Fields medalist Timothy Gowers: “This is the first clear case of AI solving a famous, unsolved mathematical problem — and the first mathematical breakthrough achieved by AI autonomously.” The model that solved it was not math-specific — it was a general reasoning system
On May 20, 2026, OpenAI released a statement that sent shockwaves through both the mathematics and AI communities: its internal general reasoning model autonomously disproved the Erdős unit distance conjecture — a classic problem posed by Paul Erdős in 1946 that had remained unsolved for nearly 80 years.
This wasn’t the first time OpenAI claimed AI had solved mathematical problems. Seven months ago, former VP Kevin Weil posted on X: “GPT-5 found solutions to 10 (!) previously unsolved Erdős problems.” It turned out GPT-5 had merely rediscovered solutions already in the literature — earning taunts from Yann LeCun and DeepMind CEO Demis Hassabis, and a swift deletion of Weil’s post. But this time, the mathematics community is standing behind OpenAI.
The Erdős Unit Distance Problem: Simple Enough for a Napkin, Hard Enough for Five Generations
The problem’s statement is deceptively simple:
Given n points on a plane, what is the maximum number of pairs that can be exactly distance 1 apart?
- Three points can form an equilateral triangle — every pair is distance 1
- Four points can’t just be arranged in a square, because the diagonal isn’t 1
- For nearly 80 years, mathematicians reached a core consensus: the optimal arrangement looks roughly like a square grid
In mathematical terms, they believed the growth rate of unit distance pairs was approximately O(n) — essentially linear. Written formally: u(n) ≤ n^(1+o(1)), where o(1) approaches 0.
Known as the Erdős unit distance conjecture, this is one of the most famous and enduring unsolved problems in discrete geometry.
OpenAI’s Breakthrough: Approaching Through Algebraic Number Theory
OpenAI’s general reasoning model didn’t take a geometric approach. Instead, it came at the problem from algebraic number theory, constructing an entirely new family of point arrangements that breaks the 80-year consensus on what “optimal solutions look like.”
Specifically, the model discovered a new family of constructions that outperforms the square grid arrangement in terms of unit distance pairs, thereby disproving the conjecture.
This means mathematicians’ 80-year understanding of this problem’s “optimal solution structure” was wrong — and AI found this out without anyone pointing it in the right direction.
Crucially, the model that achieved this was not a math-specialized system — it was a general reasoning model, meaning AI’s reasoning capability has reached a point where it can autonomously explore uncharted intellectual territory.
The Mathematics Community Responds: This Time, It’s Real
OpenAI didn’t make this claim alone. The company simultaneously published supporting statements from several prominent mathematicians:
- Noga Alon (Princeton Professor, Israel Academy member)
- Melanie Wood (Harvard Professor, first woman to achieve a perfect score at the International Math Olympiad)
- Thomas Bloom (maintainer of the Erdős Problems website, who previously called Weil’s post “a dramatic misrepresentation”)
- Timothy Gowers (Fields medalist)
Gowers’ endorsement is particularly significant: “This is the first clear case of AI solving an extremely famous, unsolved mathematical problem, and the first mathematical breakthrough achieved by AI autonomously.”
Bloom’s statement carried its own weight: “AI is helping us more fully explore the cathedral of mathematics we have built over the centuries. What other unseen wonders are waiting in the wings?”
These public endorsements stand in stark contrast to the “false alarm” seven months ago — and signal that the mathematics community is beginning to take seriously AI’s role as a “research partner.”
The Significance of a General Model: Beyond Mathematics
OpenAI emphasized that the model was a general reasoning model, not a system designed specifically for mathematics or geometry. This detail may be underappreciated:
- If AI can autonomously discover blind spots in human mathematical understanding — without domain-specific priors — it implies comparable or even greater potential in biology, physics, engineering, and medicine, fields that similarly depend on complex reasoning chains
- General reasoning means AI is no longer a “domain specialist” but a “cross-domain explorer”
- This marks a transition from AI as “tool” to AI as “research partner” — not just executing assigned tasks, but proposing directions humans hadn’t considered
Noam Brown, who leads the general reasoning model effort at OpenAI, has stated that the model will be released as soon as possible.
The Broader Context: AI for Science Is Accelerating
OpenAI’s mathematical breakthrough doesn’t exist in isolation. On the same day, other major AI industry developments were unfolding:
- Anthropic announces first profitable quarter: Q2 revenue projected at ~$10.9 billion, achieving operating profit for the first time
- Anthropic signs $40B compute deal with xAI: Paying $1.25 billion per month through 2029
- OpenAI pushing toward IPO: Sam Altman aims for a September listing
- Nvidia hits another record: $81.6 billion quarterly revenue, $43 billion in startup holdings
- AMD’s Lisa Su in Shanghai: Declares “China is core to AMD’s roadmap”
Together, these events paint a larger picture: the AI industry is entering a “full-speed” phase — from fundamental research breakthroughs to commercial deployment, from compute arms races to full-blown capital market explosions.
Looking Back and Forward: From the GPT-5 Debacle to a Genuine Breakthrough
Seven months ago, OpenAI’s “false breakthrough” was nearly a public relations disaster. But today’s announcement shows OpenAI learned an important lesson: in scientific claims, better late than unverified.
This time, OpenAI didn’t just tweet an announcement — it simultaneously published supporting statements from mathematicians. This “verify first, announce second” approach may become the standard protocol for future AI-driven scientific discoveries.
AI Pulse View: OpenAI’s breakthrough marks AI’s entry into a new phase — from “what can it do” to “what can it discover.” When AI is no longer a tool that executes given tasks, but a “researcher” capable of autonomously exploring the unknown and overturning established human understanding, the entire paradigm of scientific research may need reconsideration. For fundamental disciplines like mathematics, physics, and biology, this could mean an era of accelerated discovery. But it also raises profound questions about “human uniqueness in scientific discovery” as a philosophical proposition.
AI Pulse View: The fact that this breakthrough came from a general reasoning model, not a specialized one, is perhaps the most significant detail. It means AI’s “generality” may be more powerful than we imagined — not just surpassing humans in specific domains, but building connections across disciplines that humans haven’t yet found. This explains why OpenAI emphasizes “general” so strongly: the future AI competition may not be about whose specialized model is stronger, but whose general reasoning ability can better cross disciplinary boundaries.