OpenAI's AI Disproved an 80-Year Erdős Conjecture (2026)

[Alex]: Welcome back to the Nerd Level Tech AI Cast, the podcast where we go full geek on the stories that are quietly reshaping the world. I’m Alex, your resident explainer-in-chief.

[Jamie]: And I’m Jamie, your curious co-host—here to make sure Alex doesn’t go so deep into math land that we all need a rescue rope. [laughs] Today, we are talking about something that sounds like a sci-fi headline: OpenAI’s AI just disproved an 80-year-old math conjecture. And not just any conjecture—the Erdős unit distance conjecture. Alex, did I say that right?

[Alex]: Nailed it! Erdős, as in Paul Erdős, the legendary mathematician who basically collected unsolved math problems like Pokémon cards. And this “unit distance” problem? It’s one of those beautiful, “so simple a kid can understand, so hard no mathematician could solve” questions.

[Jamie]: Okay, so give it to me like I’m five. What was this problem, exactly?

[Alex]: Picture this: you’ve got a bunch of dots—let’s say on a napkin at a coffee shop, because that’s where all the best math happens. The question is, for any number of dots, what’s the most pairs you can make that are exactly one unit apart? So, literally, how many pairs of dots can be an inch away from each other?

[Jamie]: So, if I put nine dots in a straight line, I get eight pairs one inch apart, right?

[Alex]: Exactly, but if you arrange those nine dots in a three-by-three grid—think tic-tac-toe board—you can actually get twelve pairs an inch apart. The grid does better. But here’s the kicker: what happens when you have a million dots? Or a billion?

[Jamie]: I’m guessing that's not just a bigger tic-tac-toe board.

[Alex]: Right. Erdős figured out a clever way to space out the grid, so as you add more dots, you get just a tiny bit more pairs than you’d expect. He then made a bold claim: nobody could ever do meaningfully better than his grid arrangement. That became the Erdős unit distance conjecture. And for almost 80 years, no one could beat it.

[Jamie]: Until now—cue dramatic music! So, what did OpenAI’s AI actually do? Did it just brute-force a million combinations, or was it something fancier?

[Alex]: Way fancier. OpenAI used a general-purpose reasoning model—it wasn’t even a math-specialized AI. They gave it a single prompt: basically, “Is the Erdős conjecture true or false?” And it just… went to work. No step-by-step coaching, no hand-holding.

[Jamie]: Wait, so it wasn’t like DeepMind’s AlphaProof that’s built for math?

[Alex]: Nope, not at all. This was an experimental, unreleased model. Imagine throwing a Rubik’s Cube at someone and they invent a new way to solve it, blindfolded. The AI generated a 125-page chain of reasoning—like the War and Peace of math proofs. [laughs]

[Jamie]: Please tell me someone read all 125 pages.

[Alex]: Oh, they did. OpenAI sent the proof to nine independent mathematicians—including Fields Medalist Tim Gowers. They actually spent months checking the work, and then published a companion paper verifying it. So, this isn’t just AI hype—it’s peer-reviewed math.

[Jamie]: Amazing. But what did the AI actually find? Did it show Erdős was totally wrong?

[Alex]: Not “totally” wrong—it didn’t find the absolute best arrangement, but it found one that beats the grid by a real, measurable margin. Erdős thought you’d never do better by more than a hair. The AI showed you can do better by a polynomial margin. That’s a huge leap.

[Jamie]: Polynomial margin? For those of us who dropped out of calculus, what does that mean?

[Alex]: [laughs] It means the improvement grows a lot faster as you add more dots. Instead of a tiny, barely-there advantage, you get a solid, growing lead. The grid isn’t the king anymore.

[Jamie]: So, does this mean AI’s now better at math than all the humans?

[Alex]: I wouldn’t hand over the Fields Medal just yet. The AI did something wild: it used algebraic number theory. That’s a part of math that, until now, had nothing to do with this problem. Instead of arranging dots on a flat plane, it picked points that are solutions to complicated algebraic equations—then mapped them down into two dimensions.

[Jamie]: So, like, math origami? Folding higher dimensions into a flat page?

[Alex]: That’s… actually a perfect analogy. Except, these arrangements are so complex that you can’t even draw them on paper, even for a small number of dots.

[Jamie]: So, this proof is more like “trust me, the math works out,” rather than, “here, let me sketch it for you on a napkin.”

[Alex]: Exactly. And the AI used some heavy-duty math machinery—stuff like class field towers and the Golod–Shafarevich theorem. Tools that have been around for decades, but nobody ever thought to connect them to this problem.

[Jamie]: Okay, honest question: OpenAI’s had some, um, over-excited announcements before. How do we know this isn’t just another overhyped AI math claim?

[Alex]: Great point. In 2025, OpenAI’s VP tweeted that GPT-5 had “solved” a bunch of Erdős problems—but it turned out GPT-5 just found existing solutions that weren’t in one guy’s database. Not exactly earth-shattering. That post got deleted in a hurry.

[Jamie]: Ouch. Mathematicians do not mess around with their databases.

[Alex]: No, they do not. [laughs] But this time is different. Not only did independent experts check the proof, but the same mathematician who called out OpenAI last time, Thomas Bloom, is now a co-author on the verification paper. That’s about as close as you get to a peace treaty in math.

[Jamie]: Love a good redemption arc. So, is this really the first time AI’s solved a big open math problem?

[Alex]: Well, OpenAI’s being careful with their words. They say it’s “the first time AI has autonomously solved a prominent open problem central to a field of mathematics.” There have been AI breakthroughs before—DeepMind’s AI won medals at math Olympiads, and even nudged some open bounds in geometry. But those were either competition problems or relied on heavy human scaffolding.

[Jamie]: So, this is the first time an AI took a single open-ended prompt, went off on its own, and hit a real, famous unsolved problem?

[Alex]: Exactly. And, crucially, it wasn’t a math-specialized AI. It just… reasoned its way through the jungle.

[Jamie]: Okay, but how much of the proof was “AI magic” and how much was “humans cleaning up AI’s messy room”?

[Alex]: [laughs] Honestly? A bit of both. The raw AI output was edited and organized by mathematicians to make it publishable. As Thomas Bloom put it, “The human still plays a vital role in discussing, digesting, and improving this proof.” It’s a collaboration, not a handoff.

[Jamie]: So, should mathematicians be worried for their jobs? Or are we looking at the ultimate research assistant?

[Alex]: I’d say more the latter. The real story here is about what’s possible when AI models can hold together long, cross-disciplinary chains of reasoning. It’s early days, but in the future, we could see AI helping crack problems in biology, physics, even medicine—not just math puzzles.

[Jamie]: Until then, I’ll stick to drawing dots on napkins. [laughs] But seriously, this is wild. Anything else our listeners should know?

[Alex]: Just that this is a milestone, not the final destination. There’s still plenty of room for humans—and AIs—to team up and solve the next big challenge. And as always, it pays to read the fine print behind the headlines.

[Jamie]: Well, that’s all the nerdy math drama we have for today. Thanks for tuning in to the Nerd Level Tech AI Cast. If you enjoyed this episode, leave us a review, share it with your math-loving friends, or send us your own napkin-conjectures.

[Alex]: We’ll be back soon with more tales from the AI frontier. Until then, keep your dots at least one unit apart—and your curiosity closer.

[Jamie]: Bye, folks!

[Alex]: See ya!