AI may have just solved a million dollar math problem...

0:00 Watch how this moves. The air rolls into a ring. It stretches and breaks into smaller curls. There's more detail everywhere you look. It almost feels random, but it's not. We can predict it.

0:14 As scientists study things like this, we get massive advancements in the comforts of our lives. Airplanes don't exist without this, just for starters. But there's this open math question about

0:26 predicting how fluid moves as it gets infinitely complex. Can we predict it forever? There's a million dollar prize specifically for this math problem. It's called the Navier Stokes problem and

0:41 scientists and mathematicians have been trying to crack it for years. And OpenAI just announced that they solved it. And that's not even the crazy part. There's more to this story. Not everybody is

0:55 happy about this. Specifically, the mathematician who has been working on this problem for years. So, in this video, I'm going to explain what Naviier Stokes even is and why it's so

1:09 important. From improving how airplanes fly to improving the efficiency of vehicles and chip cooling, predicting how weather moves, so many important things are based on this one math

1:23 formula. But the main thing I want to talk about is the fact that AI was used to solve it and why that's such a big deal. So the Navier Stokes equations are the main way that scientists try to

1:38 predict and describe how fluids move. Everything from water flowing through a pipe to air moving around airplane wings, ocean currents, smoke swirling through a room. And as of a few minutes

1:50 ago, Open AI just announced that their AI model solved this big open problem. In fact, it was such a big problem, it's part of a set known as the Millennium Prizes, and there is a milliondoll

2:04 bounty for solving it. So, here's the post from OpenAI. We're sharing a solution to the Navier Stokes Millennium Prize problem. One of the deepest problems at the frontier of mathematics.

2:14 The proof was produced by a group of agents using an open AAI next generation model significantly more capable than GPT6 Astra. You know, the model that was just released 4 days ago. So, there's

2:30 two main things I want to break down here. Number one, this insanely difficult math problem was solved using AI agents. We are now in a world in which AI is solving real insanely

2:44 difficult math problems and not just ones that are theoretical or don't have any practical application. This Navier Stokes problem is a very big deal with very real world consequences. This is

2:59 actually the part of artificial intelligence that I am most excited about. the promise of having new material science, new medicine, maybe curing all disease. This all kind of

3:12 stems from what we're seeing right in this moment. And also, I kind of like that they just casually said that they have a model that is so much better than GPT Astra, which literally came out 4

3:24 days ago and is blowing my mind along with everybody who has access to it. Now, the Navier Stokes problem has remained unsolved for 90 years. There's a million-dollar prize associated with

3:38 it, and a group of AI agents from a next generation OpenAI model just solved it. But there is some drama around this announcement. Okay, so here is what happened. We have two mathematicians,

3:51 Tristan Buckmaster and Levvent Alpaji. Sorry if I'm mispronouncing your name. and they spent a year working on this hard fluid math problem Navier Stokes. They used OpenAI's codeex tool and yes

4:06 they were also using AI to help them solve this problem and they stored all of their drafts to this math problem in codecs and then in mid August they got a real result. They were able to prove

4:20 something that was previously unproven. Then on September 3rd, the rumors started. They heard that OpenAI now knew about their work. So Buckmaster emailed OpenAI to say that this was a personal

4:35 project, not an official anthropic project. And why am I mentioning Enthropic? Well, one of those two mathematicians is also an Enthropic employee. But again, this was their

4:48 personal project to try to solve this math problem. It has nothing to do or at least they claim with Anthropic. Then on September 6th, OpenAI called them. They said that their own AI had just produced

5:02 a big proof on the next harder version of the same problem. But here's where it gets a little weird. Their AI used this very unusual approach to solving this problem, which happens to be the exact

5:15 same approach that these two mathematicians took. And Buckmaster says the first prompt that OpenAI put out towards trying to solve this problem only went out a day after they learned

5:28 about their actual progress. So think about this. OpenAI learns about the progress, decides to prompt their own model to try to solve it in the same direction that these two mathematicians

5:40 have been working on. And because it's AI, it completed it. It proved it in a very short amount of time with that little bump in the right direction. Now, this is all allegedly, by the way. I

5:52 don't actually know what happened. We're just reading from public tweets and announcements from OpenAI and these mathematicians. Now, Buckmaster explicitly asked OpenAI if their private

6:04 codeex drafts had been used, and he did not get a clear answer. OpenAI then negotiated on who gets credit for the solution, who gets to publish the solution, and one of the deals would

6:16 have left one of those mathematicians off of the credit simply because he worked at Anthropic. And then on September 8th, Buckmaster published a pretty damning post accusing OpenAI of

6:30 rushing to get the solution published before going through all of the tedious checks and writeups and explanations that he would have done simply to control who gets the credit. But OpenAI

