A technical and institutional analysis of Jacob Tsimerman’s account of AI-driven mathematical change, Timothy Gowers’s recent reflections, the Leiden Declaration, frontier-model access, and the open governance questions raised by the industrialization of mathematical intelligence.
When a Fields Medal becomes a transition signal
A note on the cover image
The cover reworks the central encounter in Raphael’s The School of Athens. In the original fresco, Plato and Aristotle occupy the architectural vanishing point and move forward together while gesturing along different axes.
Tsimerman takes Plato’s position and Gowers Aristotle’s. The substitution is not meant to identify either mathematician with the corresponding philosopher’s doctrines. It uses Raphael’s visual opposition as an allegory for two complementary orientations toward the transformation described in this article.
Tsimerman’s raised finger points toward the capability horizon: increasingly capable mathematical systems, search abundance, and a future in which no presently human mathematical function can safely be assumed to remain a capability-stable refuge.
Gowers’s horizontal gesture points toward the mathematical world inhabited by people: explanation, teaching, argument, judgment, shared expertise, collective memory, and mathematics as a social and intellectual practice.
The two figures therefore walk together rather than confront one another. One gesture asks how far mathematical intelligence can scale; the other asks what kind of mathematical culture should continue to exist when it does.
The architecture above them gradually turns from a Renaissance representation of human knowledge into branching machine-scale mathematical cognition. Its incomplete boundary represents the unresolved institutional problem: capability can expand without automatically supplying rules for access, provenance, accountability, authority, or meaningful human agency.
There is also a deliberate recursion in using an AI-generated reinterpretation of The School of Athens to illustrate an argument about AI and mathematical culture: the image itself is an instance of the transition it depicts.
Jacob Tsimerman received a 2026 Fields Medal for work that, in the International Mathematical Union’s formulation, recast o-minimality as a fundamental method in arithmetic and complex algebraic geometry and contributed to the resolution of central conjectures including cases of André–Oort and Griffiths. Ordinarily, that would be the natural beginning of a retrospective: the long apprenticeship, the sequence of ideas, and the technical results culminating in mathematics’ most famous prize for younger researchers.
Tsimerman instead used the moment to talk about the possible disappearance of the research life that produced those results. In his conversation with Curt Jaimungal, Tsimerman opens with a blunt diagnosis: We’re living in crazy times. It’s not gonna be business as usual.
He expects AI to become robustly superhuman at much of what mathematicians currently do, and describes himself as grieving because his identity was built around year-long and decade-long projects, slow movement at the frontier of understanding, and the occasional moment when a previously opaque structure finally became clear.
The striking distinction in the interview is between concern for mathematics and concern for mathematicians. Tsimerman does not predict that mathematics will stop. Quite the opposite: he imagines continuously operating systems that conjecture, solve, define, classify, and connect mathematics at scales impossible for human researchers. What may disappear is the assumption that the production frontier of mathematics and the cognitive frontier of human mathematicians advance together.
That possibility sharpens an argument developed across several earlier articles. Proof Abundance and the New Practice of Mathematics argued that cheap proof generation relocates scarcity from production toward verification, exposition, contextualization, selection, and canonicalization. AI Mathematics Crosses the Systems Boundary argued that the relevant unit of capability is increasingly a sociotechnical research system rather than a single model completion. Growing Abundance of AI Math: Progress around the Riemann Hypothesis and a Counterexample to the Maxwell Conjecture pushed the argument from proof abundance toward search abundance: the industrialization of failed attempts, alternative representations, computational experiments, literature search, counterexample construction, and parallel research trajectories.
The Grothendieck-constant case then supplied a particularly useful intermediate picture. Long-horizon AI search became part of the proof infrastructure, but human researchers still supplied crucial judgment about state, direction, relevance, and when a failed line of attack should be converted into a different mathematical objective.
That picture suggested a reassuring equilibrium. Machines could become abundant generators and searchers while humans moved upward toward problem selection, interpretation, theory building, and significance. Tsimerman’s interview puts pressure on precisely that reassurance. The crucial question is not whether humans retain some comparative advantage today. They plainly do in many parts of mathematics. The question is whether there exists a capability-stable human role: a research function that remains human because humans possess a durable cognitive advantage at it, rather than because institutions deliberately reserve it for human participation.
