The experts interviewed by Scientific American didn't seem terribly impressed, and even commented on AI's inability to create new theory. That's the kind of thing you'd expect to see from creativity, no?
OpenAI blogs about a new advisory group for math and AI. On their new internal model:
"In addition to resolving the Navier–Stokes Millennium Prize problem, this model has now resolved more than 100 long-standing open problems acros... https://www.manton.org/2026/09/21/openai-blogs-about-a-new.html
OpenAI recently claimed to solve one of math’s biggest questions: the Navier-Stokes problem. But now three mathematicians have showed OpenAI’s work dodges the question rather than solving it.
What has actually been checked?
Exact Mathematics records literature search, human review, formal checking, text–formal-source correspondence, explanatory material, and public release separately.
exactmathematics.org/verification/
Official account for Exact Mathematics with AI — an independent research initiative for mathematical discovery, formal verification, and human understanding.
Updates will be posted selectively.
exactmathematics.org
Making the human that can tell whether a math proof is any good takes years of costly, slow work. Universities paid for that space and time with teaching and/or research grants. AI has repriced that pairing in ways the university and academics must face. kylesaunders.substack.com/p/the-mathem...
Seeing a Millennium Prize Problem get "solved" by AI isn't doing a lot to help my general sense of disillusionment.
It does make me wonder if I dodged a bullet by not entering academic maths, if it was just going to get tech-bro'd.
A AI-happy mathematician trawling literature to feed into the bot to write a "report" on errors has put Atiyah's late-years preprint with a flawed, claimed outline proof of RH in and written a "paper" spelling out in painful detail what the error is.
I can't even.
The scientific debate over AI in mathematics. ~ Oswaldo Royett. https://www. oswaldosrm.com/single-post/the -scientific-debate-over-ai-in-mathematics # AI4Math
And don't even get me started on who has the advantageous passport(s) for research & conference travel v who needs to apply for visas months in advance only to get denied despite ample proof of intent.
I think I saw the thread that prompted this question. Related to that, part of the issue is that "the proof of a theorem" has to be symbol manipulation, but is not necessarily what we as mathematicians value in a paper. It has proven to be easy to train an LLM against, however...
I don't like the implication that the anti-AI crowd does not include scientists
Even if proof checking is a part of maths (it’s not, in any mainstream way) or will be (which I don’t particularly believe), it does not give the output. If we are interpreting “output” in the most restrictive way (and we shouldn’t), the output of maths is a paper or a book or a talk, not Lean code.
This is what AI companies (and their fanboys, including in academia) are trying to convince the world of. Proof checking is just proof checking (which is indeed symbol manipulation), but the rest of the maths is missing from the picture, which is what I assume @barbarafantechi.bsky.social is saying.
there's a real tension and puzzle in the essay about the relation between understanding and mathematical culture and discourse. Harris wisely stops short of saying we cannot say what it is in language, instead going with "it is hard to be rationally discursive about mathematical understanding"
#AI will end humanity: Where’s the proof? | Ed Zitron www.youtube.com/watch?v=Svwm...
If you had told me six years ago that it would soon be possible to spend $10 million to get an answer to a Millennium Prize problem, I would never have believed you. It is hard to overstate how remarkable that that is.
one of the most annoying things about the Navier-Stokes result is that what should have been a straightforward, undeniable win for the "LLMs can create new knowledge" case is tainted
cannot believe OpenAI managed to fuck up the monumental achievement of "our AI solved a Millennium Prize Problem"
Are there authors out there using AI and getting published? Yes. Are all of them caught? No. However, I think it's a dangerous, irresponsible, and reckless accusation to make when there is little or no proof, especially when the author's only fault is that their work didn't click with someone.
