---
title: 'AI Solved a Millennium Problem and Everyone''s Angry'
source: 'https://youtube.com/watch?v=eR2q8s4VB1w'
video_id: 'eR2q8s4VB1w'
date: 2026-11-11
duration_sec: 399
channel: 'Sabine Hossenfelder'
---

# AI Solved a Millennium Problem and Everyone's Angry

> Source: [AI Solved a Millennium Problem and Everyone's Angry](https://youtube.com/watch?v=eR2q8s4VB1w)

## Summary

The video discusses OpenAI's claimed proof of the Navier-Stokes Millennium Problem, addressing the negative reaction from the mathematical community and evaluating the proof's plausibility and relevance.

### Key Points

- **Negative Reaction to OpenAI's Proof** [00:02] — The creator notes that mathematicians and physicists have reacted extremely negatively to OpenAI's proof, complaining that it wasn't left to humans, despite previously criticizing private business for lacking interest in pure maths.
- **The Navier-Stokes Problem Explained** [00:57] — The problem asks whether a fluid described by the Navier-Stokes equation can develop singularities (blow-ups) in finite time. The equation is key for realistic fluids and gases, used in weather forecasting.
- **OpenAI's Approach** [01:43] — OpenAI's proof starts with a fluid field that has a blow-up and inserts it into the equation, calling whatever's left the forcing. The proof is about 160 pages and uses pulses at chosen time intervals to drive the fluid into a singularity.
- **Potential Circular Arguments** [02:56] — The creator suspects that AI models may run into circular arguments, implicitly assuming what needs to be proved, similar to saying 'if pigs can't fly, then the laws of gravity are broken'—consistent but not proving the target.
- **Plausibility and Relevance** [04:08] — The proof is plausibly correct because it focuses on a class of blow-ups that has been the focus of past research. The creator also notes a relation to general relativity and black hole horizons, arguing that pure maths has intrinsic value.
- **Final Verdict** [06:04] — The creator estimates a 90% chance the proof is correct, but a 99% chance mathematicians will continue to complain about it.

### Conclusion

The video concludes that while the proof is likely correct, the mathematical community's negative reaction is predictable, and the debate over AI's role in pure maths is just beginning.

## Transcript

Navier-Stokes Millennium Problem. It's not a secret that I've been working on the same problem in the past year without my eyes models, and I'm sure so have many others. So, what are we to make of the solution? And is it really
as useless as the critics complain? Let me say how that I've been struck by how extremely negative the reaction has been to the OpenAI proof. Like, you'd think that mathematicians and physicists would be interested in how the proof works,
how OpenAI managed to get it done, what else can be done with it. Instead, everyone's complaining that OpenAI didn't leave it to humans. For decades, they've complained that private business has no interest in pure maths. Now,
they're complaining private business does have an interest in pure maths. Okay, but first, let me tell you what we're even talking about. The problem to be solved is the question of whether a fluid, that's described by the
Navier-Stokes equation, can develop singularities or blow-ups in finite time. The Navier-Stokes equation is the key equation that describes realistic fluids and gases, and it's, for example, used for the weather forecast. The
equation looks like this, where U is the fluid itself, and P is the pressure. This new is the viscosity, that's a constant, and this F here is what's called the forcing, that's the force that acts on the fluid in each point.
the fluid to blow up if you make the forcing large enough. So, if you want to solve the problem, you have to show that the forcing itself remains regular. The cleanest way to do this is to use zero forcing. That's also the hardest way.
The OpenAI proof now, instead, starts with a fluid field that has a blow-up. You just insert this into the equation and call whatever's left the forcing. remaining forcing itself remains regular. The entire proof is 160 pages
or so. I did not check it. I'll leave that to competent people, but I can make some incompetent observations. The construction works by supplying the fluid with pulses that come in at specifically chosen time intervals that
adapt to the fluid dynamics. So, it's a deliberately engineered driving of the fluid into a singularity. The simple singularity they looked at is a type of vortex that becomes increasingly stretched and focused on its center
axis. This is what the image shows. This is totally allowed by the criteria of the millennium problem, but it puts the burden on having to prove that the forcing does not supply unlimited energy. This means they have to somehow
keep track of what the pulses do and sum them all up and estimate the errors and so on. This is what most of the proof's about. It's lemma after lemma about how to estimate the total contribution from the forcing. Now, look going by what
OpenAI said, they had literally thousands of AI agents working on this. So, it seems highly unlikely to me that the proof is directly wrong. However, what I've noticed is much more likely to happen with the publicly available
OpenAI models is that they run into circular arguments where they implicitly assume what needs to be proved. They might, for example, assume a certain regularity or solvability condition for a lemma that actually contained the
bound they had to prove. The issue is that this sort of argument has no actual mistake in the step-by-step procedure. You see, it's like saying if pigs can't broken. And if the laws of gravity are broken, then pigs can fly. And that's
totally correct, but it doesn't prove that either pigs fly or the laws of gravity were broken, just that both are consistent with each other. So, it might be that the proof is consistent in that sense and yet circular in not proving
the thing they need to prove. I found that even GPT-5.6 still struggles with this. By the way, my impression is that the new Astra is worse than GPT-5.6 when it comes to maths, but that's just an aside. Let me put it like this.
Navier-Stokes equation does have blow-ups, and the class that they looked at in the proof is the case that most efforts in the past decade have focused on. So, it's plausibly correct. Still, I'll not take it for granted that the
proof's correct until humans have checked it. Then, let me say something about the question of whether this is relevant. I see a lot of people commenting, "Oh, no one cares. This isn't useful for anything." and so on.
Well, the reason I've been interested in this, besides the million-dollar prize, is that the Navier-Stokes equation has a known relation to some problems in general relativity, especially concerning the physics of black hole
horizons. Now, you could say again, "Okay, but this isn't useful for anything." at which point I'd like to ask, "Well, what's the use of kicking a ball into a net?" Personally, I think if some people find it interesting, that's
the use in itself. Can you make money with it? I mean, other than winning this prize? Probably not. And I guess that this will soon dampen Oop May I ask maths enthusiasm. But some Altmans are here now be looking at room temperature
superconductors, which I think will keep GPT occupied for a while. That problem's on a completely different level because it's not mathematically well-defined to begin with. Finally, let me add that I'm aware there's some controversy about
what I wrote May I ask model sneaked in on someone else's research. I have to admit that I'm super not interested in people stories. I really just care about science. You might say, "Okay, but if you care about science, you need to keep
people motivated to do it, so you must care about people. And if their results get stolen, they won't continue doing it." Well, I guess that's the question we have to answer. Do we need to keep people motivated by giving them a
million-dollar prize, or do we just leave it to AI to begin with? And will AI companies continue to be motivated if they notice there's no money to make with it. In summary, I think there's a 90% chance this proof is correct, but a
99% chance that mathematicians will continue to complain about it. If you haven't yet reached your daily limit of uninformed opinions about useless topics, please subscribe and buy a t-shirt.
t-shirt. Thanks for watching.
