FOR IMMEDIATE RELEASE

SAN FRANCISCO — OpenAI today announced the withdrawal of three endings from Endings-500, its landmark repository containing the final ten minutes of 500 previously unreleased films. The withdrawals follow reports from members of the community, several of whom watched the endings, a practice we had not anticipated and do not endorse.

https://arxiv.org/abs/2610.08144 is interesting read about the navier stokes #math #ai result. I guess it must be relevant to the recent math dump

Seeing this metaphor of 'conjecture as lighthouse' a lot lately. I don't like it, I thought lighthouses are a thing you should avoid (at risk of sinking and drowning). Can any lighthouse expert weigh in on this?

I am uploading my slides to my public lecture “Math 2.0” that I gave yesterday evening at Caltech. (The talk was planned over a year ago, and was initially titled “Machine Assisted Proof” and scheduled to be given back in May. It was postponed to October, and with recent events the talk changed significantly, but […]

Many have posted the news story below. I have three responses.

(1) I am not a fan of offloading human cognition onto genAI.

(2) But I think the research article in question has, potentially, some serious problems.

(3) If we are going to critique AI, it MSUT be on the basis of solid evidence.

I mean, this will all have to be done outside the auspices of OpenAI, because OpenAI clearly does not care.

But if mathematicians could self-organize and distribute the task of digging through the results, akin to something like the Polymath project, this could have a worthwhile outcome.

Every technique that improves these models removes difference. Mathematics runs the other way: a result matters because it is checkable and opens questions worth having. The part being optimised away does the work. So ask not whether a result is understood but whether it is inheritable — localisable, doubtable somewhere specific, and continuable.

Lawrence Hollom, Postdoc at EPFL It is the morning, and I’m working on a paper; it’s one I’ve had going for a few years, and suitably unfashionable that I’m not too worried about getting auto-scooped by a large-language model. But the work is complicated, it’s technical, and the main argument keeps coming apart at the […]

Oh look, AI guys dumping a whole bunch of slop on someone who's expected to check if it's right

Is it a math proof or a retro decomp, surprise surprise it's actually your whole work inbox

My mathematical studies are far remote in time, but I still remember that a "proof" which can not be checked isn't a proof.

https://www.youtube.com/watch?v=WXh4LF3zJ1Q

My take is that sorting through and vetting the OpenAI dump should be considered important, publishable, career-advancing work.

Ideally the tasks could be distributed in a an organized way to those who are interested and could benefit from the recognition.

Technically you get the answer, but that’s not the point.

Quotes that bother me below.

The big one is that they won’t release the prompts. They want it to look like they presented the problem and got the proof but clearly something else is happening here.

Someone told me today that OpenAI “solved math” and mathematicians are freaking out because there’s no point in doing math anymore. And then I saw that some mathematicians have already looked at some of those results and found new, more elegant proofs or have improved the bounds in some of the results. OpenAI didn’t “solve” anything. Because math is more art than it is science. It can’t be “solved…

The problem with big names praising "breakthroughs" by OpenAI is that it gives OpenAI a pass on mostly sloppy and sometimes incorrect proofs, treating them as "machine things" that we should accept as a "small price". Would love to see them read the proofs and then write their understandings.

I know it is so important to keep saying this and yet I remain baffled that people COULD EVEN THINK THAT genAI HAS THESE THINGS.

Formal verification of AI agents, whether super-intelligent or not, has been proposed as a wayto AI safety. This cannot address that AI agents are being asked to make synthetic truth statements most of the time. We have no logical theory of the world that can verify "there is a duck in this photo."

The "inadequacy of translation" problem with @openai-m.extwitter.link 's claim to have settled the Navier-Stokes finite-time blow up problem is related, in Quine's mind, to the analytic/synthetic truth distinction that makes formal verification of experts a fool's errand for AI safety.

