A verified proof is not an understood proof
According to Scientific American, mathematicians from several expert groups challenge OpenAI’s claims. Gonzalo Cão-Labora of EPFL says a close examination reveals a different approach, with similarities to work by Córdoba and Martínez-Soroa only at the surface.
The questions around the work began before publication. New York University mathematician Tristan Buckmaster accused OpenAI of relying on results from his work and work by Anthropic’s Levent Alpöge, produced with publicly available models. OpenAI denied this and said it had not used Buckmaster’s data.
Camillo De Lellis of the Institute for Advanced Study says the truth may be impossible to establish conclusively: language models operate as “black boxes.”
The dispute is unfolding amid a much larger stream of results. In early October, OpenAI published around 400 more, in an unsystematic collection on GitHub. Many mathematicians were angered, arguing that this way of publishing devalued their field. The volume makes independent scrutiny harder, not less necessary.
Ten thousand agents, unclear contribution
For the Navier–Stokes problem, OpenAI used around 10,000 instances of a model for about 88 hours and 130 billion tokens. OpenAI researcher Noam Brown says nobody knows how much this large swarm of AI agents contributed to the result.
There are no systematic comparisons with thousands of agents: those tests are too expensive. Nobody checked how long one agent would take to solve the same problem, and it is almost impossible to identify the specific advantage of 10,000 agents over 1,000.
Brown estimates the multi-agent setup accounts for less than 10% of the success. In his view, the base model matters more: it can work on tasks for a long time and operate in parallel.
That makes the swarm a striking part of the story, but not necessarily the important one. My guess is that the harder question for OpenAI is not how many agents it can run, but whether it can explain what they added. Without comparisons, the scale is a resource figure, not evidence that more agents made the proof possible.
Daily AI news
Every day we pick what actually matters in AI and explain it plainly — no hype, no filler. Subscribe if you want to follow where the industry is going.
Only what matters — every day
Follow on X