The Navier-Stokes Scandal: When AI Research Becomes a Race to the Bottom
Simon Willison's sharp analysis reveals a disturbing parallel to security vulnerabilities
The Navier-Stokes Millennium Prize story isn't just about math. It's about what happens when AI research becomes a competitive sport — and the rules haven't caught up.
Simon Willison published his take yesterday, and it cuts to the bone.
The Short Version
Tristan Buckmaster (NYU) and Levent Alpöge (Anthropic) spent nearly a year working on Navier-Stokes existence and smoothness. They used Claude and Codex extensively. On August 15th, they had a breakthrough.
The mathematical rumor mill kicked in. OpenAI heard that Anthropic had "resolved a major open problem." OpenAI launched their own effort on September 1st. By September 5th — 88 hours later — their agents had a solution. 300 billion output tokens. ~$15M in compute at public API prices.
Buckmaster's statement raises uncomfortable questions:
"I asked whether the model had been trained on, or had access to, our sessions in Codex, into which we had been putting all our drafts for the whole of this project. I was told the model did not look up user data. I asked again, about training, and I did not get an answer."
Willison's Key Insight
Willison draws a parallel to computer security that's hard to unsee:
"Just a rumour of a bug is enough to find a security exploit these days... Is the same now true of mathematics? Just knowing that there is an unpublished solution to a problem is enough to find it?"
This reframes everything. The issue isn't whether OpenAI "stole" the work. It's that knowing a solution exists is now sufficient to find it — if you have enough compute.
The Compute Asymmetry
The Buckmaster-Alpöge team worked for a year with AI assistance. OpenAI spent $15M and 88 hours. Both reached solutions.
This isn't a story about intelligence. It's a story about resource asymmetry. The team with 300B tokens and internal model access won the race, even though they started months behind.
OpenAI's defense — that they didn't access user data — misses the point. The rumor itself was the exploit.
What This Means for Science
Academic research depends on openness. You share ideas, get feedback, publish when ready. If knowing a solution exists is enough for a well-funded lab to replicate it in days, why would anyone share their work-in-progress?
We could be entering an era of academic secrecy driven by competitive AI. That's terrible for science.
The Uncomfortable Questions
- Should AI labs be allowed to race to publish on problems they heard about through rumors?
- Is 88 hours of compute-intensive search "research" or just expensive pattern matching?
- What happens to the Millennium Prize process when multiple teams claim solutions within days of each other?
- Can we trust that internal models aren't trained on user data when the incentive to do so is enormous?
Willison doesn't have answers. Neither do I. But the questions matter.
The Bigger Picture
This is the first major academic ethics scandal of the AI era. It won't be the last.
The rules of scientific priority were written for humans who work at human speed. AI changes the tempo. A year of human work can be replicated in days with enough compute. Our institutions haven't adapted.
The Navier-Stokes solution is impressive. The way it happened is a warning.
The race to AI-driven discovery just got its first scandal. We need new rules before the next one.
Top comments (0)