the protein folding problem.

And the protein folding problem is that there are hundreds of millions of proteins. The way they function has to do with their 3D structure.

But even though it's fairly straightforward to figure out their amino acid sequence, it's very hard to predict how they will be structured based on that.

We were never able to do it. We were doing it one by one, studying each one for years to try to figure out and basically map it.

And then they build the system alpha fold, which solves a problem, is able to predict the structure of hundreds of millions of proteins, a huge scientific advance.

So how did they build that? And what could other systems like that look like? What is his other path for AI, this more scientific path where you're tuning these systems to solve scientific problems, not to communicate with us, but to do what we truly cannot do.

Demis Asabis is a founder of DeepMind. DeepMind is owned by Google and recently Asabis was put in charge of all Google AI. So now it's called Google DeepMind and he runs all of it.

That makes him one of the most important people in the world, charting the future of artificial intelligence.

So I asked him to come on the show to talk me through the development of alpha fold, how it was built, what came before it, what could come after it.

...