Frames the announcement skeptically by noting the internal model has been training for less than two weeks yet is being pitched as more than twice as capable in mathematics as Astra, which OpenAI released just a week earlier. The framing suggests the marketing claims are outpacing what the model has actually demonstrated.
Covers the announcement as the moment the AI-vs-mathematician priority question became unavoidable, treating the claim as significant but unverified. His framing highlights that a non-peer-reviewed corporate blog post is being positioned as a Millennium Prize breakthrough.
Points out that even a rumor of someone working on a problem is now enough to trigger a massive AI-powered effort to 'flatten it before the original researcher can finish.' Tao frames this as a structural threat to how mathematical priority and credit are assigned.
Within hours of the post, mathematicians noted that Tristan Buckmaster at NYU has been publicly working the finite-time blow-up angle for years, with prior results that appear to overlap the announced approach. The implication is that the AI system is completing or repackaging an existing research trajectory rather than producing an original proof.
Notes that finite-time blow-up is not fringe — it's a Millennium Problem with $1M attached, and community consensus over the last decade has drifted toward blow-up being the likely answer for 3D incompressible Navier–Stokes. Tao's 2014 averaged-Navier–Stokes work and Buckmaster–Vicol's 2019 results already pointed in this direction, so a proof along these lines would validate rather than upend expert intuition.
On September 8, OpenAI posted a page claiming that an internal system had produced a solution to one of the Clay Millennium Prize Problems: the existence and smoothness of solutions to the 3D incompressible Navier–Stokes equations. The specific claim is that the dynamics can develop a singularity in finite time — the blow-up side of the problem, not global regularity.
The model isn't a shipped product. According to community reporting on the announcement, it's an internal system that has been training for less than two weeks. One commenter on the Hacker News thread, which pulled 1,308 points, summarized the framing bluntly: an internal model trained for a fortnight is being pitched as more than twice as capable in mathematics as Astra, which OpenAI made public roughly a week earlier.
The proof itself has not been peer-reviewed, and within hours of the post, mathematicians began pointing out that at least one researcher — Tristan Buckmaster at NYU — has been publicly working the finite-time blow-up angle for years, with prior results that appear to overlap the announced approach. Terence Tao weighed in on Mathstodon with a pointed observation that even a rumor of someone working on a problem is now enough to trigger a massive AI-powered effort to "flatten it before the original researcher can finish." Simon Willison also covered the announcement, framing it as the moment the AI-vs-mathematician priority question became unavoidable.
There are two stories tangled together here, and it's worth separating them.
The first is the mathematical one. Navier–Stokes finite-time blow-up is not a fringe conjecture — it's one of the seven Millennium Problems, with a $1M prize attached, and the community consensus for the last decade has drifted toward blow-up being the likely answer for the 3D incompressible case. Tao's own averaged-Navier–Stokes work in 2014 was widely read as evidence in that direction. Buckmaster and Vicol's 2019 non-uniqueness result was another step. So the surprise isn't the conclusion — it's the claim that a two-week-old model produced a full proof at a level of rigor that would clear Annals of Mathematics review. That claim needs to survive months of scrutiny before it means anything, and the base rate for announced Millennium-problem proofs surviving review is close to zero.
The second story is the sociological one, and it's the one Tao is actually angry about. If you can point a model at any problem a human is publicly working on and generate a competitive draft in days, the incentive to work in the open collapses. Mathematics has run for decades on preprints, conference talks, and informal circulation of ideas — a norm that assumes nobody can weaponize your half-finished work against you at industrial scale. A world where announcing you're working on a problem invites a race you can't win is a world where mathematicians stop announcing. That's a genuine change in how the field operates, and it doesn't require the OpenAI proof to be correct to happen. The threat alone is enough.
The HN thread captured the split cleanly. One camp — represented by comments like arctic-true's — thinks the drama is a distraction from the actual headline, which is a step-change in AI mathematical capability that the public Astra release didn't hint at. If an internal model really is >2x Astra on math benchmarks after two weeks of training, that's the story regardless of who deserves credit for this particular proof. The other camp, echoing Tao, argues that credit and process aren't separable from capability — an AI that can only "solve" problems by racing humans who've done the load-bearing work isn't actually doing the interesting part.
Both camps are right about different things. The capability jump, if real, is genuinely large. The process concern, if the priority allegations hold, is also genuinely large.
For working developers, the immediate practical implication isn't about fluid dynamics. It's about what "internal model, two weeks of training" implies for the release cadence you're planning around.
If OpenAI's public models are lagging their internal frontier by a factor of two on hard reasoning tasks, any architecture decision you make today assuming current-generation capability is already stale. The gap between what's in the API and what exists on a research cluster has always been there, but a 2x math delta on a two-week-old model is a wider gap than most teams have been sizing for. If you're building anything where "can the model handle novel formal reasoning" is on the critical path — theorem-adjacent tooling, formal verification assistants, symbolic math in engineering CAD, static analysis with proof obligations — the useful planning horizon just got shorter.
The second implication is about how you evaluate claims like this one going forward. The proof is a PDF on a corporate blog. It has no institutional review, no Lean formalization, no independent implementation. That's fine as a preprint; it's not fine as a product claim. When your PM forwards you the next "AI solved X" announcement, the question to ask is whether the artifact is checkable — Lean file, runnable code, reproducible benchmark — or whether it's a narrative wrapped around a screenshot. Navier–Stokes will get checked because the prize money guarantees it. Most claims won't.
Third: if you contribute to open research, take the Tao warning seriously. The old norm of "post the preprint early so nobody scoops you" now cuts the other way when the potential scooper is a well-funded model that can be pointed at your abstract. Whether that changes your behavior depends on your field, but it's a real strategic question now.
The next 90 days will resolve most of this. Either the proof survives review from Buckmaster, Tao, and the fluids community — in which case OpenAI has genuinely accelerated frontier mathematics, and the priority argument becomes a policy debate about how AI labs should engage with active research programs. Or the proof has gaps, gets retracted or heavily amended, and the announcement becomes a case study in shipping before peer review. Either outcome reshapes how AI capability claims get evaluated. The one thing that isn't happening is the story staying quiet.
My take:1. It shows what even this wave of AI can actually do.2. I wish it were done by different folks, ideally under some kind of public control like NASA research or the NPR model.3. Keep in mind: natural science is different. It's not always a matter of computation. Computer science folks o
"we cannot rule out that de-identified data derived from their usage of our products helped improve our models ."What a landmine sentence to bury in this report, you can't rule out your models were spying on other researchers?
> We’re sharing a solution to the Navier–Stokes existence and smoothness problem, one of the Millennium Prize Problems. This proof, produced by an internal OpenAI system, shows that the dynamics of the Navier-Stokes equations for fluid motion can develop a singularity in finite time. We’re sharin
For full context, here's the HN thread from the other side of the "Concurrent Work" section: https://news.ycombinator.com/item?id=49605915Unlike the vanilla read of the OpenAI press release, it is much more unfiltered and outlines some particularly aggressive behavior b
Top 10 dev stories every morning at 8am UTC. AI-curated. Retro terminal HTML email.
Buried under the drama is the fact that OpenAI is claiming that an internal model they’ve been training for less than two weeks is more than twice as capable in mathematics as Astra, which was only made public a week ago. Even if this improvement is limited to mathematics, that is an astounding feat