A 23-year-old with no math degree typed an unsolved conjecture into ChatGPT on a Monday afternoon. Eighty minutes later, he had a valid proof that defeated professional mathematicians for sixty years.

Fields Medalist Terence Tao confirmed it. The math world is losing its mind.

The Kid Who Didn’t Know It Was Hard

Liam Price wasn’t trying to make history. He was “vibe mathing” — his term for feeding random open problems into GPT-5.4 Pro to see what happens. He’d been doing it for months with his friend Kevin Barreto, a Cambridge undergrad, pulling unsolved conjectures from erdosproblems.com like lottery tickets.

“I didn’t know what the problem was,” Price told Scientific American. “I was just doing Erdős problems as I do sometimes, giving them to the AI and seeing what it can come up with.”

Problem #1196 had been open since 1968. Stanford PhD students had built entire dissertations around it. Nobody cracked it.

Price cracked it with a single prompt.

What the Problem Actually Is

Paul Erdős — the most prolific mathematician in history — posed this conjecture with collaborators Sárközy and Szemerédi. It concerns primitive sets: collections of whole numbers where no number divides any other.

The conjecture asks whether, as the smallest number in a primitive set grows toward infinity, a specific mathematical sum converges to exactly 1. Technical, yes. But it had resisted every attempt by every serious mathematician for nearly six decades.

The AI Took a Road Nobody Saw

Here’s the part that made Terence Tao sit up.

Every mathematician who attempted this problem started with the same “standard sequence of moves.” It was the obvious approach. It was also a dead end.

“The humans that looked at it just collectively made a slight wrong turn at move one,” Tao explained.

ChatGPT didn’t make that wrong turn. It couldn’t — it doesn’t know which approaches are “standard.” Instead, it pulled a formula from a completely different area of mathematics and applied it in a way no human had considered.

The AI didn’t just find the answer. It found a new way to think about the problem.

Not Perfect — But the Spark Was Real

Let’s not oversell this. Jared Lichtman, the Stanford researcher who’d spent years on this exact problem, said the raw ChatGPT output was “actually quite poor” and “required an expert to kind of sift through and actually understand what it was trying to say.”

The AI didn’t write a clean proof. It produced a messy, promising sketch containing a genuinely novel insight buried in noise. Humans had to extract it, clean it, and formalize it.

But the core idea — the unconventional connection — was something no human had found in sixty years of trying.

The Scoreboard Is Getting Ridiculous

Since January 2026, 15 Erdős problems have moved from “open” to “solved.” Eleven of them credit AI models in the process. Tao maintains a public database tracking every AI contribution.

Not all claims hold up. In October 2025, an OpenAI VP tweeted about a ChatGPT Erdős solution, then quietly deleted it when the “solution” turned out to be a known result. Problem #1196 is different: genuine difficulty, novel method, expert validation.

Tao thinks the new technique could be “the seed of a new mathematical theory.” That’s not just a solved problem — it’s a new tool.

What This Actually Means

This isn’t AI replacing mathematicians. It’s something stranger.

AI brings cognitive diversity — freedom from the groupthink that trapped every human expert into the same dead-end approach. Humans bring judgment, rigor, and the ability to recognize when something matters.

Neither could have solved this alone. Together, they did.

The “vibe math” approach sounds absurd. Feed problems to AI, see what sticks. But it’s working. And it’s democratizing mathematical discovery in ways nobody predicted. You don’t need a PhD anymore. You need curiosity, a subscription, and enough literacy to spot when the AI says something interesting.

The Caveats

The success rate is still low. Price has prompted ChatGPT with hundreds of problems. Most outputs are garbage. The AI doesn’t understand what it’s doing — it generates text that happens to contain novel insights because its training data lets it recombine mathematical structures in unexpected ways.

And as Tao noted: “I think the jury is still out on the long-term significance.”

One solved conjecture doesn’t prove AI can reliably do mathematics. But five years ago, an AI producing a novel proof technique that impressed a Fields Medalist was pure fantasy.

Now it’s Tuesday.

The Real Question

Every expert mathematician made the same wrong turn on this problem for sixty years. A kid with ChatGPT didn’t.

That’s not a fluke. That’s a signal. The question isn’t whether AI will change mathematics — it already has. The question is whether the establishment will embrace it, or keep making the same wrong turn until someone with a $20 subscription proves them wrong again.

Sources: Scientific American, Futurism, ErdősProblems.com