Imagine spending your whole career on one maths puzzle. Then, one Monday morning, you read that an AI may have solved it. That’s roughly what happened to a lot of researchers on 26 October, when OpenAI announced three new results its models had produced. These OpenAI math proofs cover pure maths, computer science and physics.
Big claim, right? So let’s slow down and look at what OpenAI actually said, who’s checking it, and why you might care even if you haven’t touched an equation since school.
So what did OpenAI actually claim?
There are three results, and each one targets a problem experts have chewed on for decades. OpenAI hasn’t named the model behind them. It only says “its models” did the work.
The maths one: waves that can’t stay focused
First up is the local smoothing conjecture, which Christopher Sogge proposed in 1991. In plain words, it’s about how waves behave when they bunch up, and it suggests extreme focusing can’t last for long. According to Interesting Engineering, OpenAI says its models proved it in every spatial dimension. Mathematicians had cracked the two-dimensional case back in 2019.
OpenAI also says the same proof settles two related puzzles, the Kakeya and Stein restriction conjectures, in all dimensions. The model produced the proof on 18 October. Then the team checked it in Lean the next day. Lean is a proof assistant, which is software that checks each step of a maths argument by computer.
The computer science one and the physics one
Next, OpenAI claims its system disproved the Exponential Time Hypothesis, or ETH. That’s a long-held belief about how hard certain logic problems are to solve. The disproof came on 16 October, and it went through Lean a day later. Finally, the physics result is about proving that certain theories of particles and forces in four dimensions actually exist on solid mathematical ground.
Why the OpenAI math proofs are a big deal (if they hold up)
Let’s start with ETH, because it’s the one that made computer scientists spill their coffee. Hundreds of papers in fields like graph theory, computational biology and databases build on it. So if it falls, a lot of “nobody can do this quickly” results need a second look.
MIT’s Ryan Williams told Al Jazeera that ETH has been “widely believed true since the late 1990s.” He also called it “a shocking development” and said it’s “hard to predict what will be possible next.”
The maths result hit closer to home for Alex Iosevich, a mathematician at the University of Rochester. Sogge, who first posed the conjecture, was his PhD adviser. Iosevich called it a “historic development” and said he now feels “filled with joy and optimism.” Yet he also admitted to “a twinge of sadness,” because he expects theoretical maths to “change drastically.”
Then there’s the physics. The claim covers quantum chromodynamics, the theory of how quarks and gluons stick together. It also touches pure Yang-Mills theory, which is part of one of the famous Clay Millennium Prize problems. To be clear, nobody’s claiming the prize itself.
Hold on, has anyone checked this?
Not fully, and that’s the honest headline. As of this week, no outside group has confirmed any of the three results. In fact, Williams put it plainly: “We are working to verify and examine their proof and determine if it holds up.”
OpenAI set up three expert groups, which it calls “shepherds,” to handle each result. For example, Russell Impagliazzo and Ramamohan Paturi of UC San Diego will lead the ETH review with Williams. OpenAI has also agreed to fund a conference on the claim. Meanwhile, a physics group led by physicist David Gross will write its own explanation before releasing any of the AI-made material.
We like that OpenAI admitted the downsides of its own approach. It said that picking the first shepherds “creates asymmetry of information,” since some experts see the work before others. That’s a fair point, and it’s rare to hear a company say it out loud.
Still, some caution makes sense. Al Jazeera notes that when Anthropic announced an enzyme discovery last month, some scientists said the company had exaggerated its significance. So we’d wait for the experts before popping any champagne.
What changes for the rest of us?
Short term? Not much you’ll notice. Williams said there are no short-term real-world applications, though he sees possible “future real-world benefits for society.” Also, refuting ETH wouldn’t automatically give us faster software for existing problems.
But the bigger shift is in how people do science. AI here isn’t writing essays or summarising emails. Instead, it’s producing work that top experts treat as worth serious checking. That’s a different job from the everyday tasks we covered in our guide to large language model use cases.
It also fits a pattern. Recently, we wrote about how AI helped find a decade-old web flaw. So AI is moving from “helpful assistant” to “occasional discoverer,” at least in fields where a machine can check the answers.
For students and young researchers, including here in Armenia, that matters. The skill that rises in value is knowing how to check, explain and build on results like these.
Okay, so should you believe the OpenAI math proofs?
Our honest take: be excited, but not yet convinced. The Lean checks are a strong sign, because a computer has gone through each step. Even so, experts still need to confirm the proofs say what OpenAI claims they say. That’s why the review groups exist.
Also, keep the source in mind. This came from OpenAI’s own announcement, and the same company is busy selling products like GPT-6.1 Terra. That doesn’t make the claims wrong. It just means outside checks matter more.
If even one result survives, though, we’ll look back at this week as a turning point. So here’s the question worth sitting with: if an AI can solve problems that stumped experts for decades, what should people spend their time learning next?