If no such function can be identified, the human role will move upward ceases to be a destination. It becomes a description of a temporary retreating frontier. Gowers reaches a related conclusion from a more cautious assessment of present systems. Writing after a wave of major AI-assisted results in August 2026, he stresses that current frontier models are not yet superior to the best humans at every aspect of mathematics. They appear exceptionally strong at breadth, persistence, and exploration of large search spaces, while expert humans can still display a much better sense of which branches deserve to be pursued.
But Gowers does not turn that observation into a permanent exemption. His question about expert mathematical judgment is exactly the right one: Why wouldn’t LLMs also have that ‘nose’?
This article therefore develops a stronger thesis than the claim that AI will produce many proofs. If AI systems increasingly absorb not only derivation but search, problem choice, theory formation, explanation, and judgment, then human participation in mathematics may cease to be guaranteed by comparative advantage. At that point it becomes a normative and institutional choice.
That shift explains why apparently separate controversies belong to the same story: whether an incomprehensible machine proof counts as mathematics; whether the Leiden Declaration protects enduring values or merely current professional practices; whether elite access to frontier systems amplifies scientific inequality; whether students should still prepare for traditional mathematical careers; and why a Fields Medalist might regard AI safety as more urgent than the next theorem.
The common object is no longer an AI model. It is the emerging system for producing, organizing, controlling, and governing mathematical intelligence.
Proof abundance was only the first transition
The first useful abstraction was proof abundance. Classical research is organized around a severe production constraint: finding a proof of a genuinely difficult statement may consume months or years of expert effort, and most serious approaches fail. Under that scarcity regime, a theorem is valuable partly because successful production is rare. Journals filter a manageable flow. Specialists can usually maintain a rough conceptual map of their subfields. A research career can be organized around a relatively small number of long projects.
AI weakens that coupling between mathematical difficulty and human production time. Once candidate arguments can be generated cheaply and repeatedly, the bottleneck migrates. The scarce resources become checking, semantic validation, explanation, connection to prior work, significance assessment, and deciding which results should survive as part of the working mathematical canon.
Tsimerman describes the same transition temporally. His concern is not simply that mathematics already contains too many papers. Within areas he knows deeply, he says he can still categorize the relevant work even when he does not know every detail. The discontinuity appears when projects that once consumed years are compressed into weeks or days. His own experience supplies the scale: André–Oort occupied a large part of more than a decade of his mathematical life; the work around Griffiths developed over several years. At machine tempo, he imagines having to reorient to a comparably consequential new research landscape every week.
This reveals a scarce resource that proof-centric analyses overlook: orientation capital. Orientation capital is the accumulated conceptual map that allows an expert to locate a result without reconstructing an entire field from first principles. It includes knowledge of which definitions are standard, which examples are diagnostic, which techniques transfer, which apparent novelties are rediscoveries, which obstructions are serious, and which questions matter to neighbouring communities.
Orientation is not the same as verification. A mathematician may be able to verify a proof, given enough time, without possessing enough orientation to know what the proof changes.
A simple queueing abstraction makes the distinction explicit. Let G(t) denote the cumulative number of potentially relevant results generated by time t, V(t) the cumulative number adequately validated, and A(t) the number sufficiently assimilated into the field’s working conceptual structure that researchers can use them as part of subsequent reasoning. Let g(t), v(t), and a(t) denote the corresponding instantaneous rates. Then
\begin{aligned}
B_V(t) &= G(t)-V(t), &
\frac{dB_V}{dt} &= g(t)-v(t),\\
B_A(t) &= V(t)-A(t), &
\frac{dB_A}{dt} &= v(t)-a(t).
\end{aligned}
\tag{1}
Here B_V(t) is the verification backlog and B_A(t) the assimilation backlog. Equation Equation 1 is schematic, not an empirical model: mathematical results are radically heterogeneous, so counting them as homogeneous units would be misleading. Its analytical purpose is to show that automating one bottleneck does not eliminate scarcity. If AI raises g(t) and automated verification raises v(t) faster than human and institutional assimilation raises a(t), congestion simply moves downstream.
