Wednesday, November 18, 2026

Top 5 This Week

Related Posts

OpenAI refutes 80-year-old Erdős conjecture


The essentials in 30 seconds

  • On May 20, 2026, OpenAI announces that one of his models of reasoning refuted the unit distance conjectureposed by the Hungarian mathematician Paul Erdős in 1946.
  • During 80 years oldmathematicians believed that square grids were the best configuration. The model found a completely new build family which beats the grid, and proved it mathematically.
  • The detail that changes everything: it’s a general modelnot a system trained specifically for mathematics nor a solver dedicated to this problem.
  • Math community verdict: “perhaps the most significant contribution of generative AI to mathematics to date”. But the real scope is more nuanced, detailed in the article.

This is an announcement that the mathematical community was waiting for without really believing in it. On May 20, 2026, OpenAI publishes a research note claiming that one of his reasoning models has refuted the planar conjecture of unit distances by Paul Erdős, one of the best-known open problems of discrete geometry, posed in 1946.

The information was reported by TechCrunch, Slashdot, GIGAZINE and an official OpenAI press release on May 21. For mathematicians, it is an event. For the trajectory of large language models, it is probably even more significant.

Erdős’ unit distance problem (1946)

The problem posed by Erdős in 1946 is deceptively simple to state.

Given n points in the plane, how many times can we observe the distance 1 between two of these points? In other words, how many pairs of points are exactly distance 1 from each other? What is the maximum possible of this quantity as a function of n?

Erdős showed that a square grid configuration gives at least n × (log n)^c pairs at unit distance (for a constant c). During almost 80 years oldmathematicians have tried to improve this bound, without significant success. The implicit consensus: the square grid was probably the optimal configuration, or very close to it.

The problem is central in discrete geometry because it concerns the combinatorics of point configurations, and has applications in graph theory, cryptography and theoretical computer science. Several of the greatest mathematicians of the 20th century broke their teeth over it, including Erdős himself who discussed it in dozens of articles until his death in 1996.

The detail that sets this announcement apart from all other AI math exploits can be summed up in one sentence: the model was not trained for that.

According to the OpenAI press release, the proof comes from a general reasoning modelnot a system specialized in mathematics (as DeepMind’s AlphaGeometry was), not a scaffolded solver to run through proof strategies, and not a system targeted on the problem of unit distances in particular. This is the same type of model that OpenAI offers in its consumer and enterprise products, faced with an open problem with an open set of instructions.

The model produced a completely new family of geometric constructions which corresponds neither to a square grid, nor to a triangular grid, nor to any of the classic configurations studied for 80 years, and has formally demonstrated that they reach a number of unit distances greater than that of grids. The proof was verified by several mathematicians in the days that followed.

To compare, DeepMind’s AlphaGeometry (announced in early 2024) solved Olympic geometry problems by relying on a symbolic solver combined with a LLMLarge Language Model, a language model trained on billions of parameters to generate and understand text.. Here, the approach is radically different: a single model, a single round of reasoning, an original proof.

Why do mathematicians talk about turning points?

Several mathematicians cited by TechCrunch and the specialized press do not hesitate to talk about first major mathematical contribution of a generative AI. Nuance matters.

Until now, models like ChatGPT or Claude were capable of solve existing math problems, provided the solution is already documented somewhere in their training corpus. They were capable of explainof rephraseand sometimes generalize known results. But none had, to our knowledge, produces an original proof of a new theorem on a problem that has been open for decades.

A mathematician quoted by GIGAZINE summarizes: “AI has gone from the status of assistant to that of contributor”. The boundary between “asking questions” and “answering open questions” has just been crossed in a specific case, and it is this boundary that separates a productivity tool from a true research partner.

The symbolic impact is such that the math community, traditionally rather skeptical towards AI announcements, has reactedArtificial General Intelligence, a hypothetical AI at the level of human intelligence, capable of solving any intellectual task. with rare positive unanimity. The result is being submitted for publication in a peer-reviewed journal, the standard condition for a mathematical discovery to be recognized.

Reservations: what is a “contribution” really?

A few voices qualify the enthusiasm and deserve to be listened to.

First reservation: this is a terminal improvementnot a resolution of the problem as a whole. Erdős had asked the question in its most general form, and finding a family of configurations that beats the grid improves the lower bound, but does not give the actual upper bound. The problem remains open in its most complete form.

Second reservation: the part of the model in the discovery is not yet clearly established. Has the OpenAI team suggested any ideas? Did the model fumble through thousands of hours of computing before arriving at the right construction? The details of human-model interaction are poorly documented at this time, and this is exactly what the peers will seek to clarify.

Third reservation: generalizability. A victory on a problem, no matter how famous, does not mean that the models will now solve Riemann, P vs NP, or the Hodge conjecture. Many open problems persist for structural reasons very different from the unit distance problem.

What this reveals about the trajectory of the models

For AI observers, the result is valuable less for its mathematical content than for what it says about capacity trajectory.

Three lessons stand out. First, the generalization continues to improve : a generalist model that produces an original proof means that the abilities of extended reasoning and combinatorial creativity begin to generalize outside the specifically trained domains. Afterwards, classic benchmarks underestimate models : none benchmarkA standardized test that measures the performance of models on a task (e.g. SWE-bench for coding, MMLU for general knowledge). math (MATH, GSM8K, AIME) does not contain this type of open problem, which means that these benchmarks only measure a fraction of what the models can do. Finally, the human-machine boundary in fundamental research is shifting : if a general model can, under human supervision, produce a real mathematical contribution, the roles in research teams will evolve.

For French players in AI research, it is also a signal. The competition is now played on extended reasoning skills, not on classic benchmarks. Read our comparison of the large 2026 models to see where ChatGPT, Claude, Gemini and Mistral stand on this dimension. And our AlphaEvolve file, the Gemini agent which is shaking up scientific research for the Google counterpart.

What to remember

OpenAI reached a symbolic milestone on May 20, 2026: a general reasoning model produced an original proof refuting a mathematical conjecture that has been open for 80 years. The result is currently being validated by the community, the details of the human-model interaction will be scrutinized, and the exact scope remains to be measured.

The essential signal is not the Erdős conjecture itself. This is because the line between “solving documented problems” and “contributing to open problems” has just been crossed on a verifiable case. This does not mean that models will replace mathematicians tomorrow. This means that a new type of research collaboration, where AI is a contributor and not a simple assistant, becomes technically possible. To watch: the next announcement of this kind, which will probably not be long.

Shredder Smith
Shredder Smith
Shredder Smith is the lead curator and digital persona behind topaitools4you.com, an AI directory dedicated to "shredding" through industry hype to identify high-utility software for everyday users. Smith positions himself as a blunt, no-nonsense reviewer who vets thousands of emerging applications to filter out overpriced "wrappers" in favor of tools that offer genuine ROI and practical productivity. The site serves as a watchdog for the AI gold rush, providing categorized rankings and transparent reviews designed to help small businesses and creators navigate the crowded tech landscape without wasting money on low-value tools.

LEAVE A REPLY

Please enter your comment!
Please enter your name here

Popular Articles