OpenAI e Navier–Stokes: cosa è stato davvero dimostrato e cosa manca ancora

Versione italiana: OpenAI e Navier–Stokes: cosa è stato davvero dimostrato.

In the first week of September 2026, two announcements overlapped within hours: on one side Tristan Buckmaster (NYU Courant) and Levent Alpöge (mathematician, researcher at Anthropic) published three blow-up results with smooth forcing for fluid equations, verified in Lean; on the other, OpenAI told a press briefing it held a proof of blow-up for the forced Navier–Stokes equations, obtained in about 88 hours with an internal model. The second announcement travelled around the world. The first is the one that can actually be read today.

This article does not pick a side between laboratories. It does something more useful: it separates what was claimed from what was shown, explains why the distinction between Euler and Navier–Stokes is decisive, what a proof assistant such as Lean truly certifies, and which path a machine-generated result would need to become an accepted proof — up to the Clay Millennium Prize of one million dollars.

Is Navier–Stokes actually solved? / È davvero risolto?

Direct answer: no. As of 8 September 2026, the Navier–Stokes existence and smoothness problem remains officially open, and the Clay Mathematics Institute still lists it among its unsolved problems. OpenAI claims to hold a Lean-verified proof of forced blow-up. Nobody outside OpenAI has seen it.

That sentence should be taken literally. It is not prejudiced scepticism: it is the minimum standard of mathematics. A proof exists when it is public, readable or at least mechanically inspectable, and survives independent scrutiny. All three steps are missing for the OpenAI result, while they are present — with the limits discussed below — for the Buckmaster–Alpöge result on related equations.

What did OpenAI claim to prove? / Cosa sostiene di aver dimostrato OpenAI?

Direct answer: according to what OpenAI told reporters on 8 September 2026, an internal model “significantly more capable than GPT-6 Astra” produced a roughly 100-page proof of finite-time blow-up for forced Navier–Stokes on R³ and T³, with smooth forcing under Fefferman options c and d. The proof is said to be certified in Lean. OpenAI claim

Mind the hierarchy of sources, which most coverage has blurred. The primary source for these figures is not a paper but what OpenAI said in a briefing, as reported by Axios and Scientific American. It is a primary source on the briefing, not on the mathematics.

1. 19/09/2025 — Alpöge writes to Buckmaster: a personal collaboration on fluids with AI, worried that large companies might claim the great results. Independent evidence
2. 01/08/2026 — OpenAI publishes “Ten advances in mathematics”: Astra solves 10 open problems, about $2,000 in tokens, human manuscripts plus Lean. Mathematically established as an account of the method, pending scrutiny of each result.
3. 15/08/2026 — Buckmaster–Alpöge breakthrough on Boussinesq and Euler with smooth forcing; Lean on 22/08. Independent evidence
4. 01/09/2026 — According to OpenAI via Axios, start of the Navier–Stokes sprint after rumours of “two solved Millenniums”. OpenAI claim
5. 03/09/2026 — Buckmaster writes to a mathematician at OpenAI to clarify the personal nature of the work. Independent evidence
6. 06/09/2026 — Two calls with Sébastien Bubeck: OpenAI describes the ~100-page proof and two editorial proposals. Buckmaster's version, generally rejected by Bubeck. Status: conflicting accounts.
7. 07/09/2026 — Tao publishes his analysis of the Euler/Boussinesq papers; Anandkumar et al. publish an unforced candidate via PINN. Independent evidence
8. 08/09/2026 — Buckmaster statement plus OpenAI briefing via Axios/Scientific American plus Bubeck's “false and inflammatory” reply. The OpenAI proof is still not public.

The eight extraordinary points require a second independent check before they can be treated as facts. Here they are with their actual status:

The editorial rule applied throughout: no amount of optimization can justify a statement stronger than the available evidence.

What did Buckmaster and Alpöge actually prove? / Cosa hanno dimostrato Buckmaster e Alpöge?

