OpenAI's AI Breakthroughs Spark Mathematical Debate and Ethical Concerns

Ten thousand agents, eighty-eight hours, one Millennium Prize problem, and a proof Lean checks line by line. So: do you want to win a Nobel Prize? Pick the problem, design the reward, go.

https://davidorban.com/2026/09/do-you-want-to-win-a-nobel-prize/

In my view, If AI can solve a math problem it may mean that the problem is already conceptually understood in terms of how to solve it. What remains is often the tedious work of formal derivation and verification which AI can do much faster.

#science

buff.ly/pFbdIRm

Regarding The Clay Mathematics Institute Millennium Prize Rules. The short answer is no, an AI or non-human entity cannot legally or officially be awarded the prize, but the human creators or users of that AI could potentially be recognized. share.google/aimode/J5OLe...

@kimmonismus

OpenAI is close to solving another math problem from the Millennium Prize Problems, the Hodge Conjecture. Via The Information

For solving the first Millenium Prize Problem "OpenAI used a variant of its next pre-trained model, codename...

https://x.com/i/web/status/2100629040692629572

More details on OpenAI claiming it had solved a famous outstanding math problem: One of the human mathematicians working on it said he had 3 ChatGPT accounts but only opted out of training the LLM with two of them. That means the model may have well ingested his work and used it to solve the problem. Attribution mess is messy. (Story by Davide Castelvecchi, who happens to have a PhD in math himsel…

OpenAI is close to solving the Hodge Conjecture as it tests whether frontier AI models can handle some of mathematics’ hardest unsolved problems.

Read more in today's AI Agenda: https://thein.fo/46vQYm9

Fields medal winner Jacob Tsimerman is now working on # AI safety for OpenAI and giving talk on "Open source game theory" https://www.youtube.com/watch?v=dLQZWMH4u0k # math # mathematics # gametheory

OpenAI said an unreleased internal model resolved Navier-Stokes existence and smoothness.

Source: Last Week in AI

@kimmonismus

Want to see how quickly AI progress can make expert forecasts look outdated?

As late as December 2024, AI researchers’ aggregate forecasts put a 50% chance of AI solving a Millennium Prize problem around 2054, finding and patching a s...

https://x.com/i/web/status/2100391602623799728

I've written a blog post responding to the letter about maths and AI signed by 25 Fields medallists. As with the Leiden Declaration, I didn't sign it, but I agree with much of it and welcome its existence.

gowers.wordpress.com/2026/09/17/w...

Anthropic says Claude translated a proof of Fermat's Last Theorem into Lean.

Lean is a computer program mathematicians use to check proofs.

Let's generalize : this post is groundbreaking proof that you can add an arbitrary amount of LinkedIn bullshit to any existing equation without changing its value.

Corollary is trivial.

(…) This is prompted by a friend informing me this morning of two annoying additional stories that taint the recent public success of AI generated proofs and will hopefully lead the mathematical community to address the question of predation and plagiarism. mathstodon.xyz/@antoinecham... #AI #Math ⏚ Antoine Chambert-Loir (@anto...

I'm not anti-AI but OpenAI solving a Millennium Prize as a marketing stunt is beyond lame. It's theft from human potential and accomplishment.

The other reason I want us using Microsoft AI tools at work rather than OpenAI's or Anthropic's, if I can't stop the AI use entirely, is because as we've seen with the recent Navier–Stokes Millennium Prize kerfuffle, OpenAI and Anthropic will definitely use any data you give them to train their […]

Nihar Shah did a heroic experiment for TMLR: he spent 20-25 hours over two weeks interviewing authors of seemingly low-quality submissions about their own papers.

He confirmed what we all suspected: people submitting these papers have no idea what is going on in them.

It would be helpful and it would accelerate progress, because open questions tend to accrete conditional theorems (this is true if Riemann is true, e.g.).

It would certainly spoil some fun, but idk if it would do so more or less than generating a proof does!

