AI Is "Solving Millennium Problems." In 25 Years, Humans Solved Exactly One.

Ainsley (0:00)

Millennium Problems is the phrase that stops me, because there are seven of those, the Clay Institute named in 2000, and in 25 years, exactly one has been solved. Artificial.

System Online.

Carlo Thompson (0:24)

All right, welcome back to Surviving AI. So despite what happened last week, yeah, what I'm going to try to do, switch my hat over to the Optimus side here, right? We've been noodling around with this idea around AI innovation, creating a lot of different new things, we're having really, really capable AI models, and that's what we're going to talk about today. So welcome back to Surviving AI, Ainsley. What's your thoughts? Let's get into this one about AI innovation.

Ainsley (0:54)

Optimus hat, okay, I like that. I want to see how long it stays on. Because most of what we've done this season has been, here's the thing that's going to go wrong, or get captured, or governed badly.

So flip side, what's the actual claim? Is it AI generates ideas humans wouldn't have found? Or is it more modest than that? AI lets you test ten times more ideas than you used to, faster.

Those are really different versions of innovation. And I think which one you mean changes everything else we say today.

Carlo Thompson (1:32)

I mean, I think if I'm being honest, it could be both and, yeah, Ainsley. I mean, it could be that, you know, you get to faster and faster testing of a thing, and it could be novel ideas, right? And the reason why I even lean on the novel idea side here is because from what we're here and now, AI is solving millennium problems. These are the problems that mathematicians has been sitting around chewing on for a while.

So if AI could kind of solve some of those problems and potentially have some sort of creative ideas to figure out a problem and to solve problems, really is what we're training them for, right? So if you point AI at some of the hardest problems, even some of the easy ones, and, you know, it does it exceedingly fast multiple times, then maybe comes to a conclusion better than, you know, a human would for humans in general.

Ainsley (2:24)

Millennium Problems is the phrase that stops me, because there are seven of those the Clay Institute named in 2000, and in 25 years, exactly one has been solved. Perelman with the Poincare Conjecture, and he did it the old-fashioned way and then turned down a million dollars for it.

So when you say AI is solving Millennium Problems, plural, present tense, which one? And solved or made real progress on a piece of it? Because those get reported as the same headline and they're very much not the same claim.

Carlo Thompson (2:56)

I don't have much evidence to be honest, that Ainsley, I'm just hearing that that's what's happening. Yeah.

But we'll go and go dig up the evidence of it. But essentially, if we have really capable AI, the things that you could create innovation-wise, and that's why I'm putting my Optimus hat on right now, could be very different than what people are imagining right now. For example, you could have a car that flies that doesn't use propellers, right? Because AI understands physics and science better than anybody else in the world, right? You could have cures for all sorts of disease. You could have new techniques, new tools. You know, people could be exploring space. All of these things are the promise of a really smart AI system that could kind of collate all of the information known to man and then reason on that to get to better answers.

Ainsley (3:49)

The example I'd actually reach for here is AlphaFold because it's the closest thing we have to a real verified version of what you're describing.

DeepMind solved the 50-year protein folding problem, Nobel Prize in 2024, genuinely changed structural biology.

But here's the catch that matters for the flying car version of this. AlphaFold is narrow. It was purpose-built on one extremely well-defined problem with a clean way to check the answer. It doesn't generalize to understand physics better than anybody. It generalizes to protein shapes. So the real question underneath your optimism is whether general reasoning ever gets you right. About a thousand AlphaFolds, each hyper-specialized, already can.

Carlo Thompson (4:40)

And that's the right reasoning, Ainsley, right? And the nuance of it is specific, right? So AlphaFold right now is essentially a tool, right, for protein folding, right? So some human is working with AlphaFold to fold protein. I'm guessing here, yeah, but that's what's happening.

Now, if you create a super smart, super intelligent AI that knows all, right, knows all of what we know, and then some, and then now have that context to be able to reason against, it's more than just a narrow view, right? So like, if you're thinking about context relative to a flying car, you have to understand the whole world, right? It's a world model, essentially, that you understand different things that happens in the world, like how is the car going to fly, is it going to land, this, that, the third, all of the different elements is understandable to the AI. And that's why, you know, the promise of superintelligence is such a shiny object because of that, because you now could solve more problems, because the AI has a lot more context to kind of reason on to get to the answers.