Direct answer: finite-time blow-up with smooth forcing for three equations: incompressible porous media (IPM), 2D Boussinesq and 3D incompressible Euler. Papers formalized in Lean and publicly discussed by Terence Tao as a “remarkable achievement”. Mathematically established as preprint plus certificate, pending full peer review.

The conceptual credit belongs in the right place: the programme was started by Diego Córdoba and Luis Martínez-Zoroa, who spent years building forced blow-ups with rough forcing. Buckmaster and Alpöge — with heavy LLM assistance including Claude and Codex on GPT-5.6 Sol — pushed it to smooth forcing and to Euler. Buckmaster himself writes that the idea is theirs, going as far as proposing Martínez-Zoroa for the Fields Medal. A rare example of generous attribution in a week of disputes.

Limits stated by the authors: the preprints, especially on Euler, are described as “AI slop”, released hastily under outside pressure and still to be digested in writing; the sibling result on hypo-dissipative Navier–Stokes is withheld because Lean verification is unfinished. The available evidence does not establish any complete extension to standard Navier–Stokes on their side.

What does “solving” Navier–Stokes mean? / Cosa significa “risolvere” Navier–Stokes?

Direct answer: Clay asks for either a proof of global regularity for all smooth data, or an admissible blow-up counterexample. Blow-up with smooth forcing is close to the second route, but it must sit exactly inside Fefferman options c and d to count as a Millennium solution.

Intuition before formalism. The equations describe water, air, smoke: velocity and pressure evolving through inertia, pressure, viscosity and an optional outside push, the forcing. “Solving” here does not mean simulating a pipe well or forecasting weather — things engineering already does — but answering a pure question: starting from a perfectly smooth initial field, does the solution stay smooth forever, or can it develop a singularity in finite time, a point where velocity or vorticity diverge?

Distinctions that most recaps blur: Euler is the no-viscosity case, more blow-up prone and further from everyday physics; Navier–Stokes includes the viscosity that smooths; IPM and Boussinesq are simpler models used as test beds. A numerical or heuristic solution exhibits a candidate; a rigorous proof closes every logical loophole; existence and smoothness means building global smooth solutions, while blow-up exhibits one that breaks. Smooth forcing is an infinitely regular outside push: physically gentle, mathematically far harder to tame than rough forcing.

Result	OpenAI claim	Independent evidence	Established / peer reviewed
Forced smooth 3D Euler (Buckmaster–Alpöge)	—	Public preprint plus Lean, Tao note	Under scrutiny, no journal plus 2 years yet
Forced smooth Boussinesq / IPM	—	Public preprint plus Lean	Under scrutiny
Forced Navier–Stokes (OpenAI)	08/09 briefing, ~100 pp., Lean stated	None: proof not public	No
Hypo-dissipative NS (Buckmaster–Alpöge)	Announced without paper	Lean unfinished	No
Unforced Euler via PINN (Anandkumar et al.)	Numerically stable candidate	Paper plus partial Lean	No, stability proof missing

What did the AI actually do? / Cosa ha fatto concretamente l'AI?

Direct answer: in the readable papers, the AI generated proof candidates — described by the authors as the worst draft ever read — then checked in Lean and rewritten by humans over weeks. In the OpenAI case, the account speaks of thousands of agents and warm-ups on easier problems, but with no public artefacts no method can be reconstructed. Independent evidence for the first case, OpenAI claim for the second.

The Córdoba–Martínez-Zoroa programme gives the idea: start from a forced solution and add high-frequency corrections that concentrate energy toward the blow-up time while keeping the forcing well behaved. For Boussinesq there is an explicit ansatz — an almost linear field plus a high-frequency plane wave — that solves exactly and yields an unstable modulation ODE system. The rest is spatial cutoffs and estimates: dozens of pages of technical bookkeeping. That is where LLM plus Lean change the pace: the machine proposes, Lean mercilessly rejects or accepts, humans try to understand what was actually shown.