6:44 says that's false. Here's OpenAI's post. We congratulate Levent, Alpagy, and Tristan Buckmaster on the remarkable mathematical work. We the researchers and the agents did not see any of their

6:56 work through any means until they released it publicly. In particular, no specific user data was accessed in order to solve this problem. So, they are categorically denying it. However,

7:08 there's even nuance to that. While unlikely, we cannot rule out that deidentified data derived from their usage of our products helped improve our models. And this has massive

7:22 implications. If you're doing anything on top of OpenAI's models, on top of Anthropics models, you just have to assume they have your data and they can come compete with you in the future. But

7:35 I'm going to get to that in a little bit. And this leads me to recursive self-improvement. If you haven't heard this term before, it basically means artificial intelligence that is

7:45 discovering ways to improve itself. And in fact, both Anthropic and Open AAI have been using the beginnings of RSI, recursive self-improvement, to improve their own models. They've talked about

7:59 how GPT 5.5 helped build GPT 5.6, just as an example. And just a few days ago, they put out an entire blog post about how they're using AI to accelerate their own research into AI. Now, it's not a

8:15 fully closed loop quite yet. And a fully closed loop means that AI literally comes up with the experiments, runs the experiments, finds the solutions that work well, and then literally patch

8:28 itself or improve itself and then continues that indefinitely. But they do talk about how AI is accelerating basically every aspect of their research. And then less than an hour

8:38 ago, Sebastian Bubck, who has been leading the Naviar Stokes initiative at OpenAI, posted receipts of his conversation with the mathematicians. So listen to this. He said he reached out

8:50 to Levent to coordinate the releases because they had come to this solution at about the same time. Then he never asked for Levent to be removed from authorship of his own work. As indicated

9:02 by my text, OpenAI's intention was to do everything possible to celebrate their mathematical achievements. And then last, on a personal level, it was incredibly difficult to have these

9:12 conversations. Levent refused to attend any of the meetings despite my repeated asking. So, it's very interesting how all of this went down. And the crazy part is that OpenAI's next generation

9:24 model solved the problem in less than 5 days. This open problem that has been open to humans for 80 plus years was solved by AI in 88 hours. They began their work on September 1st, less than a

9:41 week ago, and finished it on Saturday, September 5th. In total, 4.9 million messages between the agents and 300 billion output tokens. And so now when we think about this Navier Stokes

9:56 problem, think about AI being able to discover new math that it can use to improve itself. And I really think we're on the cusp of seeing that because if AI can solve a Millennium Prize problem, it

10:10 can solve any math problem. And this is both scary and awesome. It's scary because hopefully we're going to be able to control AI once it's improving itself. Because once that happens, its

10:23 improvement is exponential. Now, it's awesome because there's this incredible future in which all disease is cured and we have incredible material science discoveries which helps us with

10:36 environmental impact and helps us predict weather and get more efficient flights. Maybe instead of a flight from LA to New York taking 6 hours, it only takes 1 hour. And there's just so many

10:49 incredible things that can come from this. But here's another angle. What does it mean for humans? If AI is better and faster at discovering new knowledge, where does that leave us? Is all credit

11:02 going to go to AI for future innovations and discoveries? Is there even a place for humans anymore? I mean, think about it like this. AI has more or less solved chess. The best AI engine is

11:15 significantly better than the best human at chess. Now apply that to all knowledge work. The only reason we still have chess competitions is because people like seeing other humans compete

11:28 in chess. It's not fun to watch AI absolutely demolish Magnus Carlson, but it is fun to watch Magnus Carlson demolish every other human chess player. And so with chess, we like to see humans

11:42 play each other. But with math and other frontier discoveries, there is no delight in watching humans do not as well as AI could do. Really, what we want is the best knowledge discovery as

11:58 quickly as possible. And that is going to be the realm of AI. Now, that brings me to the final point. If you are building your business on top of open AI or anthropics models, you should be a

12:11 little bit nervous. There is immense platform risk, something that I've talked about at length on this channel. And Open AI more or less has admitted that is a thing. So, think about it this

12:24 way. You have all of your expertise about your business and now you're using Open AI's models to help power your business. Now, in that process, you are giving Open AAI all of your data. Listen

12:37 to it right here. We cannot rule out that deidentified data derived from their usage of our products helped improve our models. So, basically, as you're using Chat GPT, the future

12:48 versions of Chat GPT will already know about your business. It will already have your expertise built into itself. This is a huge reason to think about open-source models to really examine how

13:02 these frontier AI labs are using your business data to train their future models. So there is so much to this story which is just fascinating from AI solving math problems that were

13:14 previously unsolved for decades. We have the drama around sharing your data with these model companies and what that actually means. And then we have recursive self-improvement, which I've

13:28 covered a ton on this channel.