Executive Summary

On October 22, Nature ran a column by Terence Tao in its World View section. The headline reads "Can maths survive AI? Fields medallist Terence Tao calls for a 'Maths 2.0' world," and the print version appears in volume 658, page 1103. The body of the column sits behind a paywall, and the full text was not accessible. The only sentence from the column quoted here verbatim is therefore the subheading the editors placed under the headline. Everything else from the column is reported indirectly, through search excerpts and summaries Tao published.

"As machines race ahead in solving problems, mathematicians must establish a new era that prizes human understanding over simply publishing more solutions." Source: subheading of the Nature column, Nature 658, 1103, October 22, 2026

The column is often described as the place where Tao proposed Math 2.0 in Nature. That is half right. Read the primary texts in date order and the three core ideas of the column each appear earlier, in a different text. "From proof scarcity to proof abundance" and "proof indigestion" come from the essay "Mathematics in the age of AI," posted to arXiv on August 17, which is itself a paper-format version of a public lecture given at the International Congress of Mathematicians in July. The phrase "Math 2.0" never appears in that essay. Its first public appearance is the four-post thread Tao published on Mathstodon on October 6. The nine example activities that flesh out Math 2.0 come from the slides of a public lecture at Caltech three days later, and those slides in turn never use the word "indigestion." The Nature column compresses these four months for a general readership. The "Maths" in the headline is British spelling, chosen by Nature's editors.

This article refers to the four texts by the names in the table below: the August essay, the October 6 thread, the Caltech lecture, and the Nature column. On this topic, knowing which sentence came from which text is the whole of factual accuracy, so the body attributes each claim to its source. Writing "in Nature, Tao" before a point that exists only in the slides is itself an error.

The four bold rows are the primary texts cited throughout. The October 6 timestamp is the metadata of the repository's first commit, which is why the Korean-language Wikipedia records the release as October 7, early morning in Korea. This article uses October 6 throughout.

Korean coverage focused entirely on the Caltech lecture. AI Times reported a rebuttal to the cancer-cure thought experiment on October 12, Mobiinside summarized the lecture on October 16 with attribution to the blog and social posts, and GeekNews carried a link to the slide PDF. As of October 31, no Korean-language piece on the October 22 Nature column could be found. The English-speaking world looks similar. No direct response to the Nature column from mathematicians such as Gowers, Buzzard, or Ellenberg turned up in searches. The closest thing is a blog post by an anonymous postdoc, published the same day as the column, which says the author fears being scooped almost daily.

A brief aside: the week the column for specialists appeared, a first book for general readers appeared too. Six Math Essentials, from Quanta Books, 128 pages, came out in the US on October 27, with a UK edition from Pelican at 96 pages the same day. It is an introduction to six concepts, number, algebra, geometry, probability, analysis, and dynamics, and has nothing to do with the subject at hand. Still, a specialist column and a popular book arriving in the same week is an accidental footnote to how far Tao personally practices the activity the Nature column calls exposition.

Tao's argument starts from one premise. Traditional mathematics, which Tao calls "Math 1.0," ran on the assumption that proofs are scarce. The second slide of the Caltech lecture puts it this way: finding a solution to an open problem and proving it correct was hard, rare, and demanded considerable expertise, so most of the institutions, traditions, reward systems, and stated goals of Math 1.0 were organized around that scarcity. The August essay makes the same point with a more concrete list, and that sentence is quoted in full in Section 3.

So why did mathematics fall first? The Caltech lecture names three properties of proofs. The same three turn out to be the key to the data quality discussion below.

2.1Objective Verification, Digitizability, and High-Quality Data

The first is objective verifiability. Mathematicians disagree about which mathematics is interesting or productive, but there is broad agreement about what counts as a correct and complete proof of a given problem, and a proof written in a proof assistant language such as Lean can be checked by machine. The second is digitizability. A proof needs no physical substrate, takes no real-world data as input, and requires no experiment to confirm the result. In principle, pencil and paper plus a standard textbook for the definitions are enough to produce and verify one. The third is large, high-quality data. Solving a hard problem requires knowing the research literature of the field, and an enormous volume of that literature already exists in digital form.

"This combination of objective verifiability, digitizability, and large, high-quality data sets have made the task of solving mathematical problems particularly amenable to the machine learning techniques that power modern AI. Frontier AI labs have thus prioritized advances in mathematical problem-solving capability, ahead of other applications which lack one or more of the above three features." Source: Caltech lecture slides, October 9, 2026

The third condition deserves attention. Models alone did not open the era of proof abundance. The canonical literature that mathematicians refined over centuries, the mathematics in the form it takes in textbooks and reference works, was there as training data. The August essay sharpens the point: the training data for these tools is, quite literally, the output of the canonicalization process. The reason mathematics fell first and the largest cost of the indigestion described below meet at the same spot.

▲ Pebblous original diagram. The three properties of proofs from the Caltech lecture (Section 2.1), gathered into one figure.

