An unreleased OpenAI model has produced a proof for Navier-Stokes, a problem mathematicians have not closed in more than 100 years, and Johan Land, who read the proof, says it looks accurate.
New models are normally judged on benchmark scores. Land says benchmarks built by humans are finished as a measure, because the models now clear them, and that the only test left is problems no human has solved.
"So, you need to measure something the humans can't do. And that's math."
Land is chief product officer at Samsara, competed in the International Math Olympiad, and says he has used swarms of released models to close four problems that had gone unsolved for 50 years.
I listened to the full segment so you can skip it.
Here are the 8 takeaways that matter.
👤 Guest: Johan Land, Chief Product Officer at Samsara, a former International Math Olympiad competitor who uses swarms of AI models to attack unsolved mathematics
🎙️ Host: Akash Pasricha, who anchors TITV, The Information's live weekday news show at 10 a.m. Pacific
📰 Published: 9 September 2026 on YouTube (The Information)
🔴 YouTube | ⏱️ 12 min
Key Takeaways
Land read OpenAI's Navier-Stokes proof and says it is accurate, with only a small part of the math community objecting
The open question was whether a fluid of limited viscosity can reach unlimited speed; OpenAI built a scenario where it does
Two researchers who had been working on the same problem for a year used OpenAI's Codex along the way, and now ask whether OpenAI saw their work
OpenAI's push came after it heard a rumor of their breakthrough, in Land's account spending "10 billion plus on like 10,000 plus agents"
Math falls to AI before any other science because a proof can be checked by a computer
Chemistry and physics need real-world experiments, which are slow and uncertain
Human-built benchmarks are exhausted, so the models are now tested on problems humans have not solved
Land says he has solved four unsolved problems himself, using a swarm of released models rather than OpenAI's unreleased one
He compares resistance to AI-assisted proofs to the reception of electronic music
He cannot construct a logically consistent scenario in which AI does not get out of control
He says safety is not the labs' job but society's, and that no one can enforce it because anyone can defect
1. AI Cracked Navier-Stokes
OpenAI announced that its latest unreleased model solved one of mathematics' Millennium problems — a problem, the host said, that mathematicians have not been able to solve for more than 100 years.
Navier-Stokes itself is settled and useful: it is the equation that describes fluid dynamics, and Land said there is no dispute about that part of it
The open question is about extreme behavior. Under certain conditions, does the equation produce a singularity?
His illustration was a cannonball jump into a pool: the water is pushed to the side, comes back together, and shoots upward very fast
Stated precisely, the question is whether a fluid with limited viscosity can reach unlimited speed, and settling it means a proof either way: "You need to prove either that it exists or that it doesn't exist"
OpenAI's model produced a scenario in which the singularity happens, and formalized it completely
Land checked it himself: "And I looked at their proof. I looked at the formulation and like to me it all looks accurate"
A very small part of the math community is challenging the result. His read on the rest: "It seems like this is accepted"
2. The Codex Data Question
The dispute is not about the mathematics. It is about whose work fed the proof, and Land laid out the sequence before saying where it stops being documented.
Two researchers he named as Tristan and Levent — one at NYU, one at Anthropic — had been collaborating on the problem for the past year, using a lot of Anthropic models and also OpenAI's Codex
About three weeks ago they got a breakthrough and thought they probably had it; a week later they had formalized it, but did not like their own proof and took time over it. Land believes they did not yet have the full solution
OpenAI then heard the rumor and moved at scale. In Land's telling, about a week ago it began "spending like 10 billion plus on like 10,000 plus agents trying to solve the six remaining Millennium problems," and got traction on Navier-Stokes
The pair heard the rumor in turn and approached OpenAI. Everything to that point, Land said, is fact and well documented
What follows is blurry. OpenAI appeared to invite a collaboration, but in his account did not want both of them because of the Anthropic affiliation, and the approach was rejected
OpenAI published its proof. The two researchers then said it looked similar to theirs, and asked, since they had been using Codex, "did you actually look at our work when you created the proof"
The host's reaction: "can't make this up." Asked whose side he is on, Land declined to take one
3. Math Falls to AI First
Land's answer to the whose-side question was that the argument is a symptom of something larger: the science of mathematics is falling to AI, and people are working out how to relate to that.
His reason math goes first is verification, not difficulty. A proof can be checked by a computer, and if the check passes, the result is true and the matter is closed
Nothing else in science works that way. In chemistry or physics a claim has to be tested by running an experiment in the real world, which takes time and returns an uncertain answer
On that reading, the priority fight between two labs is what the transition looks like from inside, rather than an isolated dispute about one proof
4. Benchmarks Humans Can't Do
The host put the obvious mismatch to him: models close century-old problems, and then hallucinate on a simple task or mishandle an order placed through a food-delivery app.
