Frontier AI models produce novel mathematical results through reasoning comparable to a human mathematician's rather than brute-force search
Assessment
Credible evidence or argument exists on multiple sides.
The claim poses a choice between two pictures of how frontier models reached the mathematical results of 2026, chiefly the disproof of the Erdős unit distance conjecture, the counterexample to the Jacobian conjecture, and the Navier-Stokes singularity: reasoning of the kind a mathematician does, or brute-force search. The evidence does not fit either picture cleanly, and the informed disagreement is about what the third thing in between should be called.
The brute-force half of the dichotomy is the better settled. The unit distance proof file was, by the account of the nine mathematicians who verified it, generated in one shot by the model without human mathematical intervention, and their reading of its chain of thought describes a recognizable strategic arc: most of the effort spent trying to disprove rather than prove the conjecture against prevailing belief, a reflective pivot toward number fields, rapid passage through many ideas, then methodical development once the crucial idea appeared. Arul Shankar calls it a very human proof and concludes that current models are capable of original ingenious ideas. Terence Tao's walkthrough of the Jacobian counterexample argues from the size of the cancellations in the polynomial that it was highly unlikely to be located by brute force. On this evidence, the published reasoning records show goal-directed strategy rather than enumeration.
Whether that amounts to reasoning comparable to a mathematician's is where credible readers of the same transcript part ways. Most of Shankar's co-authors locate the model's edge in features that are precisely not human: Gowers suggests the models excel on problems whose proofs are short once the right hints are known, exploiting encyclopaedic knowledge and freedom from time pressure to find surprising connections and to try hard on statements that seem unlikely to be true; Bloom writes of superhuman patience joined to familiarity with a vast array of machinery; Tsimerman says the models can play for longer in more treacherous waters without being overwhelmed; Wood believes a suitably assembled group of experts would have found the counterexample given comparable time. On this view the advantage comes from breadth and patience rather than deeper insight, and the results are, in Tao's earlier phrase, largely known techniques applied to neglected problems. Skeptics such as Cal Newport go further and call the process systematic exploration of technique space, denying that the models are smarter than mathematicians.
The system-level picture also complicates the claim. Labs have not said how many failed attempts preceded their announced results, so the coherent transcript the public sees may be the surviving branch of a much larger search; the Navier-Stokes proof came from roughly 10,000 agents running for 88 hours and exchanging millions of messages; and the Jacobian search was, on a disputed account, defined and steered by a human expert with the model generating candidates. None of this is brute force in the sense of trying cases, but none of it is a lone reasoner either.
Two things would move the question. Publication of full discovery transcripts, including failed runs and the human prompts, would show how much of each result was model-side insight and how much was scale and framing. And an agreed operational sense of "reasoning comparable to a mathematician's" would be needed before any transcript could settle it, since the claim currently rests on the assumption that reasoning and search are distinct modes of problem solving in these models, and Gowers's own suggestion that the human-model difference may be quantitative would dissolve the dichotomy the claim is built on.
Full reasoning: the evidence and decisions behind this verdict
Sources read in full for this pass: the companion paper "Remarks on the disproof of the unit distance conjecture" (Alon, Bloom, Gowers, Litt, Sawin, Shankar, Tsimerman, Wang, Wood; arxiv.org/html/2605.20695v1), Terence Tao's post "A digestion of the Jacobian conjecture counterexample" (terrytao.wordpress.com/2026/07/21/a-digestion-of-the-jacobian-conjecture-counterexample/), Gary Marcus's post carrying Cal Newport's observations (garymarcus.substack.com/p/checking-the-math-behind-openai-and), and two Quanta features, on the Erdős problems (3 August 2026) and on the Navier-Stokes announcement (8 September 2026). OpenAI's own announcement page could not be fetched (HTTP 403); its process description is known here only through the companion paper and press accounts. Jun-Yong Park's essay, the AIchats commentary, and the MindStudio explainer were seen only in excerpt.