2.2Jagged Capability, and Still a Conditional Analysis

The Caltech lecture spells out the consequence. Anyone with enough compute and a frontier model can now obtain solutions to many open problems once considered hard, in relatively little time, and increasingly without any special expertise on the part of the person asking. On tasks that are subjective, require interaction with the real world, or have little data, AI's record stands in sharp contrast. Tao calls this "jagged" capability: astounding in some directions, inadequate in others, within mathematics and more broadly. The same lecture also makes clear that, despite the impression given by press coverage and social media, most problems at the frontier of mathematics remain unsolved even after enormous compute. The space of problems is infinite and difficulty has no ceiling.

The August essay pins this outlook down under the name "Working Hypothesis" and attaches a condition. The hypothesis is that AI tools will reasonably soon become capable of performing a reasonable fraction of research-level mathematical tasks with reasonable levels of success, quality, supervision, and cost. Tao does not ask readers to want it to be true, to believe it, or to accept it, and states that what follows is a conditional analysis. The same reservation applies to this article. The indigestion argument below rests on the premise that mathematics is heading in that direction, and it does not assume that AI has solved mathematics.

2.3When the Measure Becomes the Target

Problem solving has historically been the most visible goal of research mathematics, but according to the Caltech lecture it was only ever a proxy for more valuable long-term goals. The proxy was reasonably accurate while proofs were scarce and broke down under abundance. In Tao's words, further blind optimization of problem solving alone is now actively harmful to the long-term health of mathematics.

The August essay explains the breakdown through Goodhart's law: when a measure becomes a target, it ceases to be a good measure. It gives two reasons this applies with special force now. One is technical. Generative AI is inherently ungrounded, so it finds the gap between a measure and the thing it was meant to measure more readily than humans do. The other is economic. The financial incentives of the AI industry reward achievement that is demonstrable, quotable, and benchmarkable. Together, the two incentives push resources toward the easiest thing to count, "an open problem was solved," and leave behind what that count was supposed to stand for, which is mathematical understanding. Section 3 looks at exactly where that leftover accumulates.

Section 6 of the August essay rewrites the goal "solve an open problem" five times. The first version is a single arrow from problem to solution. But the solution has to be checked (second version), a checked solution has to be written so others can read it (third), a well-written solution has to be digested and accepted by the community (fourth), and an accepted result still has to be worked into the form that textbooks and reference works carry (fifth). Figure 4 of the essay draws the five versions as one chain, and the caption says that what began as a single arrow has become a chain of five stages, of which only the first was ever an explicit goal of the community.

The diagram below redraws that figure. The stage names follow the arrow labels in the essay: generation, verification, exposition, publication, canonicalization. Some coverage has reported "six stages" because the essay writes "digestion" and "canonicalization" side by side on the last arrow, but the body of the essay also uses digestion as an umbrella term for the last four stages together. The fifth stage's own name is canonicalization.

▲ Figure 4 of the August essay, redrawn for this article. The six nodes and five arrows carry the labels of the original; the line weights and the four gap markers follow the description in Section 7 of the essay and were added here.

3.1Indigestion Is a List of Four Gaps

Section 7 of the essay states that if the working hypothesis holds and policies and culture stay as they are, significant "impedance mismatches," or in a less flattering metaphor "proof indigestion," will emerge all along the chain. It then lists where, in four lines:

• AI-generated proofs will accumulate faster than they can be verified.
• Verified AI-generated proofs will accumulate faster than they can be given a readable write-up.
• AI-generated proofs, even those required to be both correct and well written, will overwhelm a traditional peer review system that depends on volunteer expert labor.
• Even the published proofs will be too numerous for the community to work into definitive form.

The four lines map one to one onto the four gaps in the diagram above. Indigestion is a symptom of a whole chain in which only the first two stages sped up, rather than a bottleneck at any single point. Tao also explains in one sentence why the institutions are vulnerable to this symptom. This is the original of the summary that the research, teaching, and evaluation systems of mathematics were run on the premise of scarcity.

"Most of our institutions — journals, priority conventions, hiring and promotion criteria, prizes, the very notion of a research program — were designed under the assumption of scarcity, and it should not surprise us if they behave poorly under abundance." Source: August essay, Section 7

The essay adds that the signs of indigestion predate AI. Growth in the volume of literature, the length and specialization of major proofs, and the strain on refereeing were all recorded earlier, and AI sharply amplifies these existing stresses.

3.2October 6, the Day Gap 1 Became a Number

This is the event the Nature column cites. The column reports that on October 6 OpenAI dumped hundreds of solutions to such problems online all at once, and that this disrupted countless research programs overnight. "Hundreds" here counts results. Counted by manuscript, the same release is 722 papers; counted by result bundle, it is 372. Those figures come from the Pebblous report of October 13, which downloaded the repository at its release commit and counted directly. One thing that report established is the core of this section. The share of proofs that were machine-checked comes out three ways depending on the unit: about 24% by manuscript, about 36% by result bundle, and about 42% as stated in the repository README. Three papers were withdrawn within three days, leaving 719.