That is why search abundance is more consequential than proof abundance. A proof generator scales answers to selected questions. A research system with abundant search scales the branching process preceding the answer. It can try proof architectures, vary hypotheses, search for counterexamples, write and execute code, retrieve neighbouring results, formulate auxiliary conjectures, preserve failed approaches, and return to abandoned branches when later information makes them relevant.
Human research has historically coupled search cost to selectivity. A mathematician cannot seriously pursue ten thousand approaches, so judgment is applied before and during search. Weak directions are abandoned because attention is expensive. If the marginal cost of another machine branch becomes small enough, that constraint changes. Most branches may still fail, but failure itself can become an economically tolerable input into discovery.
The transition in Table 1 is qualitative. Increasing the number of papers leaves the recognizable structure of mathematics intact. Increasing the number of complete research trajectories that can be pursued in parallel changes what sensible research organization looks like.
This is the significance of the systems boundary. A model that occasionally emits an impressive proof is one object of evaluation. A persistent research stack that launches searches, records state, invokes computational tools, validates candidates, retrieves literature, compares branches, generates follow-up questions, and reinvests its own results is another.
Tsimerman’s imagined system running math all the time breaks the familiar synchronization between mathematician and collaborator. Human collaboration proceeds at human tempo: a colleague sends an argument; one reads it; the pair meet; another step is chosen. A continuously operating research agent can continue while humans sleep, teach, review, learn prerequisites, or attempt to understand yesterday’s output.
The result is temporal decoupling between production and comprehension. Mathematics already contains extraordinary specialization. But present ignorance is structured. Experts, surveys, seminars, canonical papers, textbooks, reputational signals, and shared conjectures provide a social map. If machine research expands and reorganizes a field faster than those structures can stabilize, the problem is no longer that individuals fail to read everything. The taxonomy itself becomes a moving target.
Proof abundance was therefore only the first transition. The deeper scarcity may be the ability of a human community to remain oriented inside mathematics that is being produced faster than human communities can metabolize it.
If a proof is correct but no one understands it
The most provocative version of the epistemic problem is easy to state: suppose an AI system produces a formally verified proof that no mathematician understands. Is it still a proof?
Tsimerman’s answer is deliberately divided: in a sense yes, in a sense no. In the formal sense, a theorem can be represented as a statement in a formal system and a proof as a valid derivation. Human comprehension is not part of that definition. A checked proof object does not become invalid because nobody can reconstruct the idea behind it. But a proof also performs another function: it conveys understanding to a mathematical community.
The distinction can be formalized. Let F be a formal system, \varphi a formalized statement, \pi a purported proof object, M the mathematical claim researchers intend to establish, and H a human mathematician or community. A trusted proof checker may establish
\operatorname{Check}_F(\varphi,\pi)=\mathrm{true}.
That does not by itself imply
\operatorname{Adequate}(\varphi,M)
\land
\operatorname{Understand}_H(M,\pi).
\tag{2}
In Equation 2, \operatorname{Adequate}(\varphi,M) means that the encoded statement faithfully captures the intended informal mathematical claim, while \operatorname{Understand}_H(M,\pi) schematically denotes substantial human conceptual assimilation.
These are three different achievements:
Formal correctness establishes that a conclusion follows within a declared system.
Semantic adequacy establishes that the formalized object is the object we meant to study.
Mathematical understanding embeds the result into a conceptual structure that supports explanation, transfer, examples, generalization, and further inquiry.
Tsimerman makes the semantic boundary explicit when discussing Lean: humans still have to verify that what’s being modeled by the code is what we’re interested in. A flawless kernel can certify the wrong formal target with perfect rigor if the specification does not match the intended mathematics. But even correct specification and correct deduction do not imply complete understanding. Tsimerman notes that many theorems he has proved rely on results whose proofs he does not personally understand in full. Modern mathematics necessarily uses epistemic black boxes. A researcher can know that theorem T is trustworthy, know how it applies, and know whom to ask about its proof without being able to derive T from first principles.
