AI Independently Overturns
A Core Conjecture That Has Puzzled Academia for 80 Years
On May 20, 2026, OpenAI's internal general reasoning model independently solved the Planar Unit Distance Conjecture proposed by Erdős in 1946. Using Golod-Shafarevich theory from algebraic number theory—a field never previously associated with this problem—it executed a cross-disciplinary "dimensional strike" from geometry to number theory. This is not assistance, retrieval, or optimization—this is the first time AI has produced a mathematical discovery as an originator. This serves as the most direct justification for ResearchLinkAI's commitment to an "AI-native research roadmap."
Years as an Open Problem
Pages of Proof Documentation
Top Mathematicians Co-Signed Verification
Overturned Erdős n^(1+o(1)) Upper Bound
1. The Core Event: How an 80-Year Conjecture Was Overturned
The planar unit distance problem is stated simply: Given n points on a plane, what is the maximum number of pairs of points that can be exactly distance 1 apart? In 1946, Erdős used a Gaussian integer grid to construct a lower bound of n^(1+c/log log n), conjecturing that the true upper bound should be n^(1+o(1))—meaning any super-linear growth could only be marginal.
For eighty years, the strongest upper bound remained at O(n^(4/3)) from 1984, and the lower bound saw almost no improvement. The entire mathematics community assumed the Erdős conjecture was true. Until the OpenAI model proved: There exist infinitely many values of n such that the number of unit distance pairs is at least n^(1+δ), where δ>0 is a fixed positive constant. Princeton mathematician Will Sawin further refined this to obtain δ=0.014—directly overturning Erdős's conjecture.
"If a human had written this paper and submitted it to the Annals of Mathematics, I would have recommended acceptance without hesitation. No previous AI-generated proof has come close to this standard."
—— Timothy Gowers, Fields Medalist
2. How It Was Done: A Cross-Disciplinary Bridge from Geometry to Number Theory
The model's solution completely upended the cognitive framework of the past eighty years. Instead of continuing to refine tools within the traditional toolbox of "combinatorial geometry + discrete Fourier analysis + real algebraic geometry," it turned its attention to the infinite class field tower theory in algebraic number theory, which seemed entirely unrelated to the problem.
The construction logic is as follows: Use the Golod-Shafarevich theorem to guarantee the existence of a sequence of totally real fields F_j with Galois groups of degree 3; obtain CM fields K_j = F_j(i) by adding the imaginary unit i; use the Chebotarev density theorem to select specific rational primes that split completely in K_j; and finally project these high-dimensional lattice points onto the two-dimensional plane via Minkowski embedding. Golod-Shafarevich ensures the infinitude of the class field tower, while the Chebotarev splitting condition ensures the planar point set possesses sufficiently many unit distance pairs—the coupling relationship between the two had never been noticed by any mathematician before.
The critical turning point in the model's chain of thought appeared in this sentence: "In principle, all extremal examples can be taken to be algebraic. But the degree and height of that algebraic realization might be astronomically huge... Perhaps that enormous degree is not just a nuisance, but a source of potential counterexamples."—This cognitive flip from "huge degree is an obstacle" to "huge degree is a resource" represents a psychological threshold that human mathematicians have long failed to cross.
3. Why AI Instead of Humans
Nine top mathematicians provided explanations on three levels in the accompanying paper:
- Domain Barriers: Modern mathematics is highly specialized; discrete geometers are unfamiliar with class field towers, and algebraic number theorists would not think of geometric applications. Noga Alon stated bluntly, "This fact comes as a great surprise to experts in algebraic number theory."
- Cognitive Anchoring: The Erdős conjecture carried axiomatic cognitive weight, leading human mathematicians to automatically default to "proving" rather than "disproving." The model's CoT showed that it spent most of its time trying to construct counterexamples—a direction unencumbered by community authority.
- Aesthetic Preferences: Humans have a strong pursuit of "elegance" and tend to avoid constructions of high degree and high complexity. AI carries no such aesthetic baggage.
"The advantage of AI lies not only in its ability to try all known methods, but also in its capacity to explore more dangerous territories for longer periods without being overwhelmed."
—— Jacob Tsimerman, University of Toronto
This observation is crucial: AI's breakthrough does not stem from some mysterious "machine intuition," but from relentless systematic exploration capabilities and a natural tolerance for "massive computational overhead" and "complex structures."
4. From Assistant to Originator: A Three-Stage Leap in Research Paradigms
Placing this breakthrough in the historical context of AI participation in scientific research reveals a clear three-stage curve:
Computational Accelerator
Optimizing execution efficiency of known methods; key ideas still originated from humans.
Competition-Level Solver
AlphaProof won silver at the IMO; Aristotle/Seed-Prover won gold. Yet still confined to well-structured competition problems.
Open Problem Originator
Facing open problems with no known answers or standard methods, autonomously producing new methodologies.
Test-time compute scaling experiments released by OpenAI show: The more inference compute allocated to the model, the higher the success rate on the unit distance problem monotonically rises. This implies that the bottleneck for scientific discovery is shifting from "the scarcity of human genius" to "investment in computational resources and allocation of verification attention."
5. Why ResearchLinkAI Adheres to an AI-Native Roadmap
The significance of this event to ResearchLinkAI lies not in "AI solving another problem," but in the fact that it fundamentally alters the possibility frontier of the research services business.
Under the previous model, the ceiling for research services was determined by the "cognitive bandwidth of the expert accepting the order"—you could only serve clients in domains where your experts excelled, and you could only accept as many orders as your experts had time for. AI-native means:
- Cross-disciplinary combination capabilities are no longer limited by a single expert's training background—this is the same capability AI used to overturn the Erdős conjecture across disciplines, applied to customer service to provide professional solutions for clients with diverse backgrounds.
- Depth of exploration is unconstrained by fatigue—AI can conduct prolonged systematic exploration on client problems, turning Tsimerman's notion of "exploring dangerous territories longer without being overwhelmed" into a daily service capability.
- Scalable output—A Research Agent can serve multiple projects in parallel, with Vetted Experts intervening only at critical nodes for review, resulting in overall marginal costs far lower than purely manual models.
This is why our service matrix—MATLAB Simulation Tutoring, Bioinformatics Analysis, Molecular Docking, Medical Statistics, AI + Expert Research Collaboration—is built upon a dual-layer structure of "AI-native execution + Expert review." OpenAI's breakthrough serves as the most powerful external endorsement of this roadmap.
6. Conclusion: A Silent Cognitive Revolution
Anthropic co-founder Jack Clark predicted in May 2026: "AI will help achieve a Nobel Prize-level discovery within 12 months." OpenAI's unit distance breakthrough—occurring in the same week—has transformed this prediction from "speculation" into "evidence."
The most profound shift may lie not in what AI can do, but in the reality that "the speed of scientific discovery will no longer be limited by the scarcity of human genius, but by the resources we are willing to invest in computation and the attention we dedicate to verification". Under this new constraint, the best research service companies will be those that engineer the integration of AI's depth of exploration with the judgment of human experts.
If you are struggling with research topics, paper directions, or patent planning, feel free to chat with ResearchLinkAI—we use the latest AI-native methodologies to help you break down "seemingly impossible" research goals into clear, executable roadmaps.