The Navier-Stokes row and what you type into a chatbot

The Navier-Stokes row and what you type into a chatbot

Marketing Sideways · Data and trust

A year of secret work lasted 88 hours

Two mathematicians spent a year on a secret discovery. They used AI chatbots to help them do it. A rival company then announced the same discovery in under four days, and nobody can prove how. If you use a chatbot for work, that should bother you.

Flat 1960s style illustration of a single sheet of paper multiplying into many identical copies behind a seated figure

Two mathematicians spent a year on a secret. A rival announced the same thing in 88 hours. Nobody can prove how, and that is the part worth your time.

On Tuesday, OpenAI published a 165 page proof about a famous maths problem. Every headline is about the maths. The story underneath it is about two researchers who were working on the same thing privately, and what happened to their secret. You do not need any maths to see why that one matters to you.

What actually happened

The Navier-Stokes equations were written in the 1800s to describe how liquids and gases move. They are why we can design an aircraft wing and forecast a storm. One question about them has stayed open for roughly 90 years. Can perfectly smooth flowing water tie itself into a knot so tight that the maths describing it breaks?

In 2000 the Clay Mathematics Institute put that question on a list of seven, each carrying a million dollar prize. Before this week, one of the seven had been solved.

OpenAI says the answer is yes. It says an internal model, running as a swarm of up to 10,000 agents for 88 hours, found a scenario where a spinning column of fluid grows thinner and faster until its speed runs to infinity. The company puts the computing bill in the millions of dollars and the agents' output at 130 billion tokens, which is the unit these systems produce text in. It says it will decline the prize money.

All of that is the easy part of the story.

The part that lands on you

Tristan Buckmaster at New York University and Levent Alpöge, a researcher at Anthropic, had been working on this territory quietly for most of the past year. Buckmaster has said publicly that they used models from both Anthropic and OpenAI to do it. OpenAI says its own sprint started on 1 September, after its researchers heard a rumour that a Millennium problem had fallen.

Buckmaster has raised the question of whether OpenAI's new model had access to those private sessions. OpenAI says its agents and its people saw none of the unpublished work before it went public. The company also says that while it considers it unlikely that data from people using its products improved the models, it cannot rule the possibility out.

Read that last sentence again, because it is the whole article. The strongest available answer to "did my work leak" is that it probably did not. Two of the most careful researchers in a high stakes field are now in a position where they cannot demonstrate what did or did not travel. Suspicion on its own is the damage.

Scale that down to a business your size. A July 2026 survey of 500 employed American adults, commissioned by the California law firm Kolmogorov Law and run through the Pollfish platform, asked workers directly what they had put into personal AI accounts their employer does not control.

Work information entered into personal AI accounts, US workers, July 2026
Any work information at all38%
Internal emails, memos or documents23%
Financial or sales figures12.4%
Kolmogorov Law, surveyed via Pollfish, July 2026, 500 employed US adults. The monitoring firm Cyberhaven reaches its own figures a different way, by analysing millions of real prompts rather than asking people to recall what they typed.
Flat 1960s style illustration of a typed line of text leaving a keyboard as a ribbon and entering a grid of identical circles
Self reported numbers usually run low, because people forget. Cyberhaven, which watches what employees actually paste, reports that roughly 11 per cent of it is sensitive material.

Nobody in that 38 per cent set out to hand anything over. They were fixing a subject line, tidying a quote, asking for a faster way to word a follow up. The document arrived in the box because the box is useful.

35.6% of the same US workers knew that entering confidential company information into a personal AI account can, in some circumstances, break the law. Source: Kolmogorov Law, July 2026.

Why the story sounds so confusing

There is a second fight running underneath the first, and that one is a marketing fight.

The published proof concerns the forced version of the equations, where a smooth outside force is applied to push the fluid along. OpenAI says this establishes statements C and D of the official prize formulation, so the problem is resolved. The Clay Mathematics Institute still lists Navier-Stokes among its unsolved problems, and its rules require any solution to appear in a qualifying publication, stay in print for two years and win general acceptance from mathematicians before a prize is even considered.

The proof itself was checked by machine, step by logical step. So correctness is settled. The argument is about the word solved.

You have seen this technique at a much smaller scale. Fastest growing in its category. Award winning. Number one for customer satisfaction. Choose the frame where the claim is true, then make the claim. It works, and it is fair game, right up until somebody with more authority publishes a different frame. Then the week you booked for celebration goes on explaining yourself.

Three things to do this week

Write the do not paste list. One page, five lines, on the wall. Client names, pricing, unlaunched work, anything under an agreement, anything with a real person's details in it. A rule you can recite beats a policy nobody opens. Do this first, because it is the only one that works while you are tired and in a hurry, which is when the pasting happens.

Check the training setting on every tool, one by one. Consumer accounts and business accounts handle your inputs differently, and the difference is a toggle most people have never opened. Go through each tool your team touches this week and write down what you found next to the tool's name. The audit takes twenty minutes and it is the only version of this you can show a client.

Read your own confidentiality clause before a client reads it to you. The survey above is American and the legal points around it are American, so the Australian read is inference rather than fact. Your exposure here sits in the agreements you have already signed and in the Privacy Act, which means the answer is in the contract in your drive rather than in the news. If you handle other people's material, that clause decides what a mistake costs you.

Treat every prompt as a document you have already published.

Sources. Reporting from The Wall Street Journal, Quanta Magazine and Axios, 8 September 2026. Figures from OpenAI, the Clay Mathematics Institute, Kolmogorov Law surveyed via Pollfish in July 2026, and Cyberhaven. Statements from Tristan Buckmaster of New York University.

Mashed Avocado · Marketing Sideways · MashedAvocado.com

Back to blog

Leave a comment

Please note, comments need to be approved before they are published.