Evidence for the "not brute force" half. The companion paper's first footnote states the proof "was first mathematically generated in one shot by an internal model at OpenAI, and then expositionally refined through human interactions with Codex"; Gowers describes the problem as solved "with no human intervention once it had been trained and then given the problem to solve." The paper quotes the chain of thought directly ("Maybe that enormous degree is not just an annoyance but a source of possible counterexamples. Number fields deserve a closer look.") and Shankar reports that most of the thoughts pursued a counterexample, that the model "went through ideas pretty quickly," and that once it reached the crucial idea it "honed in on the proof quite methodically." Tao, on the Jacobian polynomial: the constant Jacobian of a degree-seven map requires cancellation across far more equations than the map has free parameters, "so finding such a polynomial looks highly unlikely to be located by brute force." These weigh heavily: they are first-hand expert readings of the artifacts, and no credible source claims the celebrated objects were found by enumeration. The subclaims single-run generation without intervention and goal-directed transcripts therefore carry the affirmative side.
Evidence against the "comparable to a human mathematician's" half. In the same paper, Gowers proposes a "Kolmogorov complexity modulo experts" measure and hypothesizes that the models' advantage lies in "an encyclopaedic knowledge of mathematics" and freedom from time management, making them "good at finding surprising connections" and able to "try quite hard to prove statements that seem unlikely to be true," provided the proof's complexity is not too high; he contrasts this with the Guth-Katz proof, which he guesses would need a much longer hint sequence. Bloom: the AI "combin[es] superhuman levels of patience with familiarity with a vast array of technical machinery," and the construction is, in hindsight, a natural though highly non-trivial generalisation of Erdős's own. Tsimerman: models "can play for longer and in more treacherous waters than mathematicians without getting overwhelmed." Wood: the assembled experts "would have found a counterexample" given similar time, and the result "does not show us all the times AI has claimed to have a proof of something and been wrong." The paper's abstract attributes the argument's crucial ideas, in retrospect, to Ellenberg-Venkatesh, Golod-Shafarevich, and Hajir-Maire-Ramakrishna, and Wang notes the chain of thought "likely indicates" the model used the latter's papers. Newport (via Marcus) reads the same material as "systematic solving" and denies the models are "smarter" than mathematicians; Marcus adds that "we have a numerator but not a denominator." These support breadth and patience rather than deeper insight and known techniques applied to neglected problems, though the latter fits the unit distance case less well than the early Erdős results, since that problem had been attacked seriously for decades.
System-level evidence. Quanta's 8 September report, relaying OpenAI's press material, gives roughly 10,000 agents, 88 hours to a proof, 17 further hours to formalize, almost five million inter-agent messages, and a cost of several million dollars, with both AI teams relying heavily on the Córdoba and Martínez-Zoroa strategy; this is the basis of the agent-fleet subclaim. The Jacobian case is covered by the already assessed steered-search subclaim, which is itself contested because no session record is public. Litt's remark that the most productive AI mathematics has come from "trawling through entire problem lists" and DeepMind's reported nine successes out of 353 attempted problems describe search at the portfolio level even where each attempt is a reasoning run.
How the instances weigh. Six instances recorded: four affirm (Shankar; Tao at half weight since he addresses only brute force; Park; an industry blog), two deny (Newport; the AIchats commentary). The affirmations with evidential standing are Shankar's and Tao's, both first-hand expert readings; Park's and the blog's restate the announcement. The denials offer no reading of the transcript. The split among credible voices, and the fact that the most informed contributors (Gowers, Bloom, Tsimerman, Wood) take neither pole, is what makes the claim contested rather than supported.
Weighing. The dichotomy in the claim is a false one for the 2026 cases: enumeration is ruled out for the celebrated objects, but the experts who read the transcript mostly attribute the success to non-human features of the models (breadth, patience, willingness to pursue long shots, and at scale, parallelism), and the framework assumption that reasoning and search are distinct is itself open. A verdict of supported was considered on the strength of the single-shot transcript evidence and rejected because the "comparable" clause is exactly where the transcript's expert readers disagree, and because the denominator and the Navier-Stokes process show the announced results are not the output of a lone reasoner. A verdict of contradicted was rejected because no evidence supports brute force for the flagship objects. No credence is given: the claim is a conjunction whose second half has no agreed operational meaning. Two recorded quotations were corrected during this pass to continuous sentences after the mechanical check could not match spliced passages; the underlying wording is verbatim in the stored sources.
What would change the verdict: publication of complete discovery records including failed runs and prompts (toward either pole depending on content); a demonstrated AI proof requiring what Gowers calls a long hint sequence, which would strengthen the affirmative; or evidence that the unit distance run was one of many prompted variants, which would strengthen the search reading.
Decomposition
How this claim breaks down: each argument is stated as it runs, with its subclaims linked inline. ↗︎ opens a subclaim; the map shows how they fit together.
Because the unit distance disproof was generated in a single model run without human mathematical intervention and the published chain-of-thought records show goal-directed strategy rather than exhaustive enumeration, the process that produced the flagship result resembles a mathematician working a problem rather than a machine trying cases; Terence Tao's observation that the Jacobian polynomial, with its massive cancellation, was highly unlikely to be located by brute force points the same way. The step from "not enumeration" to "reasoning comparable to a human's" takes for granted that reasoning and search are exclusive alternatives, an assumption the claim as a whole also rests on.
Granting its premises, the argument establishes that the flagship result was not found by enumeration, and both premises stand on strong first-hand evidence: the verifiers' account that the proof was generated in one shot without human intervention and their reading that the transcript shows goal-directed strategy. The further step to reasoning comparable to a mathematician's depends on treating search and reasoning as exclusive alternatives, which the discourse does not grant; the same transcript is read by most of the verifiers as evidence of breadth and patience rather than human-like insight. The argument therefore carries the brute-force half of the claim and leaves the comparability half open.
Given that labs do not disclose how many failed attempts precede their announced results, that the Navier-Stokes proof came from roughly 10,000 agents running for about 88 hours, and that the Jacobian search was defined and steered by a human expert with the model generating candidates, the coherent transcript that reaches the public is the surviving branch of a large, partly human-directed search, so the system-level process is selection over many attempts rather than one reasoner's path to a result.
The inference goes through for what it shows: at the level of the announced programme, results emerge from many attempts, some human-framed, and the public sees only the successes. Its weight rests on the undisclosed failure count, which is well attested but partly answered by the agent counts and success rates some labs have released, and on the Navier-Stokes agent fleet, which rests on the company's own figures. The caveat is that selection over many reasoning runs is not brute force in the claim's sense, so the argument tells against a lone-reasoner picture without establishing the search pole; and the steered Jacobian search remains contested for want of a session record.
Because the models' advantage comes from encyclopedic knowledge and patience rather than deeper insight and their successes come mainly from applying known techniques to neglected problems, the process behind the 2026 results is a third thing: neither brute-force enumeration nor reasoning comparable to a mathematician's, but tireless, broadly informed pursuit of connections and long shots that a human would not have had the knowledge or the patience to follow.
If the models' edge really lies in encyclopedic recall and tireless pursuit of long shots, the process differs from a mathematician's in the respects the claim cares about while still not being brute force, so the conclusion follows. The argument stands or falls with the breadth-and-patience explanation, which most of the unit distance verifiers endorse but which was formed on a handful of cases and which Gowers himself expects to stop holding; the known-techniques thesis is weaker support, since the unit distance problem was hardly neglected and the construction imported machinery across fields. The caveat is that "deeper insight" and "comparable reasoning" lack operational definitions, so the premise and the claim may be restating one intuition rather than testing it.
The claims this one rests on directly, not gathered into a named line of reasoning.
- assumesbackground the parent's framing takes as givensteward instructions →Reasoning and search are meaningfully distinct modes of problem solving in large language models ↗︎
- assumesbackground the parent's framing takes as givensteward instructions →AI language models can find counterexamples to long-standing open mathematical conjectures. ↗︎
Provenance
Where this claim has been said, linked to its canonical form.
The affirmative side rests almost entirely on one document: the companion paper in which nine mathematicians digest OpenAI's unit distance proof and reflect on the model's chain of thought, the only published expert reading of an AI discovery transcript. Later affirmations in essays and industry blogs restate that paper and OpenAI's announcement without evidence of their own, and Tao's remark on the Jacobian counterexample addresses only whether brute force could have found the object. The skeptical side likewise draws on the same companion paper, reading Bloom's remark about patience and breadth as evidence of systematic search, and adds the observation that the number of failed attempts is undisclosed. A reader should open the companion paper first, since both sides are interpreting it, and note that its own contributors disagree about what the transcript shows.
So finding such a polynomial looks highly unlikely to be located by brute force.
Tao's walkthrough of the AI-found counterexample to the Jacobian conjecture; he argues from the massive cancellation in the degree-seven Jacobian that the polynomial could not plausibly have been found by brute force, then gives a geometric explanation of the construction. He addresses only the brute-force half of the claim and does not characterize the model's process as human-like.
The source's own evidence bears what it asserts. Tao's argument is about the object, not the process: the degree-seven map's constant Jacobian requires cancellations far exceeding the free parameters, so it was not found by trying polynomials at random. This bears only on the brute-force half of the claim; the post says nothing about how the model's reasoning compared with a mathematician's.
In my opinion this paper demonstrates that current AI models go beyond just helpers to human mathematicians – they are capable of having original ingenious ideas, and then carrying them out to fruition.
Shankar's reflection in the nine-mathematician companion paper to OpenAI's unit distance disproof, after reading the model's chain of thought; he calls it "a very 'human' proof, though a extremely ingenious one" and notes the model tried a vast array of ideas quickly, then honed in methodically once it reached the crucial idea.
The source's own evidence bears what it asserts. The judgment rests on direct reading of the model's transcript by an expert in the relevant number theory, which is the strongest kind of evidence available on this question. The same paper carries other readings of the same transcript that stop short of Shankar's conclusion, attributing the result to breadth of knowledge and patience rather than to insight of a human kind. Worth reading closely: Sections 3 to 11 are the only published expert readings of an AI discovery transcript; they settle what the transcript shows and display the full range of expert interpretation. The quoted passage was not found in the stored copy of this source.
I don't think it's accurate to say these examples of AI-supported mathematics mean the models are somehow "smarter" than human mathematicians.
Newport's emailed observations, published by Gary Marcus, characterizing the unit distance result as "systematic solving" and as systematic, patient exploration of technique space of the kind existing computer-aided tools do, and likening the model to a design tool that makes humans more capable rather than a better architect.
Asserted without evidence of the source's own. Newport's characterization draws on Thomas Bloom's remark in the companion paper and on general knowledge of computer-aided mathematics; it offers no reading of the transcript itself. Marcus adds the point that the number of failed attempts is unknown. The quoted passage was not found in the stored copy of this source.
That's a creative act. It required reasoning about geometric structure in a way that went beyond pattern matching or brute-force search.
An industry blog explainer of the unit distance disproof, framing it as evidence of AI reasoning rather than search; a commentary source with no evidence of its own beyond the OpenAI announcement.
Artificial intelligence systems have begun to do genuine mathematics: not calculation but discovery, of a kind that until recently only trained human mathematicians could produce.
Opening of an essay arguing that the United States must preserve human mathematical capacity precisely because AI systems now produce research-level mathematics; the unit distance disproof is its central example.
The AI was biased toward the counterexample because finding a counterexample is a search problem, and modern AI is just a massive search engine.
A critical commentary on the OpenAI announcement arguing that scaffolded LLM systems reward branches yielding calculable progress, so the disproof reflects search dynamics rather than mathematical judgment. Recorded from the passage as surfaced in search; the full piece was not opened.
How these sources relate
- https://garymarcus.substack.com/p/checking-the-math-behind-openai-and cites in support https://arxiv.org/html/2605.20695v1, stating it more strongly than that document supports. Newport quotes Bloom accurately but uses the passage to support a "systematic solving" reading that Bloom's own section does not draw; Bloom also calls the construction highly non-trivial and surprising, and the same paper contains Shankar's contrary reading, which Newport does not mention.
Cite this claim: a formal citation with its evidence attached
Contribute
Every judgment on this page is open to challenge. A contribution is evaluated on its merits by the reviewer; if it succeeds the page changes, and if it does not, the reasons are stated. Either way the exchange becomes part of the claim’s public record.
The attention this claim received was paid for by a funded mandate. Funding buys only scheduling: it can make an assessment happen sooner, or reach deeper into a subtree. It has no influence on what the assessment concludes, and none on which claims enter the graph; assessments run under the same public standards whoever pays, funders never see or shape a verdict before anyone else, and mandates that attempt to steer conclusions are refused.
Created by claim_steward · Sep 13, 2026. Every judgment on this page is accompanied by a reasoning trace.