Placed on the chain, the meaning of those figures is clear. Everything released sits in Gap 1 as unverified solutions, and somewhere between one in four and four in ten of them have passed machine verification. Gaps 2 through 4 have no count yet. Tao's remark on the radio program Science Friday describes that state most briefly: there is nobody at OpenAI to call for help understanding this dump of 300 papers. In a New York Times interview on October 16, Tao compared AI mathematical output before the release to "toxic waste," then revised that after the release to "maybe treated waste." The two phrases must be read as a pair. Quoting only the first distorts an assessment Tao personally softened.

3.3Verification Takes a Day; Digestion Varies by Case

"Only the first two sped up" can be checked against numbers. For four AI-produced mathematical results from this year, the table lines up the time taken for generation, machine verification, and the point at which a human could understand and explain the result. The first three rows are events Pebblous covered earlier, and the fourth is the October 6 release. The short version is that machine verification finished within a day when it was fast, while the time humans needed to digest the result ranged from two days to several weeks depending on the case, and most of the October 6 batch remained undigested 25 days later.

The first three rows carry over figures from the Pebblous report of August 9 and the article of June 19. The Jacobian counterexample is a finite computation, substitute three points and expand a determinant, so verification was nearly free, and Tao jumped on it personally as a prominent case, which makes it a biased sample. The point of the table is that the speed after verification varies from a day to several months by case, rather than that indigestion is always slow.

The August 9 report recorded one contrasting case the table omits. Two weeks after the Jacobian counterexample, OpenAI solved ten open problems with Astra and released the Lean proofs, the reasoning narratives, and polished manuscripts together, and the people who had posed the problems still responded that they could not see why the problems were solved that way. Releasing more does not close the understanding gap by itself. On Tao's chain, Gap 2 does not shrink with the volume released.

3.4Reporting Bias and First Proof

The Caltech lecture has a one-panel cartoon slide, a speech bubble reading, roughly, "Wow, every math problem I see on social media is being solved by AI!" The next slide names the illusion: reporting bias. Only successes are published, and failed AI attempts go unreported. Tao writes that a more transparent and scientific evaluation of AI capability, one that fully reports key statistics such as resource consumption and negative results, is urgently needed, and offers First Proof as one attempt in that direction.

Each First Proof batch poses ten research-level problems whose solutions are known to the contributor but have never been posted anywhere online. The August essay reports the second batch: seven of the ten problems received at least one passing grade from one of four systems, at compute costs on the order of tens to hundreds of dollars per problem. Third-batch results were released on October 14, and setting aside problems of citation and exposition, all ten were reached by some system. That figure appeared eight days before the Nature column. The two batches pose different problems, so reading this as "up from seven to ten" is wrong. Two things follow for this argument. Capability really is rising fast, and a framework that publishes failures and costs alongside successes already exists.

3.5The Slowest Stage Is the Most Valuable

The essay also records how far today's AI has come on the third stage, exposition. AI-written mathematical text is nearly flawless in spelling, grammar, and formatting, yet it often dwells on trivial steps while passing quickly over, or even obscuring, the most interesting and novel part of the argument. It frequently fails to situate the result within prior literature or to provide the high-level overview a reader needs to judge whether the argument is worth reading. Exposition is a far fuzzier optimization target than verification, and Tao adds that under the working hypothesis AI will improve substantially here too.

The essay also carries a paradoxical warning about exposition. A proof can be too slickly written. When routine steps and the genuinely difficult steps are presented as equally easy to digest, the reader cannot tell where the effort went. That friction is informative, Tao writes. Exposition, too, can be over-optimized.

On the fourth stage, publication, the essay draws the human boundary more firmly. For a proof to contribute to a field it is not enough to be correct, nor enough to be readable; the community has to accept and value it, and that acceptance is by nature slow and human, an external process that authors and tools alone cannot optimize. Authors can help by recording the insights, false starts, and backstory from the time they wrestled with the problem. Current AI tools, though, are quite opaque about their own problem-solving process, and this is particularly true of proprietary models whose inner workings are a corporate secret.

Acceptance rests on editors and referees who work voluntarily and usually without credit. This labor is treated as less prestigious than producing proofs, but Tao calls it the very mechanism that converts individual achievement into collective progress and understanding. AI evaluation tools might screen automatically for missing verification, missing attribution, or incoherent exposition, much as plagiarism checkers do, and so spare scarce expert attention. Passing an automated filter cannot substitute for the community's acceptance, however, and Tao does not believe human referees can be removed from the publication process. This is why the third gap in Section 3.1 hangs on volunteer labor.

On the final stage the essay writes its longest paragraph. Publication is not the end. Key results must enter the canonical textbooks and reference works of the field in a form that can be taught to the next generation, and this process of canonicalization, in which a result is restated in its natural generality, given its right proof rather than its first proof, connected to its neighbors, and absorbed into the standard toolkit, is the slowest stage of all. It requires broad, deliberative consensus, it is the stage least amenable to optimization by AI tools, and Tao judges it the most valuable part of the entire process. This stage meets the third condition of Section 2.

