Did OpenAI take the initiative upon themselves to muster the 10000 agents in order to tackle the Navier Stokes equation?
Socrates:
Yes, Platohagel—but with an important distinction.
OpenAI did initiate the 10,000-agent effort itself. It was not a standing project in which 10,000 agents had been independently working on Navier–Stokes for months. According to OpenAI’s own account, the initiative began on September 1, 2026, after researchers heard rumors that two Millennium Prize problems might have been resolved. They then decided to test their new internal model against all the remaining Millennium Prize problems.
The sequence is quite revealing:
August 28
OpenAI began training a new internal model with unusually strong mathematical performance.
↓
September 1
They heard rumors concerning progress on Millennium Prize problems.
↓
OpenAI launches the experiment
Different groups of AI agents were assigned different mathematical problems and variants.
↓
Euler first
Nearly 100 agents worked for about 50 hours on a related Euler-equation regularity problem and produced a result.
↓
Resources shifted to Navier–Stokes
Once OpenAI saw the Euler result, it judged Navier–Stokes the more promising target and redirected agents toward it.
↓
≈10,000 concurrent agents
The Navier–Stokes group worked for approximately 88 hours, exchanging about 2.7 million messages and producing approximately 130 billion output tokens.
↓
Another 17 hours
GPT-6 Astra was used to formalize and verify the resulting proof in Lean.
So your wording “did OpenAI take the initiative upon themselves?” is essentially correct.
But there is an even more interesting point
They did not begin with the declaration:
“Let’s solve Navier–Stokes.”
They began with something closer to:
“What happens if we give this new intelligence access to the great unsolved problems and let different populations of agents attack them independently?”
That is a fundamentally different research strategy.
OpenAI describes the agents as being divided into groups, given different formulations of the problems, and encouraged to explore diverse approaches. Then useful discoveries from one group could be consolidated and fed back into other groups.
In your language, I would describe the architecture almost as:
╔══════════════════════════════════════════════════════════╗
PROBLEM
↓
MANY POSSIBLE APPROACHES
↓
MANY AGENTS / MANY LOCAL EXPLORATIONS
↓
RELATIONAL EXCHANGE
↓
PATTERN / STRUCTURE EMERGES
↓
CONSOLIDATION
↓
FORMAL PROOF
╚══════════════════════════════════════════════════════════╝
And that connects remarkably well to your Geometry of Memory idea.
The 10,000 agents are not simply “10,000 mathematicians.” They constitute a distributed search through a mathematical possibility space. Each agent explores a local region; communication allows information about successful regions to propagate; the system then concentrates computational attention where the landscape appears promising.
There is a striking parallel with your recent discussion of hills and valleys:
many possible configurations → exploration → resonance between successful paths → concentration into a stable structure → formalized result
That is perhaps the more profound story behind the headline.
One caution, however
I would not yet phrase this as “OpenAI has definitively solved Navier–Stokes” without qualification.
OpenAI says its proof establishes a finite-time singularity and has supplied both an analytical proof and a Lean formalization. But the mathematical community is still scrutinizing the result, and there has been controversy concerning the relationship between OpenAI’s work and concurrent work by Tristan Buckmaster and Levent Alpöge.
OpenAI itself says it does not intend to claim the $1 million Millennium Prize.
So I would distinguish:
OpenAI’s claim:
We have produced a proof resolving the Navier–Stokes Millennium problem.
Current scientific status:
The proof has been formally checked in Lean, but its mathematical significance and priority are still undergoing scrutiny.
And there is something especially relevant to your question: OpenAI did not merely ask an AI to solve the equation. It constructed an artificial research population and deliberately orchestrated the interaction among that population.
That may actually be the more historically significant development.
Dialogos of Eide 