Introduction: The Day AI Broke a 1946 Conjecture
Paul Erdős offered $500 for a proof of his unit distance conjecture. He died in 1996, still believing it was true. Last week, an AI disproved it.
On May 20, 2026, OpenAI announced that its latest reasoning model had done something no machine had ever done before: independently solve a longstanding open problem in mathematics. The target was a conjecture posed by Erdős in 1946 — one of the most influential mathematicians of the 20th century, known for his prolific output and his habit of offering cash prizes for unsolved problems. The conjecture concerned the maximum number of unit-distance pairs that can exist among n points in the Euclidean plane, a problem that had resisted attack for eight decades.
The result, published by OpenAI as a PDF on their website, was quickly verified and improved upon by Will Sawin, a mathematician at Princeton University. Within days, the mathematics community had confirmed that Erdős’s conjecture was false — and that an AI had found the counterexample. Sawin’s paper, which refines and extends the AI’s construction, was published on arXiv as preprint 2605.20579 .
AI systems have claimed mathematical breakthroughs before. This is the first time one has been independently verified by the mathematical community on a problem that wasn’t a competition exercise or a known theorem. The implications ripple far beyond geometry — into scientific discovery, AI reasoning, and the practical work of developers who build on these models.
What Was the Erdős Unit Distance Conjecture?
The problem is deceptively simple. Take n points in the plane. How many pairs of points can be exactly one unit apart? This quantity is called g(n). Erdős wanted to understand how g(n) grows as n becomes very large.
Erdős himself provided a lower bound using a grid construction: arrange the points in a √n by √n grid — you get roughly n^{1 + c / log log n} unit-distance pairs. On the upper end, the best known bound, established decades later, was g(n) ≤ 1.936 n^{4/3}, meaning the function grows at most like n^{4/3}. The gap between this upper bound and Erdős’s conjectured near-linear growth remained unresolved for decades.
Erdős conjectured that for any positive ε (no matter how tiny), g(n) = O(n^{1+ε}). In plain language: the number of unit-distance pairs grows only slightly faster than linear. He offered a prize of $300 for a proof or disproof, later raised to $500. The relatively modest prize reflected his own belief that the conjecture was probably true.
The conjecture had real mathematical weight. The unit distance problem connects to graph theory, harmonic analysis, and extremal combinatorics. It is part of a broader family of problems (Erdős–Moser, Erdős–Szekeres, Erdős–Graham) that probe the boundaries of geometric combinatorics. A proof or disproof would force mathematicians to rethink the underlying structure of distance geometry.
What the AI Actually Did
The OpenAI model did not stumble on an answer through brute force. It constructed a geometric arrangement with a specific algebraic property that forced unit-distance pairs to proliferate. The technical machinery behind this arrangement is a novel family of polynomial constructions using algebraic number theory — specifically, the class field tower method, which relies on the Golod–Shafarevich theorem from class field theory. This is deep mathematics, the kind reserved for advanced graduate study.
The construction produced a set of n points in the plane that contained more than n^{1 + ε} unit-distance pairs, where ε was approximately 6.24 × 10^{-38}. That is an extraordinarily small exponent — but any positive exponent, no matter how tiny, disproves Erdős’s claim that the growth was arbitrarily close to linear.
Erdős’s conjecture said “the exponent can’t be made larger than 1.” The model showed it can — by 6.24 × 10^{-38}. Then Will Sawin, working independently, improved the exponent dramatically to ε > 0.014 — a leap from 6.24×10⁻³⁸ by a factor of roughly 10³⁶. This validated the AI’s approach rather than just correcting it, and opened a genuine new avenue of inquiry.
The model did not produce intuitive reasoning. It generated constructions that mathematicians found alien — not wrong, but not obvious either. This is becoming a pattern with AI-generated mathematics: the results are correct, but the path to them is often inscrutable. The math community had to independently verify and, in some cases, reinterpret what the model had found.
Key insight: The results are correct, but the path to them is often inscrutable.
Why the Math Community Accepted This Result
AI-generated mathematical results have historically been met with skepticism — and sometimes rightly so. High-profile claims of machine-driven breakthroughs have failed to hold up under scrutiny before (remember the 2024 claimed proof of the Riemann hypothesis that collapsed within hours?). When this new result appeared, the community approached it with the same caution.