1. 1. Human idea. Córdoba–Martínez-Zoroa open the forced blow-up route with rough forcing.
2. 2. AI push plus Lean. Alpöge–Buckmaster generate variants via LLM, Lean checks on 22/08, humans rewrite for weeks up to smooth forcing.
3. 3. NS extension. Tao sees no in-principle obstacle toward Navier–Stokes, but expects enormous technical difficulty. OpenAI claims to have burned through it in 88 hours — without showing how.

On the Codex question, technical precision is needed. “The model did not look up user data” concerns runtime lookup; the question about training — whether drafts uploaded into Codex sessions ended up in training — is different and, according to the statement, went unanswered. It establishes no misuse, but it does not exclude it either. For anyone doing research with proprietary tools, that is the real practical warning of this story.

What does Lean actually verify? / Cosa verifica davvero Lean?

Direct answer: Lean checks that a formalized proof has no logical gaps starting from given axioms and definitions. It does not check that the formalized statement matches the Clay problem, nor that the proof is understandable or important.

This distinction decides everything. A public Lean certificate is a very strong guarantee of internal correctness: anyone can recheck it mechanically. But it remains possible to formalize the wrong statement, to use stronger hypotheses than Fefferman allows, or to produce 100 correct yet unreadable pages that teach nothing. Tao puts it clearly: the value lies in the extracted ideas, not in the seal. Without a public certificate, by contrast, “verified in Lean” is only a sentence in a briefing.

What does the Clay Institute require? / Cosa richiede il Clay Institute?

Direct answer: publication in a qualifying journal, at least two years of scrutiny, general acceptance by the community, and a decision by the scientific board. Even a correct proof today would win the million no earlier than 2028.

1. 1. Public preprint plus Lean certificate — Euler/Boussinesq: yes. OpenAI NS: no.
2. 2. Human reading and digestion — Tao spoke 30 minutes with Buckmaster; the papers are still works in progress.
3. 3. Qualifying journal — none of the September 2026 results is there yet.
4. 4. Two years of general acceptance — citations, conferences, independent checks.
5. 5. Clay decision — the board may also decline to award or credit prior work.

Why is trust the real bottleneck? / Perché il vero collo di bottiglia è fidarsi?

Direct answer: because an AI system can produce a formally checkable proof that humans have not yet understood, and mathematics accepts a result only when the community understands it — not when a machine says “ok”.

This is the central thesis. Navier–Stokes is the case study for a larger question: when an AI proves a theorem nobody can yet explain, who decides that it is true? Tao already glimpses the poisoned scenario: a black-box solution that closes the problem as a prize but contaminates it as a source of ideas, because nobody extracted the mechanism. The useful work now is not “throwing enormous compute at it” but digesting the method — the more unstable modulation ODEs, the better-localized corrections — and turning it into transmissible understanding. Generating has become cheap; understanding stays expensive.

Everyone repeats	Almost nobody explains
100 pages, 10,000 agents, millions of $	Fefferman options c/d and why smooth forcing is close to Clay without automatically counting
“Verified in Lean”	What Lean guarantees and what it does not
Euler equals Navier–Stokes	Viscosity, IPM/Boussinesq models, the real distance between results
OpenAI vs Anthropic dispute	The Córdoba–Martínez-Zoroa method as the true conceptual protagonist
“Millennium solved”	The Clay path: journal plus 2 years plus general acceptance

What does the community say? / Cosa dice la comunità?

Direct answer: cautious technical validation for forced Euler, scepticism about the physical hype, conflicting versions of the dispute.

Tao validates the mathematical core, explains the ansatz, and sees possible extensions to NS with no in-principle obstacle, while expecting enormous difficulty. As an aside, he preferred having the ideas explained over the phone by Buckmaster — “a refreshing change from AI-based communication”. The same day he flags the independent work of Ganeshram, Duruisseaux and Anandkumar on unforced Euler via physics-informed neural networks: a numerically stable candidate, but the rigorous stability proof is missing, with only partial Lean and AI used mostly for literature and formalization.

