Proof Abundance

Sources#

• No More Heroes - Terence Tao Talks to Numberphile About AI
• The Buckmaster Interview (Navier-Stokes) - Numberphile

Summary#

At the start of October 2026 OpenAI released "hundreds of proofs" at once, about 300 papers indexed at github.com/openai/math (the CONTENTS.md linked from the video description). The corpus holds the release only through Terence Tao's account of it in a ~20-minute Numberphile interview with Brady Haran, uploaded 2026-10-15 (No More Heroes - Terence Tao Talks to Numberphile About AI, practitioner-opinion, YouTube auto-captions). His name for the shift is the one he says he has been using "this whole year": "we are transitioning from an era of proof scarcity to proof abundance." For thousands of years a proof of a hard theorem required understanding, expertise, literature and conversation, so finding one also grew the field. Now "we found the shortcut… we're drowning in proofs, but we're not getting all the understanding … that we used to get." His analogy is food: scarcity meant every piece was nutritious, and abundance is "synthetic food" that is "technically … edible" and not as satisfying.

The page holds the argument's four consequences, the release as Tao describes it, and where both bear on the rest of the formal-math domain. Everything about the release is Tao's description of artifacts this corpus does not hold: none of the ~300 papers and not the repository were fetched.

The release, as one reader describes it#

• Scale against reading capacity. "I've only looked at about 10 of the 300 odd papers… There's just so many to look at." The most-consulted independent mathematician on AI results (Terence Tao) had read about 3% of the release by the time of an interview uploaded about two weeks after it. That is Verification as the New Bottleneck at the scale of a discipline: generation was a single drop, and assessment is volunteer labour with no deadline.
• The mix, by his partial reading. Some are "extensions of what was already being done", which the people already working on them would have reached "if [they] had a year or two more time". "One or two" have "some really clever new ideas", at least as far as anyone can tell. "There's a whole bunch which are still incomprehensible", and "one or two even have some errors and retracted". His summary: "there will be some good stuff in there and some rubbish but we don't know what the percentages are yet."
• No consultation, and no support afterwards. "There was no consultation with people who are already working on these problems." The courtesy he describes, discreetly checking with a colleague you know is on the same problem, did not happen, so "a lot of junior research projects … just got wiped out overnight", and "at the very least it was extremely rude and … discourteous". After release, "no one from OpenAI is helping us… as far as they're concerned they're done."
• What the source does not say. Whether the ~300 are Lean-formalized, prose, or a mix; which model or harness produced them; how many were retracted and on whose initiative; and whether any outside party checked them before release. On every axis this domain tracks — kernel check (see kernel-level proof auditing), statement fidelity (see statement drift), intelligibility (Logical vs Intelligible Proof) — the release is unassessed here.

The headline result: a zero-free region at Re = 7/8#

Tao calls one result "the biggest advance on the subject in 50 years": a "massive new zero free region" for the Riemann zeta function, which "we now call the quasi-Riemann hypothesis". The Riemann hypothesis puts every nontrivial zero on Re(s) = 1/2. "Previously we only knew that they lied a little bit to the left of one" (the classical zero-free regions shrink to Re(s) = 1 as the imaginary part grows), and now no zero lies to the right of Re(s) = 7/8: "we're one quarter of the way to the Riemann hypothesis." (The "7/8" is an ingest-time correction of the caption's "78", fixed by Tao's own "one quarter of the way" from 1 to 1/2.) His bound on the method: it "cannot get beyond three quarters without another extremely clever idea." Whether that means the line Re(s) = 3/4 or three quarters of the way from 1 to 1/2 (Re(s) = 5/8), the transcript does not disambiguate. His verdict: the Riemann hypothesis "hasn't fallen but … it's been dented."

A standard consequence the source does not state: a zero-free half-plane Re(s) > θ gives a prime-number-theorem error term of O(x^(θ+ε)), so a uniform θ = 7/8 would be a large improvement on the classical error term. That is wiki inference from textbook analytic number theory, not Tao's claim, and the result itself has no preprint, referee or second reader in this corpus. Tier: practitioner-opinion about a vendor artifact. Tao read it; nobody in this corpus has checked it.

His account of how it was made is the part most relevant to Transformative Creativity: "a combination of … three or four existing ideas arranged in a very unusual way", "not something any … single mathematician alive I think would have come up with", although "if you got the right people to work together [and] lock them in a room they might have come up with this after lots of trial and error." He adds a speculation, flagged as such: he is "sure" the labs "secretly devoted much more time to the Riemann hypothesis than what they've disclosed", "couldn't touch it", and "something beyond current AI is needed".

