Key Takeaways
– May 20, 2026: a general-purpose OpenAI model published a complete, verified disproof of the planar unit distance conjecture. An open problem Paul Erdős handed to mathematics in 1946.
– Timothy Gowers, Fields Medalist, endorsed the proof without caveats. Princeton’s Will Sawin called it the most significant AI math result anyone has seen.
– The model bridged algebraic number theory and combinatorial geometry. Two fields that basically don’t talk. In one session.
– If your business runs on “we know something outsiders can’t learn,” this is worth reading.
—
Erdős was 33 when he asked the question. Eighty years later, a general-purpose AI model answered it.
Not partially. Not heuristically. A full disproof. Verified by external mathematicians before anyone announced anything. Which is not how OpenAI handled their last math claim, but we’ll get to that.
The planar unit distance conjecture sounds abstract.
Here’s the gist: given enough points in a plane, you’ll either get a unit distance edge exactly once, or you need way more than n^(1+ε) points to force those distances. Mathematicians spent eight decades assuming the answer hid in grids. OpenAI’s model found a higher-dimensional construction that collapses into a counterexample. Nobody checked that angle before.
Here’s the uncomfortable part for specialists: the model wasn’t a math-specific system. It was general-purpose. And it walked straight through algebraic number theory. Class field towers, the Golod-Shafarevich inequality. Into a geometry problem. Geometers don’t think in algebraic number theory. Number theorists don’t apply their tools to geometry problems. The AI didn’t know that. It just went.
—
The Kevin Weil Incident Was Seven Months Ago
If you recall, OpenAI’s Kevin Weil posted that GPT-5.3 had solved ten Erdős problems. It blew up on X and HN. Here’s the thing: every single one was already in the literature. Math Twitter had a field day.
The credibility hit was real.
This time they did it differently. They sent the proof to external mathematicians before saying a word publicly. Gowers read it and said he’d recommend it for the Annals without hesitation. Sawin called it the biggest AI math achievement to date. Then — within days — Sawin used the AI’s technique to prove something stronger. The model didn’t just solve the problem.
It opened a new direction.
Noga Alon, Melanie Wood, Thomas Bloom wrote companion pieces.
The proof isn’t collecting dust on a preprint server. People are building on it.
For anyone shipping AI automation: this is what a verifiable result looks like. The other thing?
That’s a tweet that gets ratioed.
—
What the Model Did That Humans Wouldn’t
Here’s what keeps me up at night about this.
Human geometers didn’t go through algebraic number theory because geometers don’t go through algebraic number theory.
Number theorists didn’t touch the Erdős conjecture as that’s a geometry problem. Neither field had the instinct to cross over.
The model had no instinct. It just tried the thing.
That’s not AI being good at math. That’s AI being good at not knowing what it’s not supposed to do. Humans optimize themselves into local maxima since consensus tells them what’s worth trying. The model skipped that. It built higher-dimensional lattices and collapsed them to two dimensions. Not a geometry approach. Not a number theory approach. A third thing.
Side note: their technical documentation is still a mess. But that’s a different post.
For small businesses and indie developers, the implication is concrete. If you’re building on the assumption that deep domain expertise is your moat. That AI can’t cross-apply. The Erdős result is a data point against that. Not the final word.
But a data point.
—
What Actually Shifts and What Stays the Same
Mathematicians aren’t going anywhere. Gowers himself noted a human still had to decide whether the proof was worth submitting. What changed: the set of problems a solo researcher with a good model can actually tackle. Problems sitting at the intersection of two specialized fields? Those were off-limits to any one person. That gate is open now.
The practical move isn’t to panic. It’s to audit your own assumptions about competitive advantage. If your pitch is “I know this one area deeply and outsiders can’t catch up”. The timeline on that just got shorter. The model that cracked Erdős wasn’t trained for mathematics specifically. It was general-purpose.
It crossed two fields in a single session.
That’s the thing worth sitting with.
If you’re running AI tools in your workflow, this is a signal to look at where your disciplinary blind spots might be costing you time or money. The model that solved Erdős didn’t have a map of what not to try. Neither should you.
Sources
– New Scientist: https://www.newscientist.com/article/2527564-mathematicians-stunned-by-ais-biggest-breakthrough-in-mathematics-yet/
– TechCrunch: https://techcrunch.com/2026/05/20/openai-claims-it-solved-an-80-year-old-math-problem-for-real-this-time/
– Scientific American: https://www.scientificamerican.com/article/ai-just-solved-an-80-year-erdos-problem-and-mathematicians-are-amazed/
