Unit Distance from Your Own Work
In May 2026, OpenAI announced that an internal AI model had disproved a famous conjecture by Paul Erdos, the eccentric mathematician who wandered the world proving theorems and writing papers with anyone who had a good idea. The result was genuine progress on a problem that had been open for 80 years. Mathematicians called it a milestone.
The paper's sole author was listed as: OpenAI.
No individual names appeared in the byline. The blog post announcing the result contained a section titled "## Author" with nothing beneath it. The companion paper on arXiv attributes the work to "OpenAI" as the primary entity, with external mathematicians contributing "Remarks" as a separate appendix. The people inside OpenAI who designed the experiment, chose the problem, wrote the prompt, verified the chain of thought, and decided this was worth publishing are not named anywhere.
This is the cleanest example yet of a phenomenon that has been building quietly in AI research for years. When you join a frontier lab as a researcher, you do not disappear behind the company. You disappear into it. Your contribution becomes indistinguishable from the brand.
Unit distance, in a different sense
The Erdos problem was about unit distances: given n points in the plane, how many pairs can be exactly one unit apart? Erdos believed the maximum was close to what you get from a square grid. OpenAI's model found a better construction. The result was surprising and clever.
The metaphor writes itself. The distance between a researcher and their own contribution is a unit distance problem of a different kind.
In a normal academic setting, that distance is zero. Your name goes on the paper. The work is identifiably yours. You can point to it and say: I did that. Your career advances on the basis of what you have produced.
In the corporate AI lab model, the distance is different. You did the work. You sat with the problem. You had the insight or built the scaffolding that led to it. But when the paper comes out, your name is not on it. The company's name is. The contribution is identifiably theirs, not yours. The distance between you and your own output is close enough to touch and too far to claim.
The irony is that this arrangement borrows its language from the very problem it was used to solve. The researchers are at unit distance from their own work: exactly one corporate entity away.
The disappearing author
This is not unique to OpenAI. DeepMind has tightened its publication policies to protect Google's competitive advantage, restricting what researchers can publish and when. Anthropic publishes system cards attributed to the company rather than individuals. The pattern is consistent across the frontier. As the commercial stakes rise, the individual researcher becomes a liability. A named person can be hired away. A named person can be quoted. A named person can leave. The corporate entity is permanent. The corporate entity is the brand.
Melanie Matchett Wood, a mathematician at Harvard who was part of the external commentary on the OpenAI result, pointed out a related problem: the AI model itself failed to credit ideas from the literature. But the more fundamental credit problem was upstream. The company that owns the model did not credit its own people.
This matters because academic credit is the currency of a research career. If your work belongs to the company, what do you build your reputation on? What goes on your CV? What do you take with you when you leave? The frontier labs are filled with brilliant people who will never be named on the papers they helped produce.
One second take-away
OpenAI solved an 80-year-old problem about unit distances and, in doing so, created a perfect illustration of a different kind of distance: between the researcher and their own work. Close enough to have done it. Too far to be named. The ghost in the byline is not the model. It is the person who asked the question.
Links: An OpenAI model has disproved a central conjecture in discrete geometry (OpenAI) | Remarks on the disproof of the unit distance conjecture (arXiv) | AI just solved an 80-year-old Erdos problem (Yahoo/Quanta) | DeepMind Tightens Control Over AI Research (WinBuzzer)