Four consequences Tao draws#

1. Understanding and technology are two axes, and the field took the dysfunctional path#

Around 2023 (ChatGPT, then Lean) Tao saw two paths: adopt the tools while holding on to understanding, or sacrifice understanding for output. His analogy is the automobile: you can drive and stop exercising, or find a balance. He spent the period encouraging the healthy path and "we got overtaken by events": enough people and companies were "interested in just … optimizing the speed and output and performance of these tools above everything else", willing "to toss understanding to the side so that they could prove more … theorems, get more press releases". He calls this "not sustainable". Inside mathematics an unexplained result is tolerable ("the rest of the world can cope"); used to model weather or the energy grid, recommendations nobody understands "at some point … [are] going to cause a real problem". This is Outsource Your Thinking, Not Your Understanding stated at the scale of a discipline, by the field's most prominent practitioner, and it is the practitioner's version of the gap Logical vs Intelligible Proof argues philosophically.

2. The heroic age ends; an industrial one begins#

"We are seeing the end of the romantic era of mathematics, the age of heroes", something other disciplines went through "as early as the industrial revolution" while mathematics kept its quirks (blackboards, one-on-one advising, the Andrew Wiles attic story). He expects "an industrial era where … math will become much more large scale", with big teams, "more like the other sciences". His one prescription that preserves something: teach traditional methods first, "just like … we teach … kids how to do arithmetic by hand before giving them calculators." He frames it as loss with gain, and says hero worship "also has a downside": it "has winners and losers".

3. Strip-mining the training problems#

The most specific mechanism in the interview, and the one with no prior statement in the corpus. Mathematics trains its next generation on an unwritten reserve of medium-difficulty problems. A senior person could solve them, but there is an informal "agreement that … these are good problems for students", and solving one is the standard route to "career recognition". Scarcity kept the reserve full: "there were so many problems and … so few students … and problems were hard to solve." His analogy: under "sustainable fishing" nobody needs allocated grounds; now "we have this sort of strip mining of mathematics where … huge machines … like dynamite fish … hundreds of problems at a time", and "suddenly we're running out of good problems to assign to people."

Two second-order effects follow. Problem-sharing, the Erdős tradition of publishing lists to inspire students (the lists FrontierMath Erdős Benchmark is built from), now runs backwards: "people are reluctant to share their favorite problems because someone with an AI prompt will just sort of snag it before" the student or postdoc it was meant for. And the new questions that used to follow a solved problem are not arriving: "these AIs are depleting our pool of questions and not … replenishing them with … anything new." When a person solved a hard problem, "naturally there were two new questions that … showed up as follow-ups."

The same chilling effect, reported two weeks earlier from the other end of the September dispute. Tristan Buckmaster, the rival claimant on The Navier–Stokes AI Claim, told Numberphile on 2026-09-29 (The Buckmaster Interview (Navier-Stokes) - Numberphile, practitioner-opinion) that "everyone … doesn't want to talk about their open problems", because people fear the AI companies "will suddenly come in and … treat them as … benchmark problems". He says this has "slowed us down", for a result whose timing (September 8 against September 20) "would have zero impact on the world". He also reports acting on it: he and Alpöge hold "a whole bunch of results" on other problems, and he is not releasing them yet, choosing instead to spend the coming months explaining their methods. His norm is the one Tao calls etiquette: when someone else is close, you "join forces rather than accelerate", and front-running with resources the others lack is "inappropriate". Two independent practitioners, one inside the AI-for-math work and one outside it, report the same withdrawal from problem-sharing. Neither is a measurement.

This is The Tragedy of the Cognitive Commons in a domain that page does not cover, with a different depleted resource. Lovett's commons regenerates through entry-level work. In mathematics it regenerates through entry-level problems, and the machine removes the training ground by solving it rather than by replacing the trainee. The commons framing is the wiki's mapping; Tao's own words are fishing and marathons.

His remedy runs in two directions, because "at least one of us has to change." Internally, mathematicians should move "to more like collaborative projects … where instead of each person working on their one problem we have teams of people working on hundreds of problems", so that whether any one problem is scooped matters less. Externally, "the broader community has to sort of respect some of our space". It is not forbidden for the public to drive along a marathon route and beat the runners, but "it's just understood that this is … socially not a good etiquette." He does not say "leave the lighthouse problems alone"; Haran raises that position and Tao answers with the training-pipeline argument instead.

4. Incentive alignment held while the goals were out of reach#

Tao separates individuals from institutions. Many people at the labs "came from mathematics", he counts many as friends, and many hold "good balanced views", but "the culture and the incentives of these companies, especially the biggest ones … are not the same as the ones for … academia." Different goals were harmless "as long as … all the goals are really far out of reach", because everyone moved the same way, and "until about 2025, this was the case": weak tools meant the labs had to work with mathematicians, which was "win-win". Strong tools that optimize well turned that around, so "if you keep optimizing the goals that the tech companies want … they're now moving away from the goals that … the community wants": understanding, teaching the next generation, and sustaining people "who can actually … pose new questions". The labs are "racing to get their … most powerful model up and running before their rival does", and the race to AGI "is becoming a bit toxic and a bit scary", sacrificing safety as well as understanding.

The argument has a structural form worth keeping beyond mathematics: two actors' objectives can be compatible only because neither can be reached; capability removes the slack, not the goodwill. His word for it, "misalignment", matches the mathandai.org declaration's "severe misalignment", initially signed by 25 Fields Medallists, which Logical vs Intelligible Proof records from a different source. Tao mentions a group chat "with like 20 Fields medalists" but does not mention the declaration, and the corpus does not record whether he signed it.

Process is what stays human#

Asked what will change, Tao names one thing the machines lack: "these AIs … can generate the final product, you know, these nice papers, but … they're very bad at showing the process of how you get to that." Humans have been bad at it too, since the field publishes polished papers and rarely "the making of", with the Polymath projects as the exception. He predicts more open process "because that's … something that … is more uniquely human". It is a third answer to which artifact carries understanding, beside the prose and the formal proof (Which Artifact Carries the Understanding? Prose and Formal Proof as Two Objects). The derived page finds reported understanding coming from a human re-derivation; Tao locates it in the visible process, which neither AI artifact supplies.

Where this sits against the corpus#

• It is the first primary Tao document. Every earlier appearance of Terence Tao was someone else citing him or writing on his blog. The "solving problems without new insights" objection that Noam Brown attributes to him is here first-hand and softened in one respect: Tao now credits "one or two" papers with "really clever new ideas" and calls the zeta result a "very clever new technique". The objection survives in its problem-posing form: the AIs are "depleting our pool of questions and not … replenishing them". The combinational reading of the zeta result is on Transformative Creativity.
• It qualifies the corpus's only statement that no domain shows commons-scale depletion. The Data Wall and the Validation Commons Are One Supply Constraint concludes that formal mathematics shows the sub-task version of the risk (the residual human job is checking statements). Tao reports a different and larger mechanism in the same domain: depletion of the training-problem supply and of follow-up questions, with junior projects "wiped out overnight". It is a practitioner's report, not a measurement, so it qualifies the derived page's conclusion without overturning it.
• It continues the consultation pattern of The Navier–Stokes AI Claim. In September OpenAI contacted the rival lab after finishing, over credit, and Quanta records the rival team releasing early after word of their progress reached OpenAI. In October, on Tao's account, there was no contact with the academics working on the affected problems at all, before or after. Two OpenAI releases, both with contact that was absent or after the fact, and in both cases no outside party controlled the timing.
• It is a third data point on the internal-model gap. FrontierMath Erdős Benchmark measures a pre-release GPT-6 Astra at 2/68 on curated Erdős problems under a $300 protocol. Tao describes ~300 papers and a 50-year advance on the zeta function from one lab's release a month later. The model, budget and selection are all undisclosed, so the two are not comparable, and the contrast is what Latent Capability Overhang and Evaluation Horizon Versus Release Cadence predict: a lab's undisclosed process produces the headline, and the disciplined protocol produces the small number.

Evidence note#

practitioner-opinion, kept. No measurement. Tao is the best-placed independent reader the corpus has, and he says plainly how little he has read (~10 of ~300). Every figure about the release is his. Interests: he has collaborated with the labs and says so; as a senior mathematician he has a stake in the community norms he defends, and the interview catches him at the centre of the response ("every hour of my day is scheduled", articles, talks, group chats). The interviewer's questions are leading in places ("Are AI companies the enemy?"), and Tao's answers are more measured than the questions.

Transcript caveat. The source is auto-captions with incomplete speaker markers, ending mid-exchange at ~20 min. The raw's provenance note lists six caption phrases left uncorrected because the likely correction changes the meaning. None is quoted on this page; where the passage matters (the Navier–Stokes aside, the safety point), it is paraphrased around the doubtful word.

Connections#

• Logical vs Intelligible Proof — the philosophical statement of what proof abundance loses: certified answers without intelligible solutions. Tao supplies the practitioner's version and names the community goals (training, new questions) his own goals list contains
• The Navier–Stokes AI Claim — the September release this follows: same lab, same absence of prior contact with the people working on the problem, and the result Tao cites as one "no one really understands"
• The Tragedy of the Cognitive Commons — the commons argument in mathematics, with the training-problem reserve as the depleted resource instead of entry-level work
• Transformative Creativity — the zeta result as Tao describes it, "three or four existing ideas arranged in a very unusual way", is a combinational result beyond any single mathematician, and depleting questions without replenishing them is the problem-posing boundary
• FrontierMath Erdős Benchmark — the Erdős problem-list tradition that problem-sharing came from, which Tao says now runs backwards, and the disciplined counterpart to an undisclosed bulk release
• Automated Conjecturing — the generate side this release lacks: Tao's complaint is that question supply is being consumed and not regenerated, which is what conjecturing systems claim to automate
• AI-Driven Formal Proof Search — the paradigm's "verification as a filter for human review" framing, against a release where the human review is one mathematician reading 10 of 300
• Outsource Your Thinking, Not Your Understanding — Tao's automobile analogy is this prescription at the scale of a discipline
• Verification as the New Bottleneck (hub) — assessment capacity, not generation, is what binds after a 300-paper drop
• Latent Capability Overhang — the undisclosed-budget, undisclosed-model side of the capability gap
• Evaluation Horizon Versus Release Cadence — a lab releasing internal-model output faster than anyone outside can evaluate it
• The Data Wall and the Validation Commons Are One Supply Constraint — the corpus's "no domain shows commons-scale depletion" conclusion, which Tao's account qualifies for mathematics
• Which Artifact Carries the Understanding? Prose and Formal Proof as Two Objects — Tao's third candidate carrier of understanding, the visible process
• Logical Certificates Carry Neither Interest Nor Understanding: Unimplied Conjectures and Formal Proofs — places Tao's "process" and Buckmaster's promise to explain his methods in the pattern where understanding reaches the field through a person, and reads his question-depletion complaint against conjecturing queues that are certified only negatively
• Terence Tao — the source, speaking first-hand for the first time in the corpus
• OpenAI — the releasing lab
• Noam Brown — whose second-hand paraphrase of Tao's objection this source revises
• Curriculum Exhaustion and the Entry Rung: The One-Supply Argument Revisited — strip-mining read as one stock consumed twice: the medium problems are both the model's curriculum and the student's, which is where the 2026-08-17 'arrival and need are correlated' correction fails
• Post-Scarcity Macroeconomics — Tao's strip-mined medium-problem reserve is cited there as curriculum depletion before the threshold arrives

Open Questions#

• What does the release actually contain? The falsifiable form: of the ~300 entries in github.com/openai/math, how many come with a Lean formalization, how many have been retracted or corrected, and how many do outside mathematicians classify as new against extensions of existing work? Tao's "we don't know what the percentages are yet" is the baseline. Settled by fetching CONTENTS.md or by a survey from a mathematician who reads the whole set.
• Does the zeta zero-free region at Re(s) = 7/8 survive independent checking, and is the "cannot get beyond three quarters" barrier ever crossed? (Trigger: an arXiv preprint and a refereed or expert exposition of the result, a Lean formalization checked by an outside party, or a published gap.)
• Does the training-problem depletion show up in anything measurable: fewer published problem lists, fewer first papers by students on problem classes the release covered, or the team-based "hundreds of problems" model Tao prescribes appearing as a funding or publication format? (Trigger: a bibliometric study of student first papers in affected areas after October 2026, or a community statement on problem reservation.)

Sources#

• The Buckmaster Interview (Navier-Stokes) - Numberphile — Tristan Buckmaster interviewed by Brady Haran, Numberphile, uploaded 2026-09-29, 46 min, practitioner-opinion, auto-captions. Cited only for the problem-hiding report, his withheld results and his front-running norm. Full treatment on The Navier–Stokes AI Claim.
• No More Heroes - Terence Tao Talks to Numberphile About AI — "No More Heroes - Terence Tao Talks to Numberphile About AI", Terence Tao interviewed by Brady Haran, Numberphile (YouTube), uploaded 2026-10-15, ~20 min (captions end at 19:45, mid-exchange), practitioner-opinion. YouTube auto-captions deduplicated with clean-vtt.py, with 24 targeted ASR fixes at ingest (including "78" → "7/8" for the zero-free line, fixed by Tao's "one quarter of the way"). Six meaning-changing caption candidates were left unfixed, and none is quoted here. The proof release (github.com/openai/math/blob/main/CONTENTS.md) is linked from the video description and was not fetched, so every statement about its contents is Tao's.