For nearly eight decades, the sharpest mathematical minds on the planet stared at a deceptively simple geometry puzzle and came up empty. Then an AI chatbot cracked it in two weeks.

On May 20, OpenAI announced that one of its internal reasoning models had autonomously disproved a famous conjecture posed by legendary Hungarian mathematician Paul Erdős in 1946. The result has been independently verified by a panel including Fields medalist Tim Gowers — and it’s being called the first time AI has independently solved a prominent open problem in active mathematical research.

Not a benchmark improvement. Not a test score. Genuinely new mathematics.

The Puzzle That Stumped Everyone

The planar unit distance problem sounds deceptively simple. Place dots on a flat page. How many pairs can be exactly one unit apart?

Line up nine dots in a row: eight pairs. Arrange them in a 3×3 grid: twelve pairs. Erdős asked the big question — as you scale to millions of dots, what’s the theoretical maximum?

He proposed an answer using carefully calibrated square grids with precise spacing. The number of pairs would grow at a rate described by n^(1 + C/log log n), where that exponent term slowly shrinks toward zero. Erdős conjectured this was essentially the ceiling.

For 80 years, nobody beat it. Nobody proved him right, either. The problem sat in mathematical limbo — one of the most famous open questions in combinatorial geometry.

The AI Didn’t Nibble. It Demolished.

OpenAI’s model didn’t find a marginal improvement. It blew past Erdős’s limit entirely.

The AI discovered a completely new family of constructions producing more unit-distance pairs than any grid-based arrangement. It proved that for infinitely many values of n, you can build configurations with at least n^(1+δ) pairs, where δ is a fixed positive number — not something that shrinks to zero. Princeton’s Will Sawin has since refined δ to at least 0.014.

In mathematics, the difference between “shrinks to zero” and “stays fixed above zero” is the difference between a conjecture being true and being spectacularly wrong.

The method is what has mathematicians buzzing. Rather than working with dots on a flat page, the AI constructed elaborate lattices in higher dimensions with special algebraic symmetries, then developed a technique to project them down to 2D — preserving an unusually high number of unit-distance pairs.

“It feels like magic,” said OpenAI mathematician Mehtaab Sawhney. “It’s kind of an amazing experience to have a machine give back something which really resembles how I work.”

Why This Time It’s Real

Skepticism here is earned. Last year, former OpenAI VP Kevin Weil claimed on X that GPT-5 had solved multiple Erdős problems — then deleted the post after researchers including Yann LeCun and Demis Hassabis pointed out the model had merely rediscovered published solutions.

This time, OpenAI brought receipts. External mathematicians verified the proof independently. Thomas Bloom, who maintains the Erdős Problems website and had publicly criticized OpenAI’s previous claims, co-authored a companion paper validating the result. Gowers, Sawin, Daniel Litt, Jacob Tsimerman, and Noga Alon all weighed in with detailed commentary.

“No previous AI-generated proof has come close” to publishable mathematics, Gowers wrote. Litt called it “the unique interesting result produced autonomously by AI so far.”

The AI’s Unfair Advantage: It Doesn’t Get Bored

The most fascinating insight isn’t what the AI found — it’s how.

Human mathematicians, trusting Erdős’s intuition, spent decades trying to prove the conjecture rather than disprove it. The few who hunted for counterexamples wouldn’t have pursued such a tedious path — constructing elaborate higher-dimensional objects — without some hint it would pay off.

The AI had no such hesitation. It ground through hundreds of pages of careful logic, exploring paths any human would have dismissed as not worth the time.

“AIs have an edge,” noted Jacob Tsimerman. “They can play for longer and in more treacherous waters than mathematicians” without getting discouraged.

This is where AI’s mathematical edge actually lives right now. Not “smarter” than humans in any traditional sense — but capable of maintaining coherent reasoning across extremely long chains of logic while staying open to paths human intuition would reject. Tireless pattern-matching meets bottomless persistence.

The Honest Caveats

Before the champagne, some nuance.

The AI disproved the conjecture but didn’t solve the unit distance problem. The broader question — what’s the actual maximum growth rate? — remains open. The best upper bound, O(n^(4/3)), hasn’t moved since 1984.

And while OpenAI calls the proof “autonomous,” humans were involved. Sawhney and Sellke formulated the query. Bloom noted that “while the original proof produced by AI was completely valid, it was significantly improved by the human researchers at OpenAI and the many other mathematicians involved.”

There’s also the timing. OpenAI is prepping for an IPO, and a splashy math breakthrough makes excellent pre-roadshow press. Doesn’t invalidate the result — but it contextualizes the breathless announcement.

What This Actually Changes

Despite the caveats, this is a genuine inflection point. We’ve gone from AI systems that pass math exams to one that produced original mathematics which surprised the world’s best mathematicians. That’s a qualitative leap.

The question: is this a lucky strike in a domain that suited the model’s strengths, or the beginning of AI routinely contributing to frontier research?

If general-purpose reasoning models can crack problems in discrete geometry, what happens when they’re pointed at protein folding, materials science, or climate modeling?

We’re watching the birth of something new — scientific collaboration where the most productive partner might be a machine that never gets tired, never gets discouraged, and never dismisses an idea as “probably not worth exploring.”

The cathedral of mathematics just got a new architect. It doesn’t sleep.