OpenAI’s Unreleased Model Solved Math: Is Biology Next?

SPEAKER_1 (0:00)

What's up guys, I just got to Japan. I am the most jet-lagged I have ever felt in my life. And I'm sat on my hotel room floor because I need to record this video. OpenAI just dropped one of their biggest releases, and it's not an AI model. It's the announcement that they solved around 90 of the toughest problems in mathematics ever. And I know what you're thinking. You're wondering why on earth should I or you care about maths? That's something I haven't touched since high school or college. And the point of this video isn't that they solve these math problems. This will slowly and very quickly roll over into every particular field that you could think of that you probably studied to do your job today. It'll roll up into chemistry, into engineering, into physics, into whatever requires reasoning. And it's all happening much quicker than anyone can even imagine. So in this episode, I'm going to get into what exactly it's solved, but also, more importantly, what happens next and what this means for you.

All of that and more on today's episode of In the Loop, brought to you by Qualcomm.

Let me just set some quick context. Three years ago, AI could not do math at all. Two plus two didn't equal four. It equaled 11 Strawberry had five Rs in it. LLMs just weren't smart enough. And I remember commentary from Terence Tao, who is one of the leading mathematicians in the entire world. And he said, this is a reasonable attempt by AI, but I do not see how reasoning can eventually lead to solving math. Reasoning is basically how you or I or any other human reasons about a problem. We don't just take it for face value. We apply different context to it. Back then, AIs could not do that, but they were at the start of their journey of figuring out this whole thing. Now, fast forward to six weeks ago, and OpenAI, who are back then and still are the experts in reasoning for training AI models, took that to its fullest extent when they announced that they had an internal model that was only 11 days old had solved one of the Millennium Prize problems. Now, Millennium Prize problems, there's seven of them, and they're all in the field of mathematics. If anyone can go out there and solve any of these problems, you get $1 million.

Now, for 200 years, these problems have existed. People have made progress, but the truth is, there are just too few mathematicians and experts in the world to try and figure this out. We are limited by human cognition and capacity. OpenAI came along with zero experience in maths, with zero help from math researchers, and they just decided, hmm, I wonder if I just use this powerful model that's more powerful than GPT-6 Astra, and just let it go until it was able to solve the problems. There was infinite compute, infinite money, just get it done. And it solved it. In 88 hours, 10,000 agents, it was one of the most costly experiments that they've done, probably that anyone's ever done, for a research problem. It cost over 10 million dollars. And there was a few interesting things about that. One, this was the first instance that an AI model had demonstrated that it wasn't just mimicking human intelligence. Up until then, people had been thinking, okay, AI models are smart, but they're just mimicking what it's learned on its dataset. It's a next token predictor. What it means is it's trained on all human knowledge that we put out there. It's limited by the cap or the ceiling of how intelligent we already are. It's solving Navier-Stokes was the first demonstration that it could go above that level, that it knew things that we as humans couldn't understand. And when Navier-Stokes got resolved, people weren't as excited as you might expect them to be, particularly mathematicians. They looked at this result and they weren't happy. They were perplexed. They were perplexed because they couldn't make sense of the solution. Now, fast forward to today, and we have news that their new internal model has solved not just one, not two, but 90 out of 500 of the world's most hardest open math problems. That's 20% of all the hardest math problems that have existed for decades at this point. But there's a few differences between this and Navier-Stokes. It solved it in a much quicker way. On average, each problem was solved in three hours, and it also got it done at a much cheaper cost, costing only tens of thousands of dollars. So you can imagine or extrapolate in the future that these models will get faster or quicker and even cheaper to the point where the cost of solving problems that humans have spent decades trying to figure out becomes undeniable. You should just use AI and you don't need humans anymore. But I was critical. I saw this announcement from OpenAI here, and I saw that they had dumped a massive GitHub, which was interesting. They open-sourced the entire thing so all the formal verification and proofs were laid out for anyone and everyone to critique. And it got me thinking, okay, it's solved a bunch of problems, but how is this useful for you or I? Like, how should I think about this being applied to the real world? And our friend Ole actually had a great breakdown in his tweet. He ran through a few of them, but being very high level here because I'm not a mathematician. One of these problems is called Laszloff-Maxwell, and if it's truly solved, this research can then result in a way to create virtually limitless clean energy to power civilization. Another example is the inverse elasticity problem, that if it's solved, can scan and map tissue stiffness, basically allowing you to evaluate and medicate and help individuals in the medical field. There's also another one called matrix multiplication. If you've heard that term before, it's because we've probably talked about it on this show. It's the foundation of how AI models run and how they run and compute all the important data to give you an amazing result. It's also how GPUs and any other modern graphics card processor works for videos, gaming, or whatever that might be. It's a very important fundamental mechanism of computing.