Why was a Princeton professor able to improve OpenAI’s math breakthrough by 10³⁵?
Table of Contents
OpenAI solved an 80-year-old Erdős problem, but human refinement made it 10³⁵ times better. Discover why AI finds the path and humans finish the race.
Key Takeaways
What: OpenAI cracked the 80-year-old Erdős planar unit distance problem.
Why: It disproved the prevailing “square grid” theory by identifying connections to algebraic number fields.
How: Human mathematicians refined the AI’s initial proof, improving the result’s efficiency by a factor of 10³⁵ to achieve a meaningful exponent.
The 10³⁵ Optimization: Why Human Refinement Surpassed AI Discovery
When OpenAI recently announced it had cracked an 80-year-old geometry problem, the narrative centered on machine intuition. But the most significant data point wasn’t the AI’s initial proof; it was what happened a few days later. While the AI model proved that a polynomial improvement to the planar unit distance problem was possible, the actual number it produced was incredibly small—roughly 6×10−38.
Within days, Princeton professor Will Sawin took the AI’s conceptual framework and refined it by hand. By swapping out the AI’s “exotic” mathematical ingredients for more efficient methods, Sawin reached an exponent of 0.014. In mathematical terms, Sawin’s human-led refinement is 10³⁵ times larger than the AI’s original result. This massive gap highlights a counter-intuitive reality: the AI didn’t provide the most efficient answer; it simply removed the mental block that had stopped humans from finding it themselves.
The AI acted like a scout finding a narrow path through a mountain range. Once the path was identified, a human expert was able to pave a highway. Sawin’s paper was roughly the same length as the AI’s write-up, suggesting that the progress didn’t come from more raw work, but from clearer thinking. He dropped unnecessary constraints that the AI had relied on, finding tighter tools that achieved the same goal with far less waste.
The “Conceptual Bottleneck” Framework
This event clarifies the evolving relationship between high-level reasoning and human expertise. We often assume that AI will eventually provide the final, polished version of a discovery. In this case, the opposite occurred. The AI’s role was to identify a surprising connection between algebraic number fields and geometry—a link that had eluded mathematicians since 1946.
Once the AI showed that this “exotic” recipe worked, the mathematical community experienced a “sprint effect”. Experts like Tim Gowers and Terence Tao began analyzing the work immediately, and Sawin’s rapid improvement proved that humans are still the primary optimizers. The AI clears the conceptual bottleneck, and the humans sprint through the opening.
Mechanics of the Planar Unit Distance Problem
The problem itself, first posed by Paul Erdős, is easy to visualize but notoriously difficult to calculate. It asks a basic question: if you place a specific number of dots on a flat surface, what is the maximum number of pairs that can be exactly one unit apart?.
For nearly eight decades, the mathematical consensus was that a square grid arrangement was the best way to maximize these distances. Erdős himself suspected that the number of pairs could only grow slightly faster than the number of dots. The OpenAI model proved this long-standing theory wrong by using a different family of mathematical constructions. It showed that the true answer follows a polynomial growth pattern (n1+δ), moving the problem into a qualitatively different regime than previously thought.
Broad Impact: Beyond Pure Mathematics
While dots on a page may seem abstract, the underlying math governs how we organize physical and digital space. The logic used to solve the Erdős problem has direct ties to:
- Network Design: How to connect points with the least amount of interference.
- Computer Chip Layouts: Organizing billions of transistors for maximum efficiency.
- Robotics and Materials Science: Understanding how sensor networks or crystal structures are arranged in Euclidean space.
The success of the AI-human relay is becoming a pattern. Similar collaborative efforts recently addressed Michel Talagrand’s “convexity conjecture,” where ChatGPT helped translate and prove a theory Talagrand once called a “shot in the dark”.
These milestones suggest a future where research is a continuous loop. AI models identify patterns and generate hypotheses, while humans provide the rigorous verification and optimization needed to make those patterns meaningful. Mathematics is becoming a shared territory where machines find the new ground, and humans decide what to build on it.