Scott Aaronson, the theoretical computer scientist at the University of Texas at Austin, wrote something in a September 15 blog post that his 2006 self would never have said: "The Singularity has already started; it's just wildly unevenly distributed." He still unloads the dishwasher and clips his toenails, but he no longer expects to ever prove a theorem because he is needed to.
The essay, titled "The Age of Wonders and Terrors," arrives after a week of seismic news in mathematics. On September 8, OpenAI announced that an internal multi-agent system had found a construction in which the three-dimensional Navier-Stokes equations develop a singularity in finite time — one of the Clay Mathematics Institute's seven Millennium Prize Problems, open for roughly ninety years — and released a 166-page writeup plus a Lean formalization. OpenAI says it will not claim the $1 million prize, and that whether any human is eligible remains unclear. Aaronson estimates the effort burned at least $15 million in compute, and that as of his writing, "probably no human has yet read and understood" the proof.
The controversy is as loud as the achievement. OpenAI's announcement came about twelve hours after a statement from NYU mathematician Tristan Buckmaster and Anthropic researcher Levent Alpöge, who said they had resolved closely related problems first using a similar approach. Aaronson, citing Quanta's reporting, notes the two accounts disagree, in tangled fashion, about what happened and what ought to have happened. OpenAI insists its researchers and agents never saw the other team's work before public release, while conceding it cannot rule out that de-identified product data helped improve its models. What is not in dispute: both efforts built on a method pioneered by human mathematicians Diego Córdoba and Luis Martínez-Zoroa. Aaronson declines to adjudicate the priority dispute, but leaves a question hanging: "What exactly did OpenAI know, and when did it know it?"
What unsettled the math community is the list. Aaronson catalogs long-standing open problems solved or assisted by AI in the last month alone: a counterexample to the Jacobian conjecture; a Lean-verified proof of Fermat's Last Theorem (completed by Anthropic last week); a QMA/QMA(2) oracle separation and a proof of Watrous's disentangler conjecture, a problem he helped popularize back in 2007. Nearly every paper on arXiv that would interest him, he says, now carries an "AI statement" near the acknowledgments, ranging from "our main result came entirely from GPT-6, but we understood it and take responsibility for it" to "the author did not use AI for anything."
He also relays, clearly labeled as rumor, that AI companies are now sitting on solutions to "very major problems" after the hostile public reaction to the Navier-Stokes episode. Ethan Mollick, a Wharton associate professor, wrote that he does not usually spread rumors but judged the tip credible, coming from someone with inside information. NYU mathematician Scott Armstrong went further, claiming OpenAI has withheld "hundreds" of proofs since at least the International Congress of Mathematicians in July. These claims remain unverified.
The personal notes are the most striking. Aaronson says he feels it in the pit of his stomach when he puts his children to bed: what sort of future can they possibly have? His 13-year-old daughter joked that if she wants to become a mathematician, "it now looks like she has maybe two more weeks." He compares editors and program committee chairs bracing against the incoming flood of papers to the men of Gondor fortifying their city against the orc army — reviewing "will have to be done partly by AI, because otherwise there's no way to handle the orc army."
The essay offers no answer. It is a testimony: the most cautious skeptic has updated. Once "the Singularity has started" moves from a subcultural slogan to a mainstream scholar's judgment, the argument shifts from whether to at what cost. The first bill for that cost may be sitting in the withheld proofs.
[1][2]