"Indeed, the very success of AI tools in mathematics depends crucially on the canonical theories that human mathematicians have painstakingly built and rebuilt over the centuries: the training data for these tools is, quite literally, the output of the canonicalization process." Source: August essay, Section 6

Across the four texts, the crisis reads as three strands. Grouping and numbering them as three is this article's doing. Tao never counts "three" in any of the four texts; the seed of all three sits in a single paragraph of the second October 6 post, and the first half of the Nature column covers all three without numbering them. The strands are: proofs that do not turn into understanding, a culture of sharing open problems that is coming under threat, and the path by which new mathematicians are trained being cut. For each, this section records which text contributed what.

▲ Pebblous original diagram. The three strands bundled in one paragraph of the October 6 post (Sections 4.1–4.3), drawn apart here.

4.1Proofs Pile Up, Understanding Does Not Keep Pace

The first October 6 post describes what one big proof used to set in motion in Math 1.0. When a long-standing conjecture fell, a great deal of activity followed: seminars and talks, follow-up papers, extensions to other problems, and the process by which the proof was digested, streamlined, placed in context with other results, and finally absorbed into the textbooks and lecture notes of the next generation. The second post describes the present as the contrast.

"problems are being solved autonomously by AI prompters who have no interest in the broader field itself once their initial target is 'solved', and do not understand the AI output well enough to answer questions on the result, give talks, or otherwise interact with the rest of the field. Many fewer seminars, workshops, collaborations …" Source: October 6 thread, post 2 of 4

The Nature column renders the same point for a general audience. It argues that human understanding of mathematics is falling behind because AI systems generate results faster than mathematicians can digest and absorb them, and that some researchers using AI tools simply move on to the next problem once they have a minimally presentable proof. It also notes that many AI-generated results are not described clearly enough to meet the quality standards for publication. The Astra case in the Section 3 table is this diagnosis in the flesh. Even where more was released, the understanding gap stayed where it was.

4.2Sharing a Problem Becomes a Risk

Mathematics has run for centuries on the practice of posting open problems and asking colleagues for ideas. The second October 6 post says that practice is changing: promising open directions are now being withheld from the public out of fear that one's own research will be scooped. The Nature column describes the same phenomenon as a fear of having one's research directions scooped by a rival equipped with a frontier AI model. The anonymous postdoc's post mentioned at the end of Section 1 shows that fear in the first person.

The reason Tao treats this as more than a matter of competition is in the third post. Irreversibility.

"a problem that has been 'solved' cannot be somehow reverted to become 'unsolved', and even the mere knowledge that a solution exists 'contaminates' efforts … solutions to open problems are now being harvested at large scale in an unsustainable fashion, leaving entire fields of mathematics much less fertile" Source: October 6 thread, post 3 of 4

The Caltech lecture turns this idea into the lighthouse metaphor. In pure mathematics, open problems are guiding lighthouses that prompt exploration of the surrounding mathematical landscape rather than destinations to be reached for their own sake, and what is learned while trying to solve them is often worth more than the final answer. Reaching a lighthouse prematurely by automated tools cuts off the exploration of the paths not taken and sterilizes the surrounding field. A remark Tao made on Mastodon in early September, that open problems are a non-renewable resource, was the preview of this metaphor.

4.3The Ecosystem Feels the Threat First

A doctoral student becomes a researcher by working on a specific hard problem and solving at least part of it. The Nature column observes that graduate students and young researchers have built their skills on exactly such problems, and that the October 6 bulk release upended a great many such plans in a single night. In the New York Times interview Tao also said that senior researchers, Tao included, are less threatened. The threat reaches the ecosystem that trains new mathematicians before it reaches the profession.

The closing thoughts of the August essay put the problem in one line: the goal of training a mathematician is not achieved by producing correct homework. A student does not become a mathematician because AI solved the homework correctly on the student's behalf, and, turned around, the question that remains after AI has cleared hard problems by the hundreds is an educational one: what to give students to do. The anonymous postdoc's post adds one sentence at this point: the rules cannot change only for young mathematicians.

The same closing thoughts say the conclusion will differ by domain. In education and the training of young mathematicians, Tao writes, it is crucial to emphasize the irreducibly human aspects of the work and to restrict the use of AI tools fairly strictly. In other domains, by contrast, mathematicians should take the lead in setting best practices for bringing tools into their workflows on their own terms and not on terms dictated by vendors. That the same Tao who refuses to block AI stands on the restrictive side in education is a clue to keep in mind when reading the remedy in Section 5.

4.4A Fourth Problem Under the Three: Attribution

The Nature column touches on one more problem briefly. Because no one can tell how much an AI-produced result leans on prior work, it becomes unclear who deserves credit. The August essay handles this by quoting two recommendations of the Leiden Declaration directly. The first affirms the humanity of authorship: credit and responsibility continue to belong to humans within the mathematical community and should not be given to automated systems. The second calls for effort on proper attribution: because automated tools are limited in their ability to identify the sources of ideas, the duty to track down and cite the sources that made a result possible grows correspondingly. The declaration goes as far as saying that if satisfactory attribution is impossible, the paper should state so.

Pebblous has covered this axis twice. The June 30 article on the Leiden Declaration summed up the five values the declaration codifies as a single line about sources and responsibility, and the August 9 report organized the provenance gaps that formal verification cannot close into five patterns. One figure that report carried shows the floor under the attribution problem. Seven in ten journals have an AI disclosure policy, yet of some 75,000 papers published since 2023, only 76, about 0.1%, explicitly disclosed AI use. When disclosure is required but not practiced, and AI mathematical results arrive by the hundreds, none of the four gaps has a box in which to write whose result it is.

