On October 6, 2026, OpenAI published 719 mathematics manuscripts written by an unreleased model. This case follows one of them. The question: colour every point of a flat plane so that no two points exactly one unit apart share a colour — how few colours do you need? Posed in 1950, it stood at “4 to 7” for 68 years, until a biologist with a computer raised the floor to 5 in 2018. The new paper claims five are not enough: the answer is 6 or 7. The surprise is the method. Every earlier advance came from a computer checking a specific pattern of points. This one builds no pattern at all; it is 62 pages of theory from a different field. And within four days the result was already being used: a specialist posted a related paper early because of it, the 2018 pioneer rallied his group to understand it, and an outsider extended it. No expert has verified it. The case is what a method like this amplifies — and the condition under which it amplifies anything.
The problem needs no training to state. Colour every point of the plane; the only rule is that two points exactly one unit apart must differ.[17] Edward Nelson asked it in the autumn of 1950 as a student at the University of Chicago, and by November a fellow student, John Isbell, had the upper bound: seven colours always work, using a pattern of hexagons.[4] Nelson had the lower bound, four. And there it stayed. For 68 years the answer was known only to be 4, 5, 6 or 7.
In April 2018 the floor moved. Aubrey de Grey — a biologist who described himself as an amateur mathematician — constructed a pattern of 1,581 points that cannot be coloured with four colours, and confirmed it by computer.[5][6][18] A community project then shrank the pattern, to 509 points by 2020, and set itself a further goal: find a pattern that forces six colours.[7][8] It never did. Twelve days before OpenAI’s release, de Grey asked the group whether anyone had started using AI to find such patterns.[8]
The paper released on October 6 claims the next step: no five-colouring exists, so the answer is 6 or 7.[1] It says so without overreach — “The remaining alternatives six and seven are unresolved.”[1] The core statement has been checked by machine in the proof language Lean, and that statement is short enough for a lay reader to follow: there is no colouring with five colours in which points at distance one always differ.[2] The paper was released under an open licence, and it is not among the three papers OpenAI withdrew or the fourteen it revised the next day.[3]
What makes it more than one more result is how it was done. De Grey’s advance, and every refinement after it, was a search: build a finite pattern, have a computer confirm it cannot be coloured. This paper builds no pattern. It proves that if any five-colouring existed, however wild, a well-behaved one would too — a bridge built from ergodic theory, a different branch of mathematics — and then that no well-behaved one can exist.[1] A finite uncolourable pattern must therefore exist, but the paper does not produce it: “The graph and the positive constant here are existential.”[1] In 2018 the human used the computer for brute force. In 2026 the machine used theory.
The claim that a method amplifies human work is usually a promise about the future. Here there is a record, and it is four days long. On October 6 a member of the 2018 community project posted the news to the group’s thread. On October 8 de Grey wrote: “OK guys, let’s have a proper stab at understanding this paper – if anyone can do so it’s us.”[8] A specialist replied that he had been attempting something similar “and was practically certain that this is possible for the measurable case.”[8]
The same day, James Davies posted a paper proving the same kind of bridge in four or more dimensions by a different method, with the hope of extending it “possibly even” to two — the plane problem itself. It cites the OpenAI paper and explains its own timing: “Due to the very recent release of [53] by OpenAI, we have posted this paper slightly earlier than we would have liked.”[9] Davies’s paper carries its own detailed statement of how ChatGPT helped him. Two days later a self-described non-mathematician, working with a different AI model, posted a conditional extension to three dimensions and asked the group to check it.[8] One person reports rebuilding the machine-checked proof from scratch without errors — self-reported and unrefereed.[14]
None of this is verification, and the record contains its opposite. One reader called the proof “totally unbelievable alien math.”[11] A mathematician who studied it says the paper’s central definition is not new and that closely related work from 2014 goes uncited — “if I were refereeing the paper I would insist they cite” — while crediting the bridge theorem as the new idea.[10] The reference site for this family of problems still lists the answer as 5 to 7, with no proof claims recorded.[12] OpenAI is open about the limits itself: its repository says some results without a machine check “could have issues,” logs withdrawals and revisions publicly, and labels the review of whether its machine-checked statements match its papers as “unchecked.”[3]
Edward Nelson poses it at the University of Chicago and finds the lower bound, four. John Isbell finds the upper bound, seven, within weeks.[4]
The QuestionDavies posts a related proof for four or more dimensions, citing the OpenAI paper as the reason for the timing.[9]
The First UseDe Grey rallies the group; a specialist says he had tried the same route; an outsider posts an extension.[8]
AbsorptionFrom 1950 to 2018 the lower bound did not move. It took an amateur with a computer to move it once.[4][5]
De Grey’s 2018 pattern, shrunk by a community project to the 2020 record. A six-colour pattern was the stated goal and was never found.[5][7]
No pattern is constructed. The proof is theoretical and says the pattern it implies is “existential.”[1]