Land's answer was that most of those failures are product failures — "I mean some of those in my opinion is like people building bad products and not using the AI in the right way and that will catch up"
He called those growing pains, and said he is a firm believer that they get resolved
His framing is that the industry has entered what he calls the post-AGI era, which is why mathematics is now the frontier. Model releases used to be judged on benchmark percentages; the models are now hitting close to 100% on them
The reason the scores stopped meaning anything is that the benchmarks were designed by humans to measure models against humans: "The models are now better than humans"
What replaces them is unsolved problems, because a human cannot do those by definition
5. His Own Swarm of Models
The host confirmed that the model behind the proof was an unreleased version of Astra, more capable than the released one. Land agreed — "It is much more capable" — and drew the contrast with his own work.
"So I produce as well and I've solved four unsolved problems," he said, adding that they are not as prominent as the Millennium Prize problems
The problems he has closed are ones he says are 50 years old, which humans had tried and failed to solve
He does it with released models, not frontier internal ones: he named Astra, Fable 5.1, Gemini 3.8 and Kimi K3, and deploys them together to attack a single problem
The distinction he drew is that OpenAI is working with unreleased models that are more capable than anything he can use, which is why its result reached a Millennium problem
6. The Electronic Music Turn
Asked whether people shout at him for solving problems with models, Land said yes, and that a community with rivalry has formed around the question.
His analogy was music. When electronic and digital instruments arrived, people said the result was not real music
The objection did not survive the output. "But right now, no one would question is like Taylor Swift or Michael Jackson, are they real musicians?"
He described the present moment as a transition phase in which AI is being accepted as a real tool for proving mathematics, and said that acceptance is growing
7. 13 Scenarios, Few Happy
The host turned to safety, citing an Anthropic researcher who resigned this week saying the technology will get out of control and that he was unimpressed by both OpenAI's and Anthropic's approaches. Land's answer was that he has looked for a reassuring conclusion and cannot find one.
He cited a researcher at MIT who defines 13 potential end-state scenarios. Some are happy; most are negative
The happy ones do not look stable to him. "I'm not sure that I believe greatly in an awesome outcome here," he said
He does not think a decline has to be violent: "We might live in abundance and just a slow decline as the AI takes care of us and we live out our lives in a happy fashion"
His comparison was to the Neanderthals — "I mean I'm not sure that the decline of the Neanderthals was so dramatic" — living out their lives while being out-competed
The version he put most plainly was custodial: "You know they will have us live like zoo animals and give us all we need and we'll die out slowly"
Asked directly whether he believes it could get out of control, he said: "I struggled finding a scenario that is logically consistent where that isn't the most plausible outcome"
The host's response was that he could now see why Land was good at math contests. Land said it matters to speak the truth with clarity, and that "I think that the situation is alarming"
8. Nobody Can Police This
Asked whether the AI labs are doing enough on safety, Land said the question is aimed at the wrong party.
"I think it's probably not their role. Like this is society," he said
Pressed on whose role it is, he described a problem with no enforceable answer: "The problem is that this is a prisoners dilemma, right? Anyone can cheat"
Assign it to nations and any nation can defect; to companies and any company can; to individuals and there are too many of them to police
Open-weight models make the gap easy to route around, because in his estimate they are only three to six months behind the frontier
That is why he cannot find a path to a stable outcome with what he called longevity for humanity — a happy decline, in his framing, is not the same thing
Bonus Insights
The host opened by disqualifying himself: he likes math, "I did the times tables as a kid," but hated the contests he sat in middle school and high school, and says he skips every headline about a model clearing another exam. This one he could not ignore
Land competed in the International Math Olympiad when he was young, which the host took as proof he had been good at exactly those contests
Land teed up the priority dispute as entertainment before telling it: "if you got the popcorn out, this is where you really want to sit back"
The segment ended early by the host's own account — he said he had a lot more questions and would need to bring Land back to unpack them
Land's bottom line is that mathematics is the first science to fall to AI because a proof is machine-checkable, and that the fight over who solved Navier-Stokes first is a symptom of that shift rather than an argument about one result.
Products, Companies & Tools Mentioned
OpenAI (Says an unreleased, more capable version of its Astra model produced the Navier-Stokes proof, after throwing what Land describes as 10,000-plus agents at the remaining Millennium problems)
Codex, OpenAI's coding product, which the two rival researchers had been using — the reason they are asking whether their work reached OpenAI
Anthropic (Employs one of the two researchers; Land says that affiliation is why OpenAI's collaboration offer appeared to exclude him. A departing Anthropic researcher's safety warning framed the second half of the segment)
Samsara (Land's employer, where he is chief product officer)
Gemini and Kimi K3 (Two of the released models Land runs as a swarm against unsolved problems, alongside Astra and Fable 5.1)
New York University (Home institution of one of the two researchers who say the published proof resembles theirs)
MIT (Where the researcher Land cites has defined 13 possible end states for AI, most of them negative)
Books & Resources Mentioned
Millennium Prize Problems – Clay Mathematics Institute (The seven problems at issue; Navier-Stokes is one of the six that were still open)
OpenAI Math Result Stokes Data-Sharing Concerns – The Information (The show's own reporting on the dispute this segment covers, linked from the episode's notes)
If this was worth your time, send it to someone closer to the industry than you are.
Get the latest market chatter as it happens: