88 Hours, 10,000 Agents: OpenAI Says It Solved Navier-Stokes: and It’s Already Turning Sour

This initial thread does not yet concern the 88 hours, the 10,000 agents, or OpenAI’s internal model. It is dominated by another uncertainty: who actually solved the problem, and in what order?

@TimTeaFan mentions a competing rumor involving Anthropic and its Fable system:

"Weren't there rumours that Anthropic solved it with Fable?" @TimTeaFan

@mark_k’s response is inconclusive:

"There is some dispute or drama going on right now." @mark_k

In response to a similar question about the identity of the person or people who solved it, he adds:

"Currently it's a bit unclear who exactly solved it. Maybe both jointly." @mark_k

The discrepancy between the two threads lies at the heart of the story. Before OpenAI’s announcement, the debate centered on a rumor, Anthropic, priority, and the identity of the people or labs involved. Following the announcement, OpenAI turns the rumor into a quantified, technical institutional claim, without addressing the questions of authorship and provenance that @mark_k’s thread had already raised.

The announcement therefore establishes one specific point: at this stage, OpenAI is the only organization to make a detailed official claim. On its own, it does not settle the question of scientific priority.

Navier-Stokes, the problem behind the slogan

The Navier-Stokes equations are fundamental to fluid mechanics. They describe how a fluid, such as air, water, or a gas, evolves by relating several quantities:

• the fluid’s velocity, represented by a field that varies in space and time;
• pressure, which constrains and redistributes motion;
• viscosity, meaning the fluid’s internal friction;
• external forces, such as gravity or other applied forces;
• the incompressibility constraint, in the classical case of an incompressible fluid.

In simplified form, the momentum equation is written as follows:

[ \partial_t u + (u \cdot \nabla)u = -\nabla p + \nu \Delta u + f ]

Here, (u) represents velocity, (p) pressure, (\nu) viscosity, and (f) external forces. For an incompressible fluid, the condition (\nabla \cdot u = 0) is generally added.

These few symbols encapsulate extremely complex dynamics. The nonlinear term ((u \cdot \nabla)u) expresses the fact that the fluid transports its own velocity. This is one of the main sources of difficulty: motion at one scale can feed motion at other scales, distorting and amplifying it.

The real question in three dimensions: global regularity or singularity?

The Millennium Problem does not simply ask whether a flow can be simulated. It poses a question of global regularity in three dimensions.

Given sufficiently smooth initial data, can it be proved that the solution remains smooth for all time? Or can it develop a singularity in finite time, a phenomenon often called blow-up?

A singularity does not necessarily mean that a real fluid physically “explodes.” In mathematical analysis, it means that certain quantities associated with the solution cease to be controllable within the required regularity framework. The derivatives of the velocity may, for example, become infinite or lose the properties needed to extend a smooth solution.

The difficulty can be summarized as follows:

1. The initial data are regular.
2. Viscosity tends to damp small structures and dissipate energy.
3. But nonlinearity can concentrate activity at increasingly finer scales.
4. One must prove that dissipation always prevails, or construct a case in which it is insufficient.

It is this global control over the balance between nonlinear concentration and viscous dissipation that has remained elusive for nearly nine decades.

Why have such compact equations resisted solution for so long?

The equations are short. Turbulence is not.

In turbulent flow, large-scale structures transfer energy to smaller structures. Vortices interact, deform, fragment, and recombine. This cascade of scales makes analytical estimates particularly challenging.

Mathematicians already have fundamental results. In particular, they know how to construct weak solutions in general frameworks and obtain stronger results in certain dimensions, for certain regimes, or over limited periods. But the open problem concerns the universal guarantee of smooth regularity in three dimensions, given smooth initial data.

That is why a potential solution would be a major achievement. It would not simply amount to having “better computed” a simulation, but to resolving a fundamental alternative concerning the possible behavior of these equations.

Navier-Stokes Millennium Prize: what validation would require

An announcement, even an official one, is not enough to formally solve a Millennium Prize Problem. The proof would need to be made available in a sufficiently complete form to allow thorough scrutiny by independent experts.

This notably entails:

• publishing the full text of the proof and its assumptions;
• making it possible to verify every lemma, every estimate, and every logical dependency;
• critical review by specialists in partial differential equations and harmonic analysis;
• resolving any objections raised during this evaluation;
• allowing sufficient time to establish that the result is accepted by the mathematical community.

The Clay Mathematics Institute applies its own criteria for officially recognizing a solution to one of its Millennium Prize Problems. In particular, a proof must be published in a qualifying outlet, achieve general acceptance within the relevant community, and then withstand a period of scrutiny.

At this stage, OpenAI’s post provides neither a published proof, peer review, nor independent validation. It announces a claimed result whose mathematical recognition remains entirely to be established.

A Correct Proof Would Not Provide Perfect Weather Forecasts

One of the most intuitive, yet misleading, reactions is to directly link Navier-Stokes to exact weather prediction.

"Finally I’ll know for sure whether it’s going to rain tomorrow in London or not" @ChShersh

That conclusion does not follow from the announcement. Solving the theoretical regularity problem would not suddenly produce perfect weather forecasts.

Atmospheric forecasts depend notably on:

• incomplete and noisy initial measurements;
• coupled models that include humidity, clouds, radiation, and interactions between the ground and the atmosphere;
• physical parameterizations;
• necessarily discrete numerical computations;
• the sensitivity to initial conditions inherent in chaotic systems.

A proof of regularity, if established, would clarify the mathematical structure of the equations. That would be a major achievement. But it would eliminate neither measurement uncertainty, modeling errors, nor operational computing limitations.

Likewise, it would not instantly transform aeronautical, naval, or energy engineering. These fields already use Navier-Stokes under specific assumptions, with simulations, turbulence models, and experimental calibration. A theoretical advance may enrich the available tools in the long term, but it does not automatically become an industrial product.

