10,000 AI agents and a disputed proof: Brookings on science's absorption problem

What OpenAI announced

According to the essay, OpenAI said on Sept. 8 that an unreleased internal model had produced a proposed solution to the Navier-Stokes existence and smoothness problem, one of the seven Millennium Prize Problems set by the Clay Mathematics Institute in 2000. Roughly 10,000 agents ran in parallel for 88 hours. The compute bill ran to millions of dollars, which OpenAI's Sébastien Bubeck described in a press briefing as roughly a thousand times what the company had previously spent on a single problem.

The proof is not verified. Mathematicians are reading it, and under Clay's rules a proposed solution must appear in a qualifying publication, wait two years and win general acceptance before the prize is considered. OpenAI says it does not intend to claim it.

On Oct. 6, OpenAI released 722 manuscripts in one GitHub drop, claiming progress on 372 open problems in algebra, geometry, number theory and theoretical computer science, all from the same model. The company estimates the average result cost roughly three hours of ChatGPT Pro compute. Only 162 manuscripts came with machine-checkable Lean formalizations. A group calling itself the Association for Human Mathematics called the release "not a demonstration of scholarship, but a demonstration of power." A day later OpenAI withdrew three manuscripts after an internal audit found a sign error that invalidated one argument and the constructions in two papers that depended on it.

Normal science at unprecedented scale

Yaraghi argues that even if the Navier-Stokes proof holds, it fits the pattern he described in July: generative AI is well suited to normal science, meaning puzzle-solving inside a frame people have already built. The problem reached the agent swarm with a clean statement and an unambiguous success criterion. He also notes that Bubeck said OpenAI began working on the Millennium problems because of viral Twitter rumors that Anthropic had resolved two of them, so the agenda was set by competition rather than by a machine choosing an important question.

What changes, he says, is scale. He compares such systems to Galileo's telescope: an instrument that generated new questions faster than anyone could work through them, while the number of astronomers rose. The worry is that this instrument arrives just as the number of qualified people is being cut.

Production scales, absorption does not

The core of the argument is an asymmetry. Producing results is now an industrial process: add agents, compute and hours, and get more output. Understanding a result, by contrast, happens one person at a time and is bounded by human cognition and years of training. Yaraghi cites a Harvard physics trial in which students with a purpose-built tutor learned roughly twice as much as those in an active-learning classroom, in less time. Twice is a real gain, he writes, but it is not 10,000 agents in parallel.

In economic terms, he says, AI substitutes for research labor and complements absorption labor, so the ratio of results produced to results understood rises with every capability improvement.

What mathematicians did with the Erdős result

The essay points to May, when an internal OpenAI model disproved Paul Erdős' unit distance conjecture. Nine mathematicians checked the proof and published a companion paper with a short, human-verified version of the argument and signed reflections. Thomas Bloom said the original proof was completely valid, yet humans improved it substantially because discussing, digesting and drawing out consequences remains their work.

Two reflections worry Yaraghi. Victor Wang noted that verification went smoothly because the digested argument ran only a few pages, and a longer one might have fared differently. Melanie Matchett Wood said that in many cases it will be easier for AI to convince mathematicians it has a proof than to produce a correct argument, and she does not think the profession is prepared for that.

What it means

Yaraghi's conclusion is that three gaps are opening at once: more machine-generated science than qualified evaluators, no institution for verifying machine-generated claims outside pure mathematics, and a training pipeline for evaluators being cut at the wrong moment. He says the policy agenda from his July pieces lacks the instrument that matters most for this, and that the priority should shift toward capacity for verification and absorption.