Scott Aaronson says he's been told AI labs are sitting on solutions to major problems.
The same post lists seven results he credits to AI in the last month, and four of them sit on problems he has worked on himself.
Scott Aaronson published an essay called The Age of Wonders and Terrors on his blog on September 15, 2026. Most of it is about mathematics that machines did over the past few weeks. One paragraph is about mathematics that got done and then went in a drawer.
Rumors have reached him, he writes, of solutions to "some very longstanding open problems in theoretical computer science" that the AI companies are holding back. He doesn't name a company or a problem. He does say the rumored solutions leave out P≠NP and the other complexity class separations, which is the first thing anyone in his field would ask about.
I'm told that the AI companies, having been burned by the hostile response to the Navier-Stokes proof, are now sitting on solutions to some very major problems until they figure out a better way to handle things
- Who
- Scott Aaronson, who writes the blog Shtetl-Optimized
- Says
- he's told the AI companies are sitting on solutions
- Not included
- P≠NP or other complexity class separations
- The reason he gives
- the hostile response to the Navier-Stokes proof
- Also in the essay
- seven AI-proved or -assisted results, from the last month
What he says AI proved last month
The essay carries a list of seven results, and Aaronson says he kept it to "longstanding open problems I had previously known or cared about". Four sit close to his own work: one conjecture carries his name, and he introduced, popularized or studied the other three. He introduced shadow tomography in 2017 (and asked then how far the dependence on dimension could be pushed), and one of the seven pushes it further.
In the areas he knows best, he writes, longstanding open problems are now falling by the day. The first item on his list is a counterexample to the Jacobian conjecture, and it arrived as a post on X in July rather than as a paper.
hello there the jacobian conjecture is false thanx to my close friend akhil for asking about it and my other close friend fable for working during the world cup final ((1+xy)^3 z + y^2 (1+xy) (4+3xy), y + 3 x (1+xy)^2 z + 3 x y^2 (4+3xy), 2 x - 3 x^2 y - x^3 z): \C^3\to \C^3, has jacobian determinant -2, and sends (0, 0, -1/4), (1, -3/2, 13/2), and (-1, 3/2, 13/2) to (-1/4, 0, 0)
Another item is a proof of Fermat's Last Theorem in Lean, a language that checks a proof line by line. Anthropic's own page for that one, dated September 4, 2026, says Claude "worked largely autonomously over 11 days", wrote 13 million lines of Lean and proved 29,500 intermediate theorems on the way.
Navier-Stokes is the reason he gives
Aaronson's account of the Navier-Stokes proof runs at the top of the essay. He writes that the Millennium Problem appears to be solved, "with crucial contributions from both humans and AI", and that OpenAI "burned at least ~$15 million in compute to produce its 166-page solution". Probably nobody has read and understood the thing yet, he adds.
Then came the fight about credit. Aaronson writes that he doesn't find it plausible that OpenAI's models "meaningfully benefitted from being trained on Buckmaster and Alpöge's chat logs", and he still wants an answer to this: "what exactly did OpenAI know about Buckmaster and Alpöge's work and when did it know it?"
human mathematicians are forevermore dethroned as the main theorem-proving entities on planet earth
The new rules for a paper with your name on it
The part I keep going back to is about paperwork, of all things. Pretty much every paper he'd be interested in now carries an AI statement near the acknowledgments (he wishes he didn't have to scroll to the end for it, since it's often the thing he wants to know first), and the statements run from work that came out of an interaction between the human authors and AI down to "the author did not use AI for anything".
So he asks what the rules are now for putting your name on a paper. Understanding the proof, giving talks about it, answering questions about it, and then the one that bites: "Do you need to have played any role in finding the proof?" I think that's the sentence working mathematicians will argue about this week.
He ended his list by asking readers to remind him of anything he'd left out. In his own comments under the post, just after midnight on September 16, 2026, he wrote that "OpenAI has since issued a categorical denial" that the chat logs could have reached its model.