Scientific Credit, the Main Post’s Blind Spot

The most substantive response under the announcement does not only challenge the role of AI. It also concerns the visibility given to the human researchers who may have originated key elements.

@Quasilocal states the grievance directly:

"Ok but like want to mention Tristan Buckmaster and Levent Alpöge in the actual tweet where you claim credit, rather than in a separe quote-tweet not linked to from this one?" @Quasilocal

The criticism is specific. It is not merely about whether they were supposedly mentioned elsewhere. It targets the omission of Tristan Buckmaster and Levent Alpöge from the message that publicly claims the result.

OpenAI did not publicly respond to this challenge in the thread. However, @calebcterry extends this call for recognition by placing the result within a longer mathematical lineage:

"After doing more research on this, I think the human contribution is worth mentioning, too. My understanding is that this came out of years of work by Diego Córdoba and Luis Martínez-Zoroa developing the underlying approach, with Tristan Buckmaster and Levent Alpöge then using LLMs to push that research much further and eventually formalize and verify the result." @calebcterry

This statement should be read for what it is: an interpretation reported by @calebcterry in the thread, not a definitive attribution established by an academic publication reproduced here. Nevertheless, it raises the right question: what should receive credit when a system of agents produces a proof based on preexisting approaches, intuitions, formulations, and research?

@X_is_Arbitrary draws a sharper conclusion:

"This seems ballsy. If what the mathematicians claim is true, then this just bolsters claims that AI is incapable of new ideas. It needed two brilliant mathematicians to tell it how to solve the problem." @X_is_Arbitrary

This interpretation is not supported by the publicly available evidence. Based on the thread, it is impossible to determine precisely how the invention should be attributed among Diego Córdoba, Luis Martínez-Zoroa, Buckmaster, Alpöge, the LLM tools used by these researchers, and OpenAI’s agents. But the continuity between prior human research and the result claimed by the agents is a central issue, not merely a matter of communication.

OpenAI’s main post does not credit any named human researcher, while the replies make this omission the primary subject of public dispute.

Codex sessions: a serious, unsubstantiated allegation

The controversy takes a more serious turn with questions about possible Codex sessions associated with the researchers.

@Mmorgan_ML asks:

"Did you or did you not train on Buckmaster & Alpöge's Codex sessions?" @Mmorgan_ML

@melqartvantyrus essentially asks the same question. @ArtemR goes further by amplifying the allegation that research may have been stolen:

"But did OpenAI cheat by stealing the research?" @ArtemR

It is essential to distinguish facts from suspicions. These messages show that the allegation is circulating. They do not prove that Codex sessions, notes, private data, or work by Buckmaster and Alpöge were improperly used by OpenAI.

The exchanges reproduced here contain no response from OpenAI, no technical details about the relevant training data, and no corroborating evidence. Nor is the content of the external article cited by @ArtemR reproduced.

At this stage, the claim that Codex sessions were improperly used remains an unconfirmed allegation. Nevertheless, the question of the provenance of the data and research deserves a factual, public, and verifiable answer.

This issue is distinct from, though related to, the criticism concerning the system’s closed nature. Even in the absence of any established wrongdoing involving data, a lab with an internal model more powerful than publicly available tools creates a significant asymmetry in research.

What this announcement says about the state of the art

The most novel aspect of the OpenAI Navier-Stokes announcement is not merely the word “solution.” It is the claimed setup: approximately 10,000 coordinated AI agents, running for 88 hours, centered on an internal OpenAI model described as more capable than GPT-6 Astra.

This setup suggests a potential shift in how mathematical research is organized. Rather than a standalone conversational assistant, OpenAI describes a collective system capable of distributing tasks among subproblems, verification, proof attempts, searches for counterexamples, and synthesis.

But the industrialization of discovery raises a governance question. @CircumjovialLLC states it bluntly:

"Not OK to use private models trained on humanity's corpus of mathematics to solve open problems when those models are far more capable than what you released and are not accessible to all researchers. This is tech feudalism, leading to intellectual tyranny." @CircumjovialLLC

So far, no response from OpenAI addresses the release of the model, its weights, its compute cost, its energy footprint, or possible academic access. @Adamlags explicitly asks about the cost of the experiment, while @alono88 humorously asks about the token cost. These questions remain unanswered.

The criticism of restricted access therefore does not prove misconduct. It raises a structural problem: how can the community reproduce, audit, or challenge a discovery produced using a model and infrastructure to which it has no access?

Key takeaways and what remains to be established

OpenAI’s announcement is potentially historic. The lab claims to have found a solution to the Navier-Stokes problem in 88 hours using approximately 10,000 coordinated AI agents, with an internal model described as more powerful than GPT-6 Astra. This is a detailed official claim, whereas @mark_k’s earlier thread reported only a rumor and confusion over the possible authors.

But several critical questions remain open:

1. Where is the full text of the Navier: Stokes proof?
2. Which independent experts have reviewed it?
3. Does it exactly meet the criteria of the Millennium Prize Problem?
4. What precise role did the agents play: invention, exploration, drafting, formalization, or verification?
5. How will OpenAI credit the contributions of Córdoba, Martínez-Zoroa, Buckmaster, Alpöge, and any other researchers?
6. What data and materials were used to train or guide the internal model?
7. Will questions about the Codex sessions receive a public, technical, and verifiable response?
8. What access will the scientific community have to the results, methods, and computational conditions?

The real dividing line is not simply between humans and AI agents. It is between a spectacular claim and the scientific requirements of verification, reproducibility, and fair credit. OpenAI has provided figures and described a method. The thread highlights everything those figures do not yet allow us to determine.