An Org Chart Dressed as a Theorem

In the second room of my flat sleeps Ugo, and on one wall of that room sleep my curiosities: a wall of post-it notes, one for every thing that interests me and arrived at the wrong moment. In June I pinne there an article from Spektrum der Wissenschaft, by Manon Bischoff. Yesterday morning I finally took it down. Ugo read the gesture as the opening of a game; in a sense it was.

The article described an elegant chain. A model from OpenAI had solved an Erd■s problem in geometry, and the techniques of that solution had inspired human mathematicians to refute an old conjecture on sum-product sets, if only for algebraic and p-adic numbers. The machine opens a door, humans walk through and go further: the good version of the story. Then July arrived with the version that makes headlines: GPT-5.6 Sol Ultra is said to have proved the cycle double cover conjecture, open since the 1970s, in under an hour. One hour against fifty years. The ratio is built to be quoted.

What OpenAI actually describes deserves a careful reading. Sixty-four sub-agents running in parallel, deliberately kept unaware of the most promising line “to encourage independent thinking”; adversarial agents sieving every step for the standard failure modes; a 700-word prompt orchestrating the whole. Anyone who has shipped a complex system recognizes the drawing: independent teams against groupthink, a red team attacking results before release, a job description written with care. What the release describes is an org chart dressed as a theorem: management capability, applied to a language model. The celebrated intelligence lives in the process, and the process was written by people. 

The proof itself, indeed, holds. Thomas Bloom, who examined it, calls it short, elementary, beautiful; and he adds the detail that weighs most: it could have emerged in the 1980s, and it ignores the relevant literature from at least 1983. Timothy B. Lee, one of the more sober engineering readers of this field, has shown that the problem played to the model’s strengths: a vast literature to recombine, a verification that can be mechanized. Among developers, meanwhile, circulates the question the press releases omit: no one can confirm the experiment, because the model is proprietary and a replication costs what it costs. That a result can be true while nobody can narrate its genesis, we have seen already: blind thought.

Here the question stops being mathematical and becomes political. If discovery requires a fleet of sixty-four agents on a frontier model, mathematics becomes a function of the compute budget: which conjectures deserve the Sol Ultra hour is decided by whoever pays the electricity bill. Discovery gets privatized while verification gets socialized: Bloom checked the proof for free, as the reviewers of every next announcement will. It is the oldest business model in the industry: capital collects the headline, the community performs the work.

The honest concession must be made. The theorem stands proved, and the June chain, where the machine’s solution taught humans something and the humans went further, is progress without quotation marks. When the result returns to the commons, techniques readable, precedents cited, the objection falls.

The July announcement, however, sells another commodity: the machine smarter than the mathematicians, and by transitivity its owner smarter than everyone. Here a recurring trait of the techeconomy’s barons returns, the billions in the pocket and the aesthetic restyling carefully curated: the need to sit at the table of the smartest people in the room by paying the entrance fee, the invitation having failed to arrive. The attention they receive is proportional to the money they move, and to nothing else; the mathematical announcement converts that currency into another one, prestige. If the result could have come off an ordinary desk forty years ago, what has been demonstrated is the invoice.

Your field will have its Sol Ultra moment, with its own ratio built to be quoted, and the value of this story is the pair of questions it leaves ready for that day: who owns the org chart, and who pays for the verification. If the answer is twice the same company, what you are reading is an org chart dressed as a theorem, marketing with lemmas inside.