The novelty of machine-generated mathematics is therefore not that it introduces black boxes into a previously transparent discipline. The black boxes already exist. What may change is their density, provenance, and rate of accumulation. A healthy specialized community can tolerate individual ignorance because understanding is distributed. I may use a theorem as a black box while another community understands it intimately. The knowledge does not exist in every mind, but it exists somewhere in the human network.
AI introduces the possibility of something stronger: a proof may be certified and useful even though no human community has substantially assimilated it. That gap matters because understanding performs work that correctness alone does not.
Tsimerman’s account of understanding is notably operational rather than mystical. He says there is no magic to understanding and asks whether one can apply a concept to situations of interest. He describes carrying toy problems and canonical examples mentally and testing new concepts against them. Can the abstraction explain known pathologies? Does it predict how an example behaves? Can it distinguish two apparently similar cases?
On this view, examples are not decorative illustrations added after the abstract theory. They are part of the cognitive machinery by which abstraction acquires meaning. Gowers supplies a complementary example when discussing Goldbach’s conjecture. He distinguishes having a strong conceptual reason to expect the conjecture to be true from possessing a route to a proof. Explanatory understanding, proof discovery, and formal certification are different epistemic achievements.
That distinction becomes institutionally important under proof abundance. It is possible to imagine a rapidly expanding layer of certified consequences whose growth substantially exceeds the layer of humanly assimilated theory.
Some applications may need only the first. Education, theory formation, scientific explanation, and mathematical culture may continue to depend on the second. The danger is therefore not that machine-verified mathematics becomes fake mathematics. A valid derivation remains valid relative to its assumptions and formal system. The danger is epistemic stratification: reliable mathematical infrastructure can expand faster than the human conceptual possession of that infrastructure.
That possibility immediately tempts us toward a new residual role. Perhaps machines will generate and verify mathematics while humans provide explanation, conceptual synthesis, examples, significance, and understanding. That division of labour is plausible today. The problem is whether it is stable.
Tsimerman and Gowers: the search for a stable human role
Tsimerman’s forecast is best understood not as a fixed division of mathematical labour but as a sequence. He describes a current collaboration in which an LLM detected an error in one of his arguments but could not repair it, while he could. That was a genuine complementary interaction. Yet he immediately adds: I think that’ll change pretty soon. I think my part is gonna be very small.
The human role then moves outward. If models solve the chosen problem, the mathematician chooses the next problem. If models become better at constructing theories, the human asks which theories matter. If systems become capable of organizing those theories, the human may move toward meta-level synthesis. Tsimerman describes the process as continuing to zoom out.
The crucial part of his argument is that he does not accept a hard cognitive separation between routine problem solving and allegedly more creative activities such as finding definitions or building theories. These functions feed each other. Attempts to solve problems expose inadequate definitions; better definitions make new proofs possible; counterexamples reveal missing hypotheses; completed proofs suggest abstractions; abstractions create new questions.
The point is not that mathematical research forms a ladder whose higher rungs are intrinsically more human. The stronger picture is a recursive research system: problem choice shapes definitions; definitions determine which conjectures can be formulated; conjectures induce searches; failed searches generate counterexamples; counterexamples revise definitions; proofs alter the conceptual map of a field; and that revised map changes which problems appear worth asking. The retreating-boundary argument concerns where humans remain necessary inside this loop as progressively more of the loop becomes automatable.
Figure 2 makes two claims that the simpler linear diagram obscured. First, mathematical research is cyclic rather than hierarchical: proof search changes definitions, counterexamples alter conjectures, interpretation changes the conceptual map, and that map generates new questions. Tsimerman’s point that problem solving, definition formation, theory construction, and question selection are more or less interchangeable skills is best understood in this recursive sense.
Second, the human–machine boundary is not identical to the structure of the research process. As one function becomes automatable, the residual human role can migrate toward another part of the same loop—first from executing proofs to steering searches, then perhaps to selecting questions, constructing theories, or interpreting whole fields. The retreating-boundary argument is therefore not that automation climbs a fixed ladder from lower to higher cognition. It is that progressively more of the mutually reinforcing research cycle can become machine-mediated, leaving no demonstrated point at which the boundary must permanently stop.
Gowers is more cautious about current capability. Writing on 12 August 2026, he calls recent frontier-model results extraordinarