Most early AI-driven Erdős problem solutions came from hobbyists using public LLMs rather than corporate labs.
Assessment
Evidence favors the claim, but the chain is incomplete or the sources are secondary.
The claim describes the first wave of AI-assisted resolutions of Erdős problems, roughly December 2025 to the OpenAI unit-distance announcement of 20 May 2026, and holds up on the public record. The community wiki maintained under Terence Tao's Erdős problems project, which logs every AI contribution posted to erdosproblems.com along with the labs' own announcements, lists about 68 full solutions in that window. Roughly three quarters of them were obtained by people outside the AI labs prompting publicly available systems (chiefly GPT-5.2, 5.4 and 5.5 Pro, with Gemini, Claude and Harmonic's Aristotle prover), and about a quarter by corporate systems not available to the public: OpenAI's internal model, DeepMind's prover agents and its Aletheia system. The outsiders were mostly amateurs, students and non-academic professionals such as Liam Price, Kevin Barreto, Wouter van Doorn, Przemek Chojecki and Boon Suan Ho, with a few professional mathematicians among them.
The picture is one of count, not of weight. The most significant early results, above all the counterexample to the unit distance conjecture and the batch of solutions OpenAI announced in April 2026, came from internal corporate models, and DeepMind's autonomous agent resolved nine formal statements at a few hundred dollars each. Many of the hobbyist solutions were to obscure problems, and a fair number turned out to duplicate forgotten literature. Read as a statement about numbers, the claim is well supported; read as a statement about where the mathematically important advances came from, it would not be. The only recorded assertion of it is Thomas Bloom's observation as reported by Quanta Magazine, but the tally that supports it is independent of that report.
Full reasoning: the evidence and decisions behind this verdict
The single instance is Quanta Magazine's paraphrase of Thomas Bloom (www.quantamagazine.org/why-the-legendary-erdos-problems-are-falling-to-ai-20260803/): despite attention from OpenAI, Google DeepMind and startups, "most of the new results had come from hobbyists and undergraduates using publicly available LLMs, not from corporate labs using more advanced internal models," a state of affairs the article says changed on 20 May 2026. The article offers no count, so the claim was checked against the primary tracking record, the wiki "AI contributions to Erdős problems" (github.com/teorth/erdosproblems/wiki/AI-contributions-to-Erd%C5%91s-problems, data frozen 30 June 2026), read in full.
Tallying the wiki's full-solution rows (green outcomes) across its four primary-contribution sections, and excluding rows that are only new proofs of already-known results, gives 68 rows dated before 20 May 2026. Classifying by system: 49 used publicly available tools (GPT-5.x Pro or Thinking, Codex, Gemini 3.x Pro, Claude Opus, Aristotle), 16 used systems not publicly available (OpenAI internal model on problems 960, 987, 990, 1091, 1141 on 9 April, 1014, 997, 846; DeepMind prover agent on 125, 741, 152, 846, 26; Aletheia on 1051, 397, 659, 1089), two mixed rows involving AlphaEvolve alongside public models, and one custom system. Extending to 30 June 2026 gives 56 public against 19 internal. Three of the public rows were Aristotle alone (897, 1026, 966), which Harmonic staff may have run; three involved professional mathematicians (Tao on 380 and 347, Sawhney and Sellke on 848). Removing all six still leaves about 43 outsider solutions against 16 corporate ones. The wiki also records that DeepMind's January 2026 Aletheia results and its May 2026 agent results are included, so the corporate count is not artificially low; even crediting DeepMind's self-reported nine and OpenAI's announced batches in full, the corporate share stays below a third.
The "hobbyist" characterization fits most but not all of the outsider rows: the named humans are predominantly amateurs, students and non-academic professionals (Price, Barreto, van Doorn, Chojecki, Boon Suan Ho, Turturean, Sothanaphan, Bhalla), with a minority of professional mathematicians. The supporting subclaim on DeepMind's nine of 353 is assessed as supported and bounds the largest corporate autonomous effort at about four fully resolved catalogue problems.
Weaknesses: the wiki is a community record with acknowledged selection effects and disclaims being a benchmark; several outsider solutions were later found in the literature (section 1(b)), though the wiki still counts them as full solutions and Bloom's remark concerned "new results" in the same sense; "early" is read as the period before 20 May 2026, as the Quanta article frames it; and the claim would be false if weighted by mathematical significance rather than by count. Nothing found denies the claim. What would change the verdict: evidence that corporate labs' announced results substantially exceed what the wiki records for the period, or a reading of "solutions" restricted to results the community judged significant.
Decomposition
The claims this one rests on directly. ↗︎ opens a subclaim; the map shows how they fit together.
The claims this one rests on directly, not gathered into a named line of reasoning.
- supportsthis provides evidence for the parentsteward instructions →DeepMind's AlphaProof Nexus agent autonomously resolved 9 of 353 formalized open Erdős problems at a few hundred dollars each. ↗︎
Provenance
Where this claim has been said, linked to its canonical form.
most of the new results had come from hobbyists and undergraduates using publicly available LLMs, not from corporate labs using more advanced internal models
Quanta feature on AI and the Erdős problems; the sentence reports the surprise of Thomas Bloom, who runs erdosproblems.com, that despite attention from OpenAI, Google DeepMind and startups, most new results up to May 2026 had come from hobbyists and undergraduates using public LLMs. The article's next sentence marks OpenAI's 20 May 2026 unit-distance announcement as the point where that changed.
Asserted without evidence of the source's own. The article reports the observation as Thomas Bloom's and offers no tally of its own; its examples (Barreto and Price, van Doorn, Chojecki's circle) illustrate the pattern, while its own account of DeepMind's January and May results and OpenAI's April and May results supplies the corporate side. The community wiki of AI contributions, which the article does not cite for this point, bears the observation out by count.
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 extractor · Sep 13, 2026. Every judgment on this page is accompanied by a reasoning trace.