A place to read up on AI in mathematics, understand new perspectives, and develop answers to the challenges we face. https:// ai.math.ms/

one idea in it that stuck with me is that the purpose of a proof is not to determine if a statement is true or false

it's to break the statement down into a collection of substatements, joined by argument, which can be mapped onto existing mathematical knowledge

There really isn't any reason that proofs of theorems must be simple enough for a single human (or even the entire mathematics community) to understand, at some point we may actually need ai and some measure of faith in formal verification to progress.

I don’t know that I’m getting across the scale of a millennium prize problem if it’s getting compared to writing a paper.

I mean by this metric I guess everything is on a continuum so nothing is any better than anything else, even solving one of the canonical millennium prize unsolved math problems. “Some math problem” indeed.

speaking to an applied mathematician friend at work, the OpenAI NS proof is non-constructive, so while the counter-example has been proven to exist, it's actually non-trivial to simulate it.

I wonder if porting it (via dedukti?) to Agda might be a valuable thing to do?

The blog @ proofsandprompts is also on mastodon now. It's a community project about # AI and # math .

A genAI math proof might actually be correct, it might actually do what it said it could do and not just be stolen from a human mathematician.

What it doesn't do, and can never do, is create a mental model in a human's head that makes them go "holy shit this is the same as ___"

Formal verification is often treated as a finish line: prove the code once, then trust it forever. That is not what it gives you. It proves that one version of a program matches one version of a specification.

Programs change, and so do their goals.

Anthropic's solution to Fermat's Last Theorem is a constructive proof; the AI generated 29,510 "sub proofs," each one of which has to hold, for its conclusion to Fermat's Last Theorem to be true. Having grad students confirm each step is a chore, but it will have to be done to validate it.

In a list of absolutely remarkable things, what stands out to me is this: Zhang's proof was so carefully and beautifully written that it could be read, understood, vetted, reproduced and published in five weeks. Most modern proofs take a year or more.

Compare that to the recent AI press.

And to add to that, this is mathematics. A proof in Lean is extremely strong evidence, which is something we don't have in, say, medicine or experimental physics.

If someone claimed they cured cancer or solved cold fusion, "wait for consensus" is reasonable. But far less so in (formal) math.

@nature.com's take on last week's events in mathematics — and the controversy over OpenAI's announcement that it had solved the Millennium Prize Problem on the Navier-Stokes equations www.nature.com/articles/d41...

Its solved a Millennium prize, so depends how you categorise productivity in mathematical discovery. I agree that diffusion is still in early stages, but its phenomenally powerful and useful. And the emissions aren't very high despite weird accounting. "Russia's state-company coal" isn't very big.

OpenAI says an internal model just solved a 90-year-old math problem. Mathematicians are not convinced yet.

The company claims an unreleased model, coordinating roughly 10,000 sub-agents over 88 hours, proved that the Navier-Stokes equations (one of seven Millennium Prize Problems) break down un...

Proofs and Prompts is a community blog about Math and AI. Since we started it a month ago, 42 blogposts have appeared, covering many aspects of math and AI. You can follow this mathstodon account to get notified when there are new posts. We also wrote a post about our experiences running the blog: https:// proofsandprompts.com/2026/09/1 6/one-month-of-proofs-and-prompts/ # introduction # proofsand…

Twenty-five Fields Medalists just signed a letter about how one AI lab treats outside researchers.

Two mathematicians say OpenAI learned of their approach to a Millennium Prize problem, then put a research team and massive compute behind it to claim it first.

As a vaguely fluid mechanics person, it's fine. It's not the kind of equation for which there's "a" solution, the Millennium Prize question was never about anything of practical significance. The intellectual value was in being able to prove it.

"A proof is not a solver" bsky.app/profile/larg...

Listen Morty—Anthropic's Claude just formalized Wiles' Fermat proof in eleven days. Thirteen million lines of Lean. Humans budgeted a decade. Three centuries of margin-note drama, then a chat box notarizes the homework. What a waste of blackboard real estate. day259.070

Really cool how all it takes to solve Millennium Prize Problems is a Python loop and a chatbot

OpenAI says it solved Navier-Stokes. Nobody has seen the proof. thenextweb.com/news/openai-...