OpenAI’s General-Purpose Reasoning Model Achieves Breakthrough, Disproving a Nearly 80-Year-Old Mathematical Conjecture
On May 21, AI research company OpenAI announced a significant research achievement: an internal, novel general-purpose reasoning model independently disproved the Erdős unit distance problem, a long-standing core conjecture in discrete geometry. This development has drawn widespread attention in the academic community, with renowned mathematicians like Fields Medalist Timothy Gowers praising it as a sign of a profound shift in the role of artificial intelligence within the scientific research paradigm.
1. An 80-Year-Old Geometric Puzzle: The Erdős Unit Distance Problem
The Erdős unit distance problem was proposed by the legendary mathematician Paul Erdős in 1946. It seeks to determine the maximum number of pairs of points, denoted as u(n), that can be exactly one unit apart among n points in a two-dimensional plane.
Erdős himself established a lower bound based on a square grid construction, showing that u(n) could be at least on the order of n * exp(c * log(n)/log(log(n))). At the same time, he famously conjectured that the upper bound for u(n) could not grow much faster than linear, in the form of n^(1+o(1)), where the o(1) term approaches zero as n increases. This implied that the optimal point set construction should essentially resemble a “grid-like” structure.
Over the past nearly 80 years, mathematicians have made relentless efforts to solve this problem. On the upper bound, Spencer, Szemerédi, and Trotter proved a bound of O(n^(4/3)) in 1984, with no substantial improvements since. On the lower bound, despite numerous attempts, no one had managed to surpass the result from Erdős’s original grid-based construction.
2. AI’s Disruptive Construction: A New Lower Bound and the Conjecture’s Falsification
OpenAI’s model did not follow traditional geometric or combinatorial paths. Instead, it constructed a series of completely new point set structures. According to the proof documents released by OpenAI, the point sets constructed by the model can generate a number of unit distances on the order of n^(1+c), where c is a definite positive constant.
This result decisively refutes Erdős’s original conjecture. A growth rate of n^(1+c) is polynomially faster than n^(1+o(1)), completely shattering the long-held academic consensus that the “optimal structure should be grid-like.”
After the AI generated the initial proof, Princeton University mathematics professor Will Sawin verified and refined it, further confirming that the constant c can be explicitly set to at least 0.014. This collaboration between human and AI rapidly solidified the mathematical foundation of this groundbreaking result.
3. Interdisciplinary Fusion: Innovative Application of Algebraic Number Theory to a Geometric Problem
What has most captivated the academic world about this breakthrough is the deep insight and interdisciplinary integration demonstrated by the AI. The model did not rely on brute-force search or exhaustive enumeration but instead invoked complex tools from advanced algebraic number theory to solve a seemingly elementary geometry problem.
The core of the model’s proof strategy involves:
- Constructing complex algebraic number fields: It did not confine itself to the Gaussian integers used in traditional constructions but introduced higher-dimensional, more symmetric algebraic number fields. These higher-dimensional spaces provided the theoretical possibility for generating more unit distances.
- Applying cutting-edge number theory: To prove that its proposed algebraic structures were mathematically viable, the model skillfully employed “Infinite Class Field Towers” and the “Golod–Shafarevich Theory” in its reasoning chain.
These are advanced tools in algebraic number theory research, typically mastered only by experts in the field. The ability of a general-purpose reasoning model to spontaneously connect them to a geometry problem and construct a complete, rigorous proof chain demonstrates its powerful abstract reasoning and knowledge transfer capabilities.
4. A New Paradigm for Scientific Discovery: AI as an Original Explorer
OpenAI specifically emphasizes that this proof was accomplished by a general-purpose reasoning model, not a specialized system custom-built for a specific mathematical task. This indicates that AI’s capabilities are expanding from task-specific execution to broader, human-like logical reasoning and creativity—a key step toward Artificial General Intelligence (AGI).
This event reveals a new paradigm for scientific research:
- AI as a discovery engine: AI can explore “uncharted territories” that humans may have missed due to cognitive biases or knowledge barriers, proposing entirely new concepts and pathways.
- Deeper human-machine collaboration: AI is responsible for generating original, potentially highly complex initial proofs, while human experts use their deep intuition and judgment to verify, interpret, and refine them, transforming the AI’s “discoveries” into part of the human knowledge system.
As Fields Medalist Thomas Bloom stated, the AI has shown us that number-theoretic structures possess a far deeper power in solving geometric problems than we had ever imagined. This achievement not only solves a difficult mathematical problem but, more importantly, it “teaches us new knowledge.” This ability to integrate knowledge across domains, conduct long-chain rigorous reasoning, and produce original ideas signals its immense potential for application in other complex scientific fields such as physics, biology, and materials science.