OK guys, let’s have a proper stab at understanding this paper – if anyone can do so it’s us. — Aubrey de Grey, who raised the bound to 5 in 2018, October 8, 2026 [8]
| Dimension | Evidence |
|---|---|
| Quality (D5) Origin · 86 | A 76-year-old problem: the floor rises from 5 to 6, with 6 or 7 left open — the paper’s own words. The core statement is machine-checked and five lines long.[1][2]A Limit Moved, Not a Problem Closed |
| Employee (D2) L1 · 82 | The 2018 author rallies his group to understand it; a specialist says he was attempting the same route; a related human paper is posted early because of it.[8][9]The Pioneers Regroup |
| Operational (D6) L1 · 80 | No finite pattern is built; the proof is 62 pages of ergodic theory and geometry, and the pattern it implies is “existential.” The model that wrote it is unreleased.[1][3]Theory Where Search Was Expected |
| Regulatory (D4) L2 · 68 | No expert verification; the reference site still lists 5 to 7; a reader flags uncited 2014 work; OpenAI itself labels its statement review “unchecked.”[3][10][12]Nobody Has Ruled |
| Customer (D1) L2 · 60 | The question is one a child can ask. The answer is, to one reader, “totally unbelievable alien math” — and to an outsider with an AI model, something to extend.[8][11]Legible to Some |
The cascade originates in D5 (Quality): one result, on a problem anyone can state, whose machine-checked claim is five lines long.[1][2] It propagates to D6 (Operational), the method — theory where every predecessor used search, with no pattern to re-run and no model to rerun it on,[1][3] and to D2 (Employee), the people who work on the problem: a pioneer regrouping his community, a specialist publishing early, an outsider extending the result.[8][9] It lands in D1 (Customer), the wider audience the result has to be legible to — one reader’s “alien math” is another’s new tool[11] — and in D4 (Regulatory), the institutions that decide what counts as proven, none of which has yet ruled.[10][12] D3 (Revenue) is not scored: nothing in the record ties this result to money.
Two precedents mark the range of outcomes. When a machine solved protein folding in 2020, a problem open for half a century, the lasting effect was not the released structures but what others could then do: by the developer’s own count, more than 3 million researchers and 35,000 citing papers five years on.[15] When a mathematician posted a claimed proof of a famous conjecture in 2012 using methods nobody else could follow, the field never accepted it, and nothing was built on it.[16] The difference between the two was not correctness alone. It was whether other people could pick the method up. The honest limit: this case rests on a proof no expert has verified, and the amplification it reports is four days old.
-- UC-328: Five Colours Are Not Enough: 6D Amplifying Cascade
-- Hadwiger-Nelson problem, posed by Edward Nelson autumn 1950: fewest colours for the plane so points at distance one differ. Bounds 4 to 7 from 1950 (Nelson, Isbell). Lower bound 5: de Grey, Apr 2018, 1,581-point computer-checked graph; Polymath16 reduced to 509 points (Parts, 2020). OpenAI release Oct 6 2026: 719 manuscripts, 372 families, unreleased model, about 4,000 problems posed, 300 of 719 top-line results formalized. Family 158: 62-page paper dated Sep 23 2026 claims no proper five-colouring; answer is 6 or 7. Method: transfer theorem (any k-colouring iff weak measurable k-colouring, via ergodic theory) plus geometric obstruction; no finite graph constructed. Lean statement formalized; not withdrawn or revised. Oct 8: Davies arXiv 2610.12301 cites it and is posted early; de Grey rallies Polymath16. No expert verification as of Oct 11.
FORAGE five_colours
WHERE long_open_problem_bound_moved = true
AND method_differs_from_all_prior_progress = true
AND downstream_use_within_days = true
ACROSS D5, D6, D2, D1, D4
DEPTH 3
SURFACE five_colours
DIVE INTO method_as_reach
WHEN core_statement_machine_checked = true
AND expert_verification_pending = true
TRACE reach_amplification_cascade
EMIT five_colours_signal
DRIFT five_colours
METHODOLOGY 86
PERFORMANCE 34
FETCH five_colours
THRESHOLD 1000
ON EXECUTE CHIRP signal 'A 76-year-old problem a child can state: the proven minimum moves from 5 to 6, claimed by an AI model using theory where every human advance used computer search. Within four days a specialist posts a related paper early because of it and the 2018 pioneer rallies his group to understand it. No expert has verified the proof, and the model is unreleased. The method amplifies only as far as other people can pick it up.'
SURFACE analysis AS json
Runtime: @stratiqx/cal-runtime · Spec: cal.semanticintent.dev · DOI: 10.5281/zenodo.18905193
In 2018 a human moved the limit by having a computer check a 1,581-point pattern. In 2026 a machine claims to move it again with 62 pages of theory and no pattern at all. The division of labour people assumed — humans reason, machines search — did not hold.[1][5]
The evidence is not the proof; it is the four days after it. A specialist published early, the pioneers regrouped, an outsider extended the result. A method amplifies only as far as others can pick it up.[8][9]
The paper claims 6 or 7, not an answer, and says so. The honest unit of progress here is the bound — and the bound is what a method like this can be judged on.[1]
Primary sources are OpenAI’s paper, its machine-checked statement and its repository log; the community thread where the 2018 authors responded; the follow-on paper; and the original papers for the history. All were read directly. The proof has not been verified by any expert as of October 11, 2026, and this case does not claim it is correct. Sources 17 and 18 are interactive companion visuals made for this case.
This case follows one machine-written result from claim to first use, and marks exactly where verification stops. The same reading applies to any AI-produced work: what was checked, by whom, and what others can build on it. Book a discovery call to map where checking sits in your own pipeline.