Four AIs just cracked an unsolved math problem

For the first time, an AI has solved a genuinely unsolved mathematics problem — and then three more AIs solved it independently. Epoch AI just confirmed that GPT-5.4 Pro cracked a Ramsey-style combinatorics problem that only 5–10 specialists worldwide had seriously attempted. Experts estimated it would take a human mathematician 1–3 months to solve.

This isn't AI acing a test or winning a competition. This is AI doing something no person had managed to do yet.

What the AI actually solved

The problem involves hypergraphs — think of them as networks where connections can link three or more things at once, not just pairs. The challenge: build the largest possible hypergraph that avoids a specific mathematical pattern called a "partition."

In plain English, imagine trying to build the biggest possible social network where no group of friends can be neatly split into perfectly matching teams. Mathematicians had found some size limits, but nobody had proven a tighter boundary — until now.

The problem was posed by Will Brian, Associate Professor at UNC Charlotte. It comes from research on whether infinite series (long chains of numbers being added together) can all converge (settle on a value) at the same time.

One AI solved it — then three more followed

Will Brian verified the solution and called it "exciting": "The matching lower and upper bounds are quite good for Ramsey-theoretic problems." The AI's approach eliminated inefficiencies in previous attempts and matched the intricacy of the best known upper-bound constructions.

The fact that four different AI models from three different companies all solved the same problem suggests this isn't a fluke — it's a capability that frontier AI models now reliably possess.

Why this is different from every AI math benchmark before

AI has been acing math competitions for a while — NVIDIA's model recently won gold at three olympiads. But competition problems have known answers. Someone already solved them. The AI just has to find the same path.

Open problems are fundamentally different. Nobody knows the answer. There's no training data to memorize. The AI has to discover something genuinely new.

Epoch AI's FrontierMath benchmark tracks exactly this: a collection of 15 unsolved mathematical problems that have resisted serious attempts by professional mathematicians. Until today, none had been solved by AI. Now one has — and it was solved four times over.

The skeptics aren't wrong to ask questions

On Hacker News, where the story hit 131 points, the reaction was split. Supporters called it "a remarkable result" and said they were "deeply baffled by AI denial at this point."

But skeptics raised valid concerns:

• Human involvement: Researchers Kevin Barreto and Liam Price guided GPT-5.4 Pro using a "scaffold" — a framework of tools, prompts, and auto-critiquing systems around the model. How much did the scaffolding matter vs. the AI's own reasoning?
• Problem difficulty: Experts rated this problem as 95–99% solvable and estimated it would be published in a "standard specialty journal" — important, but not a Fields Medal problem.
• The real test: As one commenter put it: "Can an AI pose a frontier math problem that is of any interest to mathematicians? That would be true PhD-level research."

Where AI math goes from here

This result sits at a fascinating inflection point. AI has graduated from solving problems humans already solved (competitions) to solving problems humans hadn't solved yet (open research). But it hasn't yet started asking new mathematical questions — the kind of creative leap that defines great mathematicians.

Still, 14 of 15 FrontierMath open problems remain unsolved. The problems range across number theory, combinatorics, and algebra — including famous challenges like finding Hadamard matrices (special mathematical grids with useful properties in signal processing) of order 668, and variants of the Inverse Galois problem (a 200-year-old question about the symmetries of polynomial equations).

If AI starts knocking those down too, the conversation shifts from "can AI do math?" to "what can't it do?"

Related Content — Get Started with Easy Claude Code | Free Learning Guides | More AI News