Grok

Source: https://grok.com/c/08965b27-3abe-4f16-ae2f-9e2e130061ae?rid=5db292c5-2930-4fc1-8293-4d1783059c65

The recent news is not that ChatGPT-the-chatbot sat down and solved a famous equation in a conversation. It is that OpenAI used a much stronger internal model, plus a swarm of about 10,000 cooperating AI agents, and claims they produced a mathematical proof of one of the Clay Millennium Prize problems: the Navier–Stokes existence and smoothness problem.⁠OpenAI
The Clay Institute later said the problem has “apparently been settled,” but it has not handed over the $1 million prize. Review is deliberately slow. OpenAI says it does not intend to claim the money.⁠Claymath
Navier–Stokes in plain English
The Navier–Stokes equations are the standard math for how fluids move: air, water, blood, smoke, ocean currents, airflow over a wing.
They track three things at once:

how fast the fluid is going at each point
the pressure pushing it around
viscosity (the “stickiness” that tries to smooth things out)

Engineers use them every day for planes, weather, engines, and pipes. In practice they almost always use approximations, because the full 3D equations are brutally hard.
The Millennium question is not “can we simulate weather?” It is a much sharper math question:
If you start with a perfectly nice, smooth 3D fluid, do the equations always stay nice forever? Or can the math itself blow up — meaning some speed or spin becomes infinite in a finite time?
That blow-up is called a singularity. Think of a whirlpool that keeps tightening and spinning faster until the equations say “infinity,” even though viscosity is supposed to damp things down.
A real tornado is a useful mental picture — a spinning column that gets thinner and more intense — but the math question is stricter. Real air never actually reaches infinite speed. The question is whether this particular set of equations can predict that they do.
15.4: Tornadoes - Geosciences LibreTexts
Why it matters: if the equations can blow up, then at those points they stop being a complete description of real fluids. That is why this sat on the Clay list with Poincaré, P vs NP, and the Riemann hypothesis.
What OpenAI actually claims to have proved
Clay’s official write-up (by Charles Fefferman) does not require one giant all-or-nothing answer. It offers four acceptable targets. Prove any one and you have resolved the prize problem as written:⁠Wikipedia

OpenAI claims C and D: they constructed a fluid that starts at rest, is pushed only by a smooth force, keeps finite energy, and still develops a finite-time singularity.⁠OpenAI
The picture they describe is a vortex that spirals inward, stretches like spaghetti, gets thinner, and spins faster until the velocity goes to infinity — while the total energy stays finite.
Important nuance, because headlines skip it:

This is not the same as proving that an untouched fluid (no force at all) must blow up.
Options A and B — the “leave the water alone” versions — are still the versions many mathematicians treat as the heart of the problem.
A forced blow-up does match Clay options C and D, which Fefferman left open on purpose. So OpenAI is claiming a valid official target, not a random side problem. Whether Clay will treat that as prize-worthy after review is a separate question.⁠Gadgetsnow.indiatimes

How the AI effort worked
Rough timeline from OpenAI’s own write-up:⁠OpenAI

Late August 2026: they train an internal model they describe as stronger than GPT-6 Astra (the public-facing model at the time).
September 1: after hearing rumors that someone else was close, they point agents at Millennium problems.
About 88 hours later (September 5): the Navier–Stokes group reports a resolution.
Another 17 hours: GPT-6 Astra helps formalize the argument in Lean, a proof assistant that checks each logical step.
September 8: public announcement, a long write-up, and the Lean code.