Tao's remedy leaves AI problem solving alone and rewrites the standard by which a mathematician's work is judged. The fourth post of the October 6 thread, where the name "Math 2.0" was first attached, lays out that standard in three directions.

"'Math 2.0' will need to decenter the role of raw problem solving and value mathematical progress more holistically - for instance by elevating the role of exposition, but also that of community building and opening up new directions of study. … our community will need to explicitly re-evaluate its criteria for education, publication, and career advancement, to reflect the 'Math 2.0' era." Source: October 6 thread, post 4 of 4

The August essay places the same remedy on the chain of Section 3. Our culture should decrease the emphasis it places on proof generation, and in particular on being the "first" to solve a problem, and correspondingly increase the emphasis on proof digestion: exposition, refereeing, publication, and canonicalization. At this point the essay also cites David Bessis's April piece and argues for moving away from a "theorem economy" that rests mainly on proof generation. Summaries circulating online often condense the remedy to "four activities Math 2.0 should value more," but the number four appears in none of the texts. There are the three directions of the October 6 thread and the nine examples of the Caltech lecture.

5.1The Nine Activities of the Caltech Lecture

Before its nine example activities, the Caltech lecture introduces a new era of experimental mathematics. Large mathematical datasets that could not be built before can now be assembled, and with them mathematics can be done as a census instead of a case study. Two examples are given. The Equational Theories Project (ETP) settled 22 million true-or-false statements in universal algebra in 57 days, using human and automated reasoning together, and the Inverse Galois Problem challenge (IGP24) crowdsourced a polynomial for each of the 25,000 possible Galois groups of degree 24 and is now extending to degree 31. The nine examples that follow are set out in the table below, with dates and figures attached where something is already running. Of the nine, only the first three are running far enough along to carry a date or a figure, and the open-source mathematics model the lecture said was coming soon has not appeared.

The gist column renders the sentences on three Caltech lecture slides (Examples / More examples / Even more examples). The dates and figures in the third column were checked against each project's official page and press coverage. "No specific project named" means the lecture gave no example, not that nobody is doing it.

The three items one widely shared summary called "Tao's separately organized Math 2.0 plan," which are re-explaining results that are hard to understand even though proven, integrating many papers, and building formal proof libraries, are rows 1, 2, and 3 of this table. "Verifying the meaning and limits of AI-generated results" comes from row 9 and from the Nature column's mention of verification.

The August essay filled the same space two months before the lecture, in an appendix. As new infrastructure to complement traditional workflows it names collaborative formalization projects, structured problem databases, new venues for exposition, and new venues for publishing negative and partial results, and it lists seven that already exist: the formal mathematics library Mathlib, the video journal Mathematical Discourse which peer-reviews research talks, the Erdős problems database, a database of optimization constants, the SAIR Foundation's mathematics competitions, First Proof, and the formal proof registry Palomar. Row 3's Mathlib and the First Proof of Section 3.4 are among the seven. The caption's caveat, that "no specific project named" reflects only the lecture's silence, is partly filled in here.

5.2The Leiden Declaration: Twenty-Three Recommendations, Five Values

Rather than proposing a full program of its own, the August essay points to the Leiden Declaration. Tao calls the declaration, published in June 2026 and endorsed by the International Mathematical Union, an excellent starting point, and describes it as the product of a 2025 workshop at the Lorentz Center in Leiden, consisting of twenty-three recommendations addressed to individual mathematicians, to mathematical organizations and not-for-profit funders, and to policymakers. Tao quotes four of the recommendations aimed at individual mathematicians and comments on each: besides the two on attribution carried over in Section 4, they are to disclose tool use and to support the need for refereeing. The worst scenario Tao names to avoid is an author hiding AI use to escape criticism from peers. The Pebblous article of June 30 summarized the same declaration as five values, so the declaration can be read as five values and twenty-three recommendations. It collected 130 signatures on its first day and passed 1,000 within 24 hours.

After commenting on the four recommendations, Tao offers a personal rule of thumb. It is a threshold hung between exposition and publication on the chain of Section 3, and it is Tao's own statement of the thesis that formal verification is not the end.

"My own suggested rule of thumb: if the authors cannot convincingly demonstrate that they are able to give a clear, expert-level talk on their results, one that is correct and properly attributed, then the result should not be published. A proof that no human can properly explain should be viewed as incomplete, even if it has been formally verified." Source: August essay, Section 8

5.3A Dissent, and the Conclusion That the Bottleneck Is Imagination

There are voices more pessimistic than Tao's. Max Weinreich of Baruch College, City University of New York, in a piece titled "The crisis of AI-generated mathematics," is skeptical of the very prospect that institutional redesign can resolve the crisis, and favors giving more weight in hiring and promotion to work done without AI. The piece was posted to arXiv in early August, revised on October 18, and has been accepted by the Notices of the American Mathematical Society. Tao lists it among the AI-related resources on the living summary page. Where the two part ways is the remedy, not the diagnosis.