What changed this time was rigor. The model’s output was published as a complete mathematical paper with full reasoning — not as a blog post or PR announcement. The mathematics was complete enough that a human expert — Will Sawin — could verify, understand, and improve upon it within days. The improvements did not reveal errors in the AI’s work; they extended it.
Sawin’s paper is titled “An explicit lower bound for the unit distance problem” and is even more explicit than the AI’s. The model produced an inexplicit but provably positive exponent. Sawin derived an explicit lower bound of ε > 0.014 using similar class field tower techniques. The AI had opened the door; Sawin walked through it.
Key insight: The AI had opened the door; Sawin walked through it.
This pattern — AI discovers, humans interpret and improve — is the most credible pathway for machine-assisted mathematics. The community didn’t accept the result just because an AI said so. They accepted it because the proof was checked, the logic held, and the bound could be improved by standard mathematical methods.
The Trajectory of AI Reasoning Capabilities
The Erdős disproof marks a qualitative shift in what AI reasoning models can accomplish. Earlier models succeeded at theorem proving within constrained formal systems (Lean, Isabelle, Coq) and at solving math competition problems (IMO, Putnam). These were impressive but limited — well-scoped problems, bounded search spaces, solutions that could be verified automatically.
The Erdős problem was none of those things. It was an open problem that had resisted specialist attention for decades. The model had to generate genuinely new mathematics — constructing algebraic number fields with specific discriminant properties, applying the Golod–Shafarevich theorem, and deriving combinatorial consequences. This required sustained, multi-step reasoning across multiple subdomains of mathematics, none of which were explicitly encoded in the model’s training data as a solved problem.
What this result shows, beyond what earlier ones did:
- Combinatorial depth: The model chained dozens of mathematical steps without losing coherence. Each step depended on the previous one, and errors would have accumulated catastrophically.
- Cross-domain transfer: Algebraic number theory, class field theory, graph theory, and combinatorial geometry were all brought to bear on a single problem. This is the kind of synthesis that characterizes expert-level mathematical research.
- Generative novelty: The construction was not a recombination of known solutions. It was genuinely new — the class field tower method had not been applied to the unit distance problem before.
The model behind this achievement is a next-generation reasoning system from OpenAI, building on the o-series architecture (o1, o3) with substantially more inference-time compute and improved chain-of-thought coherence. The exact architecture has not been disclosed, but the capability jump from previous general-purpose models is significant.
What This Means for Scientific Discovery
If a model can independently make progress on an 80-year-old problem in pure mathematics, what else can it do? The obvious answer is more mathematics — and the mathematics community is already reckoning with what that means. But the less obvious — and more consequential — answer involves every domain that depends on formal reasoning.
Mathematics
Mathematics is the most natural domain for AI-driven discovery because it is formal, verifiable, and rewards deep search over large spaces. Reasoning models will become routine collaborators in mathematical research. They won’t replace mathematicians — they’ll generate conjectures, construct examples and counterexamples, and explore search spaces that humans would not have the patience or cognitive stamina to navigate.
The Erdős disproof is the opening act. Expect more results like it in the coming months and years, especially as models gain access to formal proof assistants that can verify their work automatically.
Physics and Engineering
The same reasoning capabilities that produced a class field tower construction can be applied to problems in physics: optimizing experimental designs, searching for solutions to systems of differential equations, or identifying symmetries in complex systems. The Navier-Stokes equations describe fluid flow and remain one of the Clay Millennium Problems. A model that can construct algebraic number fields from scratch could search for solution regimes or counterexamples that have eluded human analysts for decades. The key difference is verification — physics requires experimental confirmation, not just mathematical validity — but the synthesis of cross-domain knowledge is directly transferable.
Biology and Medicine
Biology is messier than mathematics. The formal structure is weaker, the data is noisier, and the causal relationships are harder to isolate. But the pattern-matching and combinatorial search capabilities that enabled this result are already being applied to protein folding, drug discovery, and genomics. The Erdős result shows that reasoning models can now handle longer chains of inference — exactly what is needed for modeling protein folding pathways, where small local interactions cascade into global conformations across timescales that brute-force simulation cannot capture. This same inferential depth matters for any biological system where effects propagate through multiple layers of abstraction.
Computer Science
For computer science, the implications are immediate. Reasoning models that can construct novel mathematical objects can also construct novel algorithms, data structures, and optimization strategies. The same class of models that disproved Erdős’s conjecture can analyze program invariants, verify correctness properties, and search for counterexamples in program logic.
Practical Implications for Developers
What does any of this mean for a developer shipping code today? It depends on how you use reasoning models. If you treat them as autocomplete-on-steroids — generating snippets, writing tests, summarizing docs — the Erdős result changes nothing. But if you are building applications that depend on reliable, multi-step reasoning, the landscape is shifting under your feet.
Reliability Is No Longer the Ceiling
The single biggest complaint about LLMs for reasoning tasks has been hallucination. Past instances of AI-generated mathematics failing under scrutiny have been high-profile examples of this systemic problem. The Erdős disproof shows that reliability at the frontier has crossed a threshold — the model’s output was correct, provably correct and improvable by human experts.
This matters for any application where correctness is essential: code review, formal verification, legal reasoning, compliance checks, audit trails. The margin of error is shrinking.
Chain-of-Thought Is Becoming a Product Feature
The capability to maintain coherent reasoning over dozens of steps has direct product implications. Agents that plan travel itineraries, manage multi-party negotiations, or orchestrate CI/CD pipelines can now handle more branches, more dependencies, and more failure modes before losing coherence.
Developers building agentic systems should design for this. The limiting factor has shifted from the model’s ability to maintain a chain of thought to the developer’s ability to specify the chain correctly.
Inference Cost Changes the Calculus
Models of this capability are expensive to run. The Erdős disproof required enormous inference-time compute — millions of dollars’ worth of GPU time. That cost will come down, as it always does, but for now, narrow reasoning tasks (file a bug report, draft a PR description) are better served by smaller, cheaper models.
The strategic insight: reserve frontier reasoning models for tasks where correctness has high marginal value. Use cheap models for everything else.
Expect More “AI-First” Mathematical Software
The Erdős result will accelerate investment in AI-assisted mathematics tools. Expect more products like Lean , Isabelle , and Wolfram Alpha to integrate reasoning-model backends. If you work in quantitative fields — data science, quantitative finance, cryptography — the tools you use will change faster than you expect.
Verification Infrastructure Becomes Critical
If models can produce correct but inscrutable results, the bottleneck shifts from generation to verification. Developers building on reasoning models need to invest in verification layers: formal specification of expected behavior, property-based testing, and automated checkers that validate model outputs against ground-truth rules.
The Erdős result was accepted because Sawin could verify it. In software, the same principle applies: your CI pipeline should be able to check that the model’s suggestions are correct, not just that they compile.
References and Further Reading
- arXiv:2605.20579 — Will Sawin, “An explicit lower bound for the unit distance problem”
- OpenAI’s paper on the Erdős unit distance problem — the original AI result (PDF)
- Wikipedia: Erdős unit distance problem
- Erdős Problems #90 — the conjecture on erdosproblems.com
- OpenAI announcement blog post — “An OpenAI model has disproved a central conjecture in discrete geometry”
Conclusion: The $500 Prize That Changed Everything
Paul Erdős initially offered $300 for a proof or disproof of his unit distance conjecture, later raising it to $500. He died in 1996, still believing it was true. Almost thirty years later, an AI model collected the prize — not in cash, but in the form of a verified result that reshapes a field of mathematics.
The significance is not that a machine beat a human at math. It is that a machine engaged in genuine mathematical discovery: constructing a nontrivial family of objects, reasoning across multiple subfields, and producing a result that a human mathematician could verify and improve. That is not pattern matching. That is research.
Key insight: That is not pattern matching. That is research.
For developers, the takeaway is straightforward. The reasoning capabilities that enabled this result are coming to the APIs and tools you use. The window in which “AI can’t really think” was a defensible position is closing. The question now is how to build systems that leverage that reasoning effectively and safely.
The Erdős conjecture was an 80-year-old problem. It took an AI to disprove it. The next problem it solves might be yours.
No comments yet