The recent formalized proofs of famous mathematical problems, carried out by entities which in the discussions are sometimes referred to as the devil, remind me of the following soviet short film called 'The mathematician and the devil'. Unfortunately I was unable to find a version with good English subtitles, maybe I will get around to create some myself. Until then, the generated ones here are n…
The future of the human mathematician anakedrhymejam.substack.com/p/the-future-o… #AI #math (No easy answers. An interesting read)
OpenAI is close to solving the Hodge Conjecture, another of the seven Millennium Prize Problems, according to a person with knowledge of the work.
Read more: https://thein.fo/46vQYm9
Formal verification still pushes the burden back onto the specification itself. If an agent generates the spec and the implementation together, it just verifies its own misunderstandings.
A proof checks against a fixed standard. An explanation is scored by whether someone else can now go on. Those are different referees, so one ledger imports the proof standard by default.
Grant Sanderson's (3blue1brown) guest post on Terry Tao's blog pretty much affirms and expands on my opinion of what's missing in maths academia: more maths communication overall, and especially towards different audiences.
Two responses to "A severe misalignment of AI in mathematics". ~ Timothy Nguyen, M. Levent Doğan. https:// proofsandprompts.com/2026/09/1 8/two-responses-to-a-severe-misalignment-of-ai-in-mathematics/ # AI4Math
Mathematicians Hate AI. They Can’t Quit It https:// fed.brid.gy/r/https://www.wire d.com/story/mathematicians-cant-quit-ai/
AI Use in Math Research 25.25% of Arxiv math papers acknowledged AI use in Aug. 2026. We use GPT-5.6 Sol to detect AI use acknowledgments and classify AI use cases, and Claude Fable 5 as a validator. The use cases can be divided into research use, which includes cases where AI may have contributed ideas or proofs, and non-research use, including writing, coding, and literature review. https:// epo…
ok but Tristan Buckmaster of NYU was the one who solved it first, with help from Anthropic employee Levent Alpoge, GPT 5.6 and Claude Opus for LEAN proof formalization
"The proof [...] was obtained in a conversation with ChatGPT-6 Astra on Tuesday, September 15, 2026 at 1:02 AM PDT. We then prepared this manuscript for public release, with the intent of uploading it on the morning of Thursday, September 17, 2026"
Shameless, and shameful. arxiv.org/abs/2609.207...
Agreed! But I think in the case of the abc conjecture, verifying that the conjecture was stated correctly in Lean wouldn't be that hard. Not nearly as hard as the amount of effort the community put into trying to understand the purported proof.
You ever think that someone actually completed some millennium prize level problem and thought “nope. Too ugly. Too embarrassing.”
Makes me chuckle.
The social scene around AI and AI safety is more unpleasant than I realized. I'm sure there are other sides to it. But when you get a lot of men with too much money who think they're saving the world, surrounded by groupies, what do you expect? Read the long Bluesky thread about this, collected on one page: https:// skywriter.blue/@segyges.bsky.s ocial/3mvom4b4dn22q I can't confirm thi…
OpenAI spending significant computation on cracking millennium prize mathematical problems sounds impressive, but I just don't care that much? It cost a lot of money. It's a narrow problem. No messy data. Relatively simple to verify. 1/2
In 2004, two mathematicians hypothesized a powerful kind of sandwich. A new proof has finally built it. www.quantamagazine.org/mathematicia...
"The five that remain are: the Birch and Swinnerton-Dyer Conjecture, the Hodge Conjecture, P versus NP, the Riemann Hypothesis, and Yang-Mills existence and mass gap." # math # bigproblems # AIbait
had a dream that someone used the axiom of choice to prove that a mile equals a kilometer. I don't remember the proof but dream!me found it very convincing
I just submitted my abstract for the # AI and the Future of # Math seminar: Title: You are right, and I apologize. Abstract: Doing research changes the researcher. In this talk I will try to document my personal story with AI. I love computers and have been an early adopter of AI with a mostly positive outlook onto the future. Recently the picture has become a lot more grim. This has two reasons.…
Even odds no other millennium prize is solved in September. If announced, has to stand expert review up till 10/31.