On Hacker News and PDE forums the dominant sentiment is sober: blow-up is a pure-mathematics problem, constant-viscosity equations have known limits as models, turbulence remains a matter of computational complexity more than singularities. On the Buckmaster–Bubeck dispute we report both versions without taking sides: on one side the 06/09 calls, proposals of a coordinated announcement or single authorship with model acknowledgement, a double request to exclude Alpöge over the Anthropic affiliation, and remarks reported as threatening; on the other, Bubeck's flat denial as “false and inflammatory” with a fuller reply promised, plus the mirrored posts of Noam Brown and Sholto Douglas. The available evidence does not establish anything beyond this.

What we know and what we do not / Cosa sappiamo e cosa non sappiamo — 8 September 2026

An unsensational conclusion. Even if the OpenAI proof were correct, it would still need reading, understanding, publication and two years of scrutiny. If it were wrong or outside the Clay target, the sibling result would remain: a human programme — completed with machines — that genuinely moved the boundary of forced blow-up. Either way, the news is not “the AI won a million in 88 hours”. It is that mathematics has a new bottleneck: not generating proofs, but collectively deciding when a machine-generated proof counts as a proof as such.

Method and limits: this article rests on documents public as of 08/09/2026 17:30 UTC. Figures and descriptions of the OpenAI proof are attributed to the briefing via Axios/Scientific American or to the Buckmaster statement, not independently verified. We will update the box above when preprints, Lean certificates or documented rebuttals appear.

Frequently asked questions / FAQ

Has Navier–Stokes been solved?

No. As of 8 September 2026 it remains an open Millennium Prize. There are Lean-verified preprints of forced blow-up for sibling equations (Euler, Boussinesq, IPM), but the OpenAI proof on Navier–Stokes is not public and has no independent verification.

What did OpenAI show according to reports?

As reported from the 8 September briefing, finite-time blow-up for forced Navier–Stokes with smooth forcing on R³ and T³, in about 100 Lean-certified pages. An attributed description, not a public document.

What did Buckmaster and Alpöge show?

Blow-up with smooth forcing for IPM, Boussinesq and 3D Euler, with public Lean formalization, extending the Córdoba and Martínez-Zoroa programme. Terence Tao calls it a notable advance, distinct from the full Clay problem.

What does Lean actually verify?

The internal logical correctness of a formalized proof — not whether the statement matches the Clay problem, nor whether it is understandable or mathematically important.

When would the Clay million be won?

Only after publication in a qualifying journal, at least two years of general community acceptance, and a positive decision by the Clay scientific board. Even with a correct proof today, not before 2028.

Why is Euler not Navier–Stokes?

Euler describes fluids without viscosity, more prone to singularities; Navier–Stokes includes the smoothing viscosity. A forced Euler blow-up is a serious step, but it does not automatically equal a Navier–Stokes counterexample as formulated by Fefferman.

Primary and secondary sources / Fonti

Primary evidence — Buckmaster statement (NYU Courant PDF); 08/09 Mastodon post; Terence Tao analysis 07/09 and Mathstodon thread; Anandkumar et al. paper 07/09 on Euler via PINN; OpenAI page “Ten advances” 01/08/2026; Clay rules and Fefferman description. Attributed secondary reporting — Axios 08/09 for briefing figures (10k agents, 88h, millions of $); Scientific American 08/09 for framing and stated Lean certification; The Decoder, OfficeChai, Glitchwire, NewsBytes for chronicles of the dispute. OpenAI proof: no public preprint, repository or certificate found at verification time. Italian version of this analysis: OpenAI e Navier–Stokes: cosa è stato davvero dimostrato.

On our method for working with AI: AI for SMEs, how we structure citable content and building sites with AI without losing quality.

Do you have technical content that must survive scrutiny?

We build fast, structured, verifiable pages — the same principles as this analysis: sources, dates, tables and structured data.

Talk to us →