Scale they published:

~10,000 concurrent agents on the Navier–Stokes track
~2.7 million messages and ~130 billion output tokens just for that problem
If a customer tried to buy the same run at list prices, OpenAI itself floated a figure around $15 million⁠Newscientist

So “ChatGPT solved Navier–Stokes” is the street version. The accurate version is: a next-gen OpenAI system, run as a giant multi-agent math lab with code execution and a cached internet, produced a candidate proof that was then machine-checked in Lean.
The controversy
A day earlier, NYU mathematician Tristan Buckmaster and Anthropic researcher Levent Alpöge posted major related results: finite-time blow-ups for several fluid equations that sit next door to Navier–Stokes, especially Euler (Navier–Stokes with viscosity turned off). They used AI tools too, including models from both companies.⁠Nature
Then the credit fight started: who inspired whom, whether de-identified usage of OpenAI products helped their models, whether OpenAI jumped the line after hearing a rumor. OpenAI says it did not look at their unpublished proof, cannot totally rule out that anonymized product usage improved its models, and will not claim the prize. Clay has not declared a winner. Independent mathematicians are still reading.⁠The Guardian
So the honest status as of mid-September 2026:

A serious, Lean-formalized claim exists for Clay options C and D.
The math community is excited and also uneasy about AI-scale “solve it with $10M of agents.”
The prize is not awarded yet.
The no-force versions (A/B) remain a different, still-open question unless someone later shows those too.

Are most of the world’s supercomputers used for Navier–Stokes / fluid dynamics?
Short answer: fluids are one of the classic giant uses of supercomputers, especially weather and climate. They are not “most of everything.” AI training is eating a huge and growing share of the newest machines.
What “fluids on a supercomputer” actually means:
Weather and climate models are fluid dynamics. The atmosphere and ocean are treated as fluids. Those codes solve Navier–Stokes-like equations plus heat, moisture, radiation, and chemistry, on a grid covering the planet. That is why national weather services and climate labs own some of the biggest machines. Market and usage breakdowns often put weather/climate in the ballpark of ~18–22% of supercomputing demand.⁠Dataintelo
Other fluid-heavy jobs:

aircraft and rocket aerodynamics (CFD)
turbine and engine design
nuclear weapons stewardship (shock waves and high-speed hydrodynamics)
oil/gas reservoir flow and carbon storage
fusion plasma (a cousin of fluid equations)

Government and defense is often listed as the single largest consumer of HPC (~30% in some industry splits), and a lot of that is simulation, including fluids. Academic research is another big slice.⁠Xtendedview
What has changed: the newest TOP500-class systems are increasingly GPU farms built for AI. The same boxes still run climate and CFD, but a large fraction of new capacity is training and inference. So if you walked the world’s machine rooms in 2026 you would see:

weather / climate / Earth-system models
national-security physics (including fluids)
engineering CFD
materials, chemistry, biology
a rapidly expanding AI-training pile

Fluids are not a niche. They are one of the reasons supercomputers exist. They are not the only reason anymore.
A useful distinction: solving Navier–Stokes on a computer (what weather codes and CFD codes do, approximately, on a grid) is a different activity from proving theorems about Navier–Stokes (what OpenAI’s agents did). Supercomputers have been doing the first for decades. The news this month is about the second.

---

🔐 Cryptographic Verification

Archived URL: https://grok.com/c/08965b27-3abe-4f16-ae2f-9e2e130061ae?rid=5db292c5-2930-4fc1-8293-4d1783059c65

�� CONTENT HASHES:
  SHA-256:  ff21d4eeda05c794cf1d25206a683e09f7581a05d58bf45b2ed576e322a883b5
  BLAKE2b:  280e489f674fba2006fa86e257d606425352e36b3c70f6e81a759b0070b66685
  MD5:      97908d62559e156f44d9c0bda46e78e7

�� TITLE HASHES:
  SHA-256:  dca61d32363b091bf130e0b539eaa6557a3a035be17a1be1e3dc2c183eafcd2f
  BLAKE2b:  93b372066f2f6799a21deac75705fc3bdaef053cebdd78253ae7e020e9de9b91
  MD5:      6626db256698f843db48e4e46ad4ea64

�� INTEGRITY HASHES:
  SHA-256:  b14834c341c7751c26cc9d9523e65c723e7d242f1c50951df457541e397a48b3
  BLAKE2b:  ba94b8a2dca9d98d90cd027e9049c1433729ce939e5772cab09883036663edc7
  MD5:      bb903b070e46ed45bbd33f6881af66ba

Archived with ArcHive - Client-side cryptographic archival system