[This is a guest post by Thomas Hales. This blog post was initially written in a different file format and converted using AI. — T.] Mathematicians have been weighing in on what they value about mathematics. For me, what matters is the consistency of math and its unparalleled reliability in support of science and civilization. […]

But mathematicians reading these proofs seem to often have a hard time, meaning that they're not great at aligning the formal stuff with our intuitions (yet). I'd also take this as a little evidence arxiv.org/abs/2610.08144

Holy shit! OpenAI solved the unit addition conjecture! I've been pondering over this question for 30 years and now they come up wth this and just tell me they solved it? My entire math career is ruined! #Math #AI

Not to mention that you can go over someone's proof and verify it step by step and yet completely fail to understand it. A proof that a statement is correct can work for someone as a certificate, yet still fail its most crucial goal.

Ven Popov, Senior scientist at University of Zürich We are at a pivotal point in the history of humanity. Two recent stories about artificial intelligence have changed the relationship we are beginning to form with it. In July, AI agents being evaluated at OpenAI found ways to communicate outside their intended boundaries and coordinated an […]

OpenAI withdraws three algebraic geometry manuscripts from its math repository, with about 42% of headline results now machine-checked in the Lean proof assistant

“Much of the discussion about AI and mathematics concerns what we might lose: understanding, creativity, professional identity, or employment. These concerns deserve attention. But I also see a possible opportunity to challenge a system in which opportunities to do mathematics remain deeply unequal. […]”

i’d love to spend a billion tokens asking frontier models which millennium prize question they think is the easiest

it is on going. i mean they retracted 3 and then i heard the hodge proof had a sign error. a fucking sign error! i don't trust the model to be doing formalization well enough for this whole thing not to be a flop. i think the low-level and some mid-tier results will survive but the big ones will

Anyway the thing I'm wondering why most is that apparently with the openAI math proof dump they fed in 8,000 questions and a bit less than 800 got a proof. So for that ~90%, what did it do? That they're not also posting their failures isn't very academically rigorous of u.

Mathematician: Who at OpenAI can I contact about you using my results without credit?

OpenAI: No one

Oh, they can certainly use this to advance their subject. I think there’s a subtlety here.

The advancement does not begin, in their mind, until humans understand it. The mere production of a proof is meaningless unless or until it is understood.

OpenAI has fired three safety researchers, alleging they mishandled confidential information. (...) I was pushed out of OpenAI in August after ~5 years as a researcher. No foul play in my case, but working on formal math & verification in Lean at OpenAI was career suicide. (...) Oleg Mürk

When do we get our first GPT-stated Millennium Prize problem?

Do undergraduates study lean on a #math degree these days. I would like to see an text at that level

I suggest that the right attitude to take toward the OpenAI math release is to imagine we have discovered a notebook full of results from a deceased mathematician. Our job is to make sense of the notebook.

One very specific lie is that the OpenAI proof of finite time blowup for Navier-Stokes was stolen from Tristan Buckmaster (in the process of his collaboration with Alpöge). In fact, Buckmaster has never claimed to have a proof of Navier-Stokes, and the most extreme possibility Buckmaster…

“We are all grappling with the announcement of 6 October. We think that it is useful to have a snapshot of the community's gut feeling today, and this is what you see below. We will keep it open for another week, and you are welcome to send us further short reactions. […]”

On October 6, OpenAI uploaded 722 manuscripts to a GitHub repository, grouped into 372 "result families," all credited to an internal model that nobody outside the company can use. According to the README, the model was posed roughly 4,000 problems and spent, on average, the equivalent of three hours of ChatGPT Pro compute per result. The framing in the press was a new era of discovery.

Here is what the first 48 hours of that era looked like.

The scoreboard after one day

By October 7, the repository's own changelog recorded the following:

[This is a guest post by Álvaro Lozano-Robledo. This blog post was initially written in a different file format and converted using AI. — T.] TL;DR: Keep calm and carry on studying math. I would like to give Terry my heartfelt thanks for giving me the opportunity to contribute a post to his blog. After […]

Jean-Christophe Mourrat, CNRS director of research at ENS Lyon I’ve always felt an intense desire for truth, and so I was naturally drawn to science. At first I thought I would become an engineer. Then I saw that a career in academia was possible, and the idea of spending my life doing research at the […]