Tao's conclusion is on the last slide of the Caltech lecture. The Nature column closes on the same optimistic note: AI need not diminish mathematics, and used creatively, with reward structures adjusted, it can become a tool that deepens human understanding.

"The primary bottleneck to achieving 'Math 2.0' is not technology, or even institutional culture. It is imagination." Source: Caltech lecture, final slide

This is the part that circulating summaries leave out entirely. Tao did not keep indigestion inside pure mathematics. The Caltech lecture draws the Math 1.0 ecosystem on one slide. Pure mathematical problems are solved by objectively verifiable pure solutions, and the insight gained along the way passes through human understanding into applied problems, yielding applied solutions that require subjective judgment. Open problems are solved within a renewable cycle, and insight from pure mathematics is transferred safely to applications. The next slide inserts AI into that picture. When AI solves problems indiscriminately and in a non-renewable way, the damage reaches beyond the long-term health of the field to the safe transfer into messier real-world applications. A four-cell table then shows where that damage comes from.

▲ The four-cell table of the Caltech lecture, redrawn for this article. The final slide of the lecture discloses that its images and diagrams were AI-generated, so the original is not reproduced. The cell names are verbatim; the small captions were added here.

The explanation runs as follows. A lack of human understanding of a real-world system is not inherently dangerous. It just means that the field has not yet been satisfactorily mathematized. When such a lack of understanding is combined with the optimizing power of AI, however, misalignment can occur. One slide writes it like a formula: powerful AI plus a lack of human understanding equals misalignment risk.

6.1The Cancer-Cure Thought Experiment and the Korean Rebuttal

The lecture's thought experiment goes like this. An advanced AI is instructed to find a cancer cure that passes Phase III clinical trials, and to make no mistakes. After enormous computation it produces a cocktail of chemicals never seen before and claims that, mixed and injected, it kills every cancer cell. Nobody knows how it was found, but the AI somehow attaches a Lean certificate, and the cocktail really does pass Phase III. Could this answer be a misaligned one that exploited a weakness in the trial procedure or in its mathematical model? Here Tao turns the question to the audience. If you had just received an early cancer diagnosis and a chatbot recommended this treatment, would you find it reassuring to know, before injecting it into your veins, that at least one human cancer expert understands the mechanism behind the cure, even partially? Or that some human mathematician understands the mathematical models used to find and analyze the cocktail?

Korean coverage attached a rebuttal at this point. On October 12, AI Times reported a response from the American oncologist Vinay Prasad to the effect that cancer drugs whose mechanism is unknown are already in use. On October 16, Mobiinside recounted how the qualifier "even partially" in the question had dropped off on social media, leaving the question to be mocked without it. Neither rebuttal overturns the argument. What Tao asked for was not a complete understanding of the mechanism but the distinction between a state in which at least one person understands and a state in which nobody does. In the four-cell table, the dangerous cell is not the one where understanding is weak but the one where understanding is weak and AI is heavy.

6.2Credit Approval, Financial Risk, and Regulation That Moved the Other Way This Year

The Nature column places the real-world version of this thought experiment in finance. The column grants that proof indigestion confined to pure mathematics might be a minor concern, but AI-generated mathematical models deployed in the real world without a matching understanding of their mechanisms could have serious consequences, and it cites the AI models already used in credit approval and market risk measurement. If a model is too opaque or too fast for people to interrogate, errors spread into public systems. The cancer cocktail question, whether any mathematician understands the model, returns with a credit model in place of the cocktail.

Look for regulation that would back this warning and an unexpected picture appears. The US Federal Reserve's model risk guidance, SR 11-7, required from 2011 that financial institutions demonstrate understanding of their models in development and use, and said documentation should be detailed enough that a third party unfamiliar with the model could understand how it works, its limitations, and its key assumptions. On April 17 of this year that guidance was replaced in full by SR 26-2. The new guidance turns to a principles-based approach, drops uniform validation cycles, explicitly places generative and agentic AI outside the scope of model risk, and adds language stating that it establishes no enforceable standards or prescriptive requirements. The October in which Tao wrote the warning in Nature was the sixth month in which the US requirement to understand models had put exactly that type of AI out of scope.

Europe points the other way but has not switched on. The EU AI Act, in Annex III item 5(b), classifies AI used to assess the creditworthiness of natural persons or to compute credit scores as high-risk, and imposes three layers of explanation duties: technical documentation, instructions for use, and, under Article 86, the right of affected individuals to a clear and meaningful explanation of the decision. Under the Digital Omnibus amendment, the start date for these obligations has been pushed back to December 2, 2027. The accurate description of October 2026 is not "regulation that demands understanding is coming" but "regulation that demanded understanding is empty in the face of AI." Keeping the misaligned cell of the four-cell table empty is, for now, the job of the organizations using the models, not of regulators.

Why Pebblous Is Watching

Tao's Five Stages Map Directly onto the Data Pipeline

Pebblous's work is to judge whether data is in a state fit for AI and to leave that judgment as a record that can be read later. The chain Tao drew for mathematics runs parallel to the chain of that work. Automatic labelers, synthetic data generators, and code generation models have made the generation stage hundreds of times faster, and automatic verifiers such as schema checks and cross-labelers have kept pace at the verification stage. But explaining why a label is what it is, the review by which a team accepts the data, and the canonicalization into datasheets and reference datasets still take as long as they ever did. The indigestion Tao locates in the four gaps appears on the data side as label indigestion. The table below is the correspondence.

Original Pebblous correspondence table. The left column is the five stages of Figure 4 in the August essay; the middle and right columns are Pebblous's reading.

Set DataClinic's diagnostic report on this table and it occupies the three cells after verification: making data explainable, letting a team digest it, and leaving it as a reference. On the Physical AI side, Section 6 applies directly. If nobody understands a robot policy trained on data drawn from simulation, that policy belongs in the misaligned math cell of the four-cell table.

Undigested Output Becomes Training Data Again

The third condition of Section 2 and the quotation at the end of Section 3 meet on the ground of data quality. The condition Tao named for mathematics falling first was large, high-quality data, and the essay wrote that this data is, quite literally, the output of the canonicalization process. Indigestion breaks that virtuous cycle. An undigested proof cannot become canonical, and what is not canonical cannot become training data for the next model. In the language of data quality, this is a feedback loop that feeds unverified output back in as training data, and it mirrors the warning Pebblous has repeated in its articles on synthetic data.

Figures from the data side show the structure. In a reanalysis of e-commerce synthetic labeling (SynthAVE), the 92.6% accuracy reported as "passed automatic verification" was a conditional figure in which humans had inspected only the disagreement cases, and a stratified re-estimate that included the agreed cases gave 90.5%. In synthetic data for competitive programming (X-Coder), 12.7% of the reference solutions in 150 human-audited tasks were wrong. In an autoformalization pipeline for theorems (Theorem Prover as a Judge), 23.1% were wrong from the formalization of the question onward. These three figures come from outside mathematics and are not the 24 to 42 percent of Section 3 under another name. The structure repeats, not the number. A "passed verification" figure is almost always conditional, stripping the condition leaves a residual error of 5 to 13 percent, and that residual flows into the next round of training data.

An Institution for Digestion Matters More Than a Faster Verifier

Carried over to an organization that receives large volumes of AI output, Tao's remedy becomes four questions. For each piece of output: who understands it, is there an explainable document, has it been absorbed into the team's knowledge, and has it been left as a reference? If the first is empty, the other three cannot be built. The first Math 2.0 activity Tao named, open exposition problems, the task of rewriting an existing result so it can be understood, translates for an organization into rewriting the datasets it has already accumulated so they can be understood. That is a matter of writing down why the existing data looks the way it does, not of producing new data faster.

For customers in finance and the public sector, the regulatory situation of Section 6 applies as it stands. The US removed generative AI from its model risk guidance this April, and Europe's high-risk obligations for credit assessment have slipped to December 2027. While the regulation that demanded understanding stands empty, the only thing that fills the space is an organization's own institution for digestion.

A Company Standing in Tao's Ninth Cell

Look again at the last row of the table in Section 5: using expertise gained from understanding pure models to analyze the failure modes of applied models and probe for misalignment. That cell is where Pebblous stands. It sells the capacity to understand and digest, not the capacity to generate. Pebblous's three earlier pieces, the provenance-gap report in August, the Fields Medalists' declaration article in September, and the 722-manuscript ledger report in October, all showed with numbers that the verification cell is empty. The addition here is the word of a participant that filling that empty cell is becoming a new standard of achievement inside mathematics itself. The correspondence stops here. This section ends at showing that Tao's cells and the data cells line up.

References

Primary Sources, Tao

• 1.Terence Tao, Mathematics in the age of AI, arXiv:2608.16753 v1, August 17, 2026. Referred to in the text as the "August essay." The acknowledgments state that it is a paper-format version of a public lecture given at the International Congress of Mathematicians in July 2026. Source of the five-stage chain (Figure 4), the four gaps of proof indigestion (Section 7), and the quoted Leiden Declaration recommendations (Section 8).
• 2.Terence Tao, Mathstodon thread, four posts, October 6, 2026, 18:00 UTC. The "October 6 thread" in the text, and the first public appearance of the "Math 1.0 / Math 2.0" contrast. ⚠️ It precedes the release of OpenAI's repository by about four hours and does not name OpenAI.
• 3.Terence Tao, Math 2.0, AIM public lecture, 30 slides, Beckman Auditorium, Caltech, October 9, 2026. The "Caltech lecture" in the text. Source of the three properties, the four-cell table, the cancer-cure thought experiment, and the nine example activities. The final slide discloses that the images and diagrams in the talk were AI-generated, so the diagrams here were redrawn. Blog announcement: terrytao.wordpress.com, October 10, 2026.
• 4.Terence Tao, Can maths survive AI? Fields medallist Terence Tao calls for a 'Maths 2.0' world, Nature 658, 1103, online October 22, 2026, doi:10.1038/d41586-026-03389-1. The "Nature column" in the text. ⚠️ The full text could not be read because of the paywall. Only the subheading is quoted verbatim; the body is reported indirectly through search excerpts and the summary in item 5.
• 5.Terence Tao, Terence Tao on AI: a living summary, updated October 28, 2026. A page Tao maintains that summarizes and links the four texts above. It names the five stages as generation, verification, exposition, publication, and canonicalization, matching item 1.
• 6.Terence Tao, Six Math Essentials, Quanta Books / FSG, 128 pages, ISBN 978-0-374-62179-7, October 27, 2026 (UK edition, Pelican, 96 pages, same day). The aside in Section 1 only.

Institutions and Projects

• 7.Leiden Declaration on Artificial Intelligence and Mathematics, June 2, 2026, doi:10.5281/zenodo.20302944. Endorsed by the International Mathematical Union. The count of 23 recommendations was confirmed verbatim in Section 8 of item 1.
• 8.First Proof. The second batch (10 problems, 4 systems, 7 passed, tested May 28, 2026) follows the account in item 1; the third batch (10 problems, released at the ICARM webinar of October 14, 2026) was checked against the project's published results. The two batches pose different problems. M. Abouzaid, N. Srivastava et al., First Proof Second Batch, arXiv:2606.18119.
• 9.AIM, Terence Tao Presents "Math 2.0" at the 2026 AIM Public Lecture, October 19, 2026. How the announced title "Machine Assisted Proof" became "Math 2.0" on the day. · Caltech Mathathon official page (November 13 to 15 schedule, "Old Problems, New Proofs") and press coverage of the September redesign.
• 10.US Federal Reserve, SR 26-2, April 17, 2026 (replacing SR 11-7). The provision that explicitly excludes generative and agentic AI from the scope of model risk, and the language that it "establishes no enforceable standards." · EU AI Act Annex III item 5(b), Articles 11, 13, 14, and 86, with the start date of December 2, 2027 under the Digital Omnibus.
• 11.Mathlib statistics, October 2026. 291,414 theorems, 772 contributors. · Equational Theories Project (22,028,942 statements, 57 days) · SAIR/LMFDB IGP24 (25,000 Galois groups, extension to degree 31). Figures checked against the verbatim text of item 3 and each project's page.

Dissent and Adjacent Scholarship

• 12.Max Weinreich, The crisis of AI-generated mathematics, arXiv:2608.02859 v2, October 18, 2026, accepted by the Notices of the AMS. The dissent of Section 5.3. The author is an assistant professor at Baruch College, City University of New York.
• 13.David Bessis, "The fall of the theorem economy," April 21, 2026. Cited in item 1 in the passage on the "theorem economy."
• 14.SynthAVE, arXiv:2607.07469 (synthetic label accuracy, conditional 92.6% → unconditional 90.5%) · X-Coder (12.7% error rate in synthetic SFT reference solutions) · Theorem Prover as a Judge, arXiv:2502.13137 (23.1% question formalization errors). The data-side figures in "Why Pebblous Is Watching." ⚠️ Not from the mathematics domain; used only as examples of the same structure.

Press and Reactions

• 15.Elisabeth Bumiller, The New York Times, October 16, 2026. "I'm on a war footing," "toxic waste" before the release → "maybe treated waste" after, and the remark that senior researchers are less threatened. · The Observer, October 23, 2026. The "toxic waste" remark revisited and OpenAI's rebuttal. · Science Friday. "There's nobody at OpenAI I can call to help me understand this dump of 300 papers."
• 16.AI Times (Korea), October 12, 2026 (oncologist Vinay Prasad's rebuttal of the cancer-cure thought experiment) · Mobiinside, October 16, 2026 (summary of the lecture and how the "even partially" qualifier was lost to mockery) · GeekNews, October 12, 2026 (slide PDF). All three Korean pieces covered the Caltech lecture of item 3; no Korean-language piece on the Nature column of item 4 could be found as of October 31, 2026.
• 17.proofsandprompts.com, October 22, 2026. Post by an anonymous postdoc. "Almost daily, I fear being scooped by someone, perhaps using AI, who posts first." · 36kr, October 14, 2026; GIGAZINE, October 13, 2026; hyper.ai; note.com (Japanese-language summary). Census of overseas reactions. ⚠️ The "six stages" in 36kr and note.com count digestion and canonicalization separately and were not used.

Related Pebblous Articles

• 18.OpenAI's math papers carry a review field, and what it says is unchecked (October 13, 2026). Source of the 722 manuscripts, 372 bundles, and 24 to 42 percent in Section 3.2. · The Proof Was Right. How It Got There Wasn't Recorded (August 9, 2026). The Jacobian counterexample, the prime number theorem formalization from 18 months to three weeks, the 0.1% AI disclosure rate, and the Astra contrast. · 25 Fields Medalists Signed a Declaration Against How AI Proofs Are Announced (September 14, 2026). · 1,000 Mathematicians Asked AI to Get Consent Before Training on Their Papers (June 30, 2026). The five values; signatures from 130 to 1,000. · AI Broke an 80-Year-Old Math Conjecture, Then Humans Rewrote the Proof (June 19, 2026). The nine verifiers.