The Referee Is the Bottleneck: OpenAI's 722 Manuscripts and the New Economics of Proof
On Monday, October 6, OpenAI published 722 AI-generated research-mathematics manuscripts to a public repository — grouped into 372 families of related results, grown out of an evaluation of roughly 4,000 open problems, averaging about three hours of ChatGPT Pro thinking compute per result.12 The drop includes papers, formal proof artifacts in Lean, and ten abridged reasoning summaries; the internal model that produced them remains unreleased.3 The claims travel faster than the proofs: a quasi-Riemann hypothesis result, a no-Siegel-zeros theorem, integer multiplication faster than n log n, and progress reported across 90 of the top 500 open problems in mathematics.4 Sam Altman called it “a new era of discovery.”4 Levent Alpöge — Harvard number theorist, the man on the other side of September’s Navier-Stokes priority dispute — went further: “It’s obviously the most significant moment in mathematical history,” before noting “sad stories” of researchers getting scooped along the way.5 The Institute for Advanced Study’s advisory group, which OpenAI consulted on the release, has been blunt in both directions: engage the mathematics, but do not use mathematical releases as marketing.6 Three weeks ago this blog asked whether September’s announcement was solved as written.7 Now the writing has shipped. Four frames on the week mathematics acquired a queue.
Five Whys — why the awe arrives pre-mixed with irritation
Surface issue: the reaction split cleanly — genuine interest in the formalized results, open irritation at the framing.
- Why the irritation? Because verification is uneven: many manuscripts carry Lean formalizations, but others do not, and OpenAI itself warns the unformalized results may contain errors.3
- Why does that sting when some results are machine-checked? Because the release splits into two epistemic classes — statements whose proofs a kernel settles mechanically, and statements whose correctness, novelty, and significance still run on human judgment.
- Why does the human half lag? Because refereeing is volunteer labor; the field’s review capacity is a small, fixed guild.
- Why didn’t that capacity ever scale? Because it never had to. Producing candidate proofs was the bottleneck, so serious claims arrived at the rate careers are made.
- Why does that regime fail now? Because the generator got industrial — three hours of compute per result — while the referee stayed artisanal.
Root cause: the binding constraint has moved from finding proofs to adjudicating them. The repository is not a library. It is an intake queue.
Explanation Ladder — from repo drop to epistemics
- High school: a computer wrote hundreds of math papers, and for many of them another program checked every single step.
- College (responding): that checker is Lean — a proof assistant whose kernel verifies a formal artifact against the stated claim. Where the artifact exists, correctness is no longer a matter of opinion.
- PhD (responding): but the kernel checks the proof against the statement; it cannot check that the statement matters or is new. Formalization certifies validity, not value — and the unformalized remainder inherits the old referee economy wholesale.8
- Philosopher (responding): mathematics was the discipline where certainty was expensive to earn and cheap to store once earned. Automation has collapsed the front of that pipeline while leaving its gate human.
- Gigabrain (responding): every trust system meets this moment eventually. When the manufacture of claims becomes cheap, the scarce good is no longer proof but a person with standing and time to look. The repository does not settle mathematics. It queues it.
Analogy Transfer — how mature claim-markets survive their producers
Structural form: a producer can now generate serious claims faster than standing institutions can certify them.
| Domain | Their version of the problem | Their mechanism |
|---|---|---|
| Assay offices / hallmarking | Makers stamp their own precious metal | Independent office assays; public mark means third party, not producer |
| Anti-doping | Athletes control their own bodies | Accredited labs certify; samples retained for years and re-tested as methods improve |
| CVE registries | Anyone can claim a vulnerability | Numbered registry; staged adjudication by assigned authorities |
Disanalogy check, mandatory: assay and doping tests converge on physical samples; mathematical novelty judgments do not converge mechanically, and Lean covers only part of the corpus. What survives the transfer is the pairing of independent certification with retention for re-verification — and note that OpenAI preserving manuscript versions as corrections arrive is exactly the retention move.2 The transferable demand: publish the certified subset as a visibly marked class, and let outside referees — not the producer — apply the marks.
Counterfactual — what if the model had shipped instead?
Actual history: artifacts released, weights withheld. Minimal intervention: release the model publicly on October 6 instead.
| Step | Consequence | Confidence |
|---|---|---|
| 1st order | Everyone runs it; claim production explodes well past 722 | Near-certain |
| 2nd order | Referee workload becomes unbounded; trust in AI-produced math degrades as a class; scooping conflicts multiply | Probable |
| 3rd order | The field routinizes bulk verification and journals reorganize around formal artifacts | Speculative |
Equilibrium check: competitive pressure pushed toward withholding regardless — the weights are the asset, the manuscripts are its advertisement. Verdict: contingent. The curated-corpus release is what holds referee workload at “months” rather than “impossible.” Call it load-shedding, not stinginess — while noting it is also marketing, which is precisely what the IAS group warned against.6
What the frames show together
Mathematics this week became the first discipline to experience mass production of serious claims. Five-whys: the constraint moved from production to adjudication. Ladder: machine-checked validity and human-judged value have decoupled. Analogy: mature claim-markets answer floods with third-party marks and retained evidence. Counterfactual: the curated release is a throttle, and a defensible one. What to watch: a named mathematician outside OpenAI reviewing specific manuscripts on record;8 formalizations landing on the unformalized remainder;3 and the scooping disputes becoming mathematics’ first machine-speed labor conflict.5 “Solved, as written” was September’s question.7 The repo is the writing. The reading starts now.
-
https://openai.com/index/sharing-ai-progress-in-mathematics/ ↩
-
https://superpowerdaily.com/posts/openai-722-math-manuscripts-with-verification-still-uneven ↩ ↩2
-
https://www.theneurondaily.com/p/openai-s-ai-produced-722-math-manuscripts ↩ ↩2 ↩3
-
https://www.latent.space/p/ainews-quasi-riemann-hypothesis-openai ↩ ↩2
-
https://finance.biggo.com/news/269cc3f5-7ade-42e3-8f0d-1630fa061477 ↩ ↩2
-
https://logicinczo.github.io/semantic-thinking/2026/09/solved-as-written-openai-navier-stokes/ ↩ ↩2
-
https://explainx.ai/blog/openai-722-math-manuscripts-github-repo-what-to-check-2026 ↩ ↩2