Vibe coding gave us apps nobody could maintain and now ‘vibe mathing’ may give us proofs nobody can actually learn from.
Some context: Andrej Karpathy, one of OpenAI’s founding members, coined ‘vibe coding’ in February 2025 to describe a style of programming where you let AI do the work and “forget that the code even exists”. Collins Dictionary made it its word of the year. Karpathy has since moved on to ‘agentic engineering’, which involves a lot more checking of the work, or what the rest of us call doing your homework.
Now maths is getting the same treatment. On 8 September 2026, OpenAI announced that around 10,000 AI agents, running on an unreleased model, had solved the Navier-Stokes problem in 88 hours. It is one of the Clay Mathematics Institute’s seven Millennium Prize Problems, each worth $1 million. The Clay Institute has yet to give its verdict.
The Navier-Stokes equations describe how liquids and gases move and they’re used in everything from weather forecasting to aircraft design. The big open question was whether they can ever break down: could a perfectly smooth flow end up, in a finite amount of time, with fluid spinning infinitely fast? OpenAI’s answer is yes, by way of a vortex that tightens and then stretches out “like spaghetti”.
The Stokes in question, by the way, was George Gabriel Stokes, born in Skreen, Co Sligo, in 1819. A Royal Society paper credits his childhood beside the Atlantic with sparking his interest in waves. According to an Irish Times profile, he never engaged in small talk (with one recorded exception). He would have made a terrible chatbot.
Anyway, how the problem was solved is messier than the equations. OpenAI admits it only started on 1 September after hearing that others were getting close. “Just like everyone, we saw rumours on Twitter,” OpenAI researcher Sébastien Bubeck told reporters.
The ‘others’ were NYU mathematician Tristan Buckmaster and Levent Alpöge, who works for rival Anthropic. They’d spent the past year on the problem, partly using OpenAI’s own Codex tool, and Buckmaster has publicly questioned whether his drafts could have influenced OpenAI’s model. OpenAI initially said it couldn’t rule this out and then changed its tune, saying on 10 September that an investigation confirmed his prompts couldn’t have influenced the result.
Both teams, meanwhile, built on a strategy developed by Spanish mathematicians Diego Córdoba and Luis Martínez-Zoroa who have been called the heroes of the story by Princeton’s Charles Fefferman, the guy who wrote the Clay Institute’s official description of the problem.
To be fair to the machines, maths has something coding lacks: a referee. OpenAI’s proof was checked in Lean, software that verifies every step and cannot be charmed by vibes. So the answer is highly likely correct but being correct isn’t the same as understanding what you’re correct about.
Terence Tao, one of the world’s greatest living mathematicians, put it pretty memorably. He compared the automated strip-mining of open problems to digging treasure out of an archaeological site with an excavator: you might get the treasure but you destroy the context that gave it meaning. He even suggested some problems might need to be declared off-limits to AI in a similar manner to how it’s universally considered socially unacceptable to divulge spoilers to films and TV shows – you try it and you’re a pariah. Maths, apparently, now needs a spoiler policy.
Before this week, only one Millennium Prize Problem had been solved, by Grigori Perelman. In 2010 he turned down the prize money, saying US mathematician Richard Hamilton’s contribution was no less than his own. OpenAI said it won’t claim the prize either. By one estimate, it spent about $1 million of its own computing power getting there (and used god knows how much power and water. Ho hum).
So one mathematician turned down a million dollars because the credit should have been shared while one company spent roughly the same after reading a rumour on Twitter. That’s the difference between doing maths and ‘vibe mathing’.
Ltd
