Key Takeaways
- OpenAI solved Paul Erdős’ 1946 puzzle with n^(1+δ) unit-distance constructions.
- Princeton verified the result, giving AI a 2026 credibility boost in mathematics.
- Tim Gowers says the advance could influence cryptography and proofs beyond geometry.
An 80-year-old geometry riddle finally budged when an OpenAI system stitched together an unlikely construction that beat long-standing expectations. The unit distance problem, posed by Paul Erdős in 1946, asks how many pairs of points exactly one unit apart can exist among n points in the plane; the AI found configurations that grow faster than the classic playbook allowed. Princeton mathematicians checked the work, and heavyweights like Tim Gowers and Arul Shankar took notice. Beyond bragging rights, the result hints at a new kind of collaborator for math, one that uses general inference to push past human heuristics.
AI cracks 80-year-old mathematical mystery with breakthrough solution
Some problems keep nudging at the edges of human patience. The unit distance problem, posed in 1946 by Paul Erdős, asked a deceptively crisp question: with n points on a flat plane, how many pairs can be exactly 1 unit apart. Generations attacked it with grids, symmetry, and grit. Progress came in slivers, never in leaps. Then, quietly, an AI stepped in.
A decades-old problem, solved at last
The classical approach arranged points in square grids, tweaking scale to coax more pairs at distance 1. That method suggested growth just above linear, roughly n multiplied by a factor that barely beats n as it gets large. The field settled around the idea that the best lower bound hovered near n^(1+o(1)), a notch above n, not a stride.
How AI outperformed conjectures
According to researchers involved, an internal model from OpenAI proposed a new family of point configurations that crosses a threshold long thought out of reach. The system produced constructions with at least n^(1+δ) unit-distance pairs, for a fixed δ greater than 0 that does not fade as n increases. That is a genuine polynomial improvement, not a rounding error.
The approach blended geometric insight with advanced algebraic number theory, a surprising toolkit for a spatial counting puzzle. It did not come from a math-specialist engine. Instead, it emerged from a general inference model under evaluation, suggesting broader reasoning capabilities that can navigate across domains when the search space is vast.
Confirmed by experts, celebrated by the field
Independent mathematicians at Princeton University reviewed the AI’s constructions and confirmed the result, per people familiar with the review. Esteemed voices, including Sir Tim Gowers and Arul Shankar, praised the advance as a meaningful step for the field. This is the case where a new lower bound, long static, finally moved because an AI found the right lens.
Implications for mathematics and beyond
What does it mean when a generalist model nudges past entrenched conjectures. For one, it hints at a workflow where machines surface candidate structures and humans stress-test them. In addition to geometry, disciplines like combinatorics, coding theory, and cryptography could see similar collaborations when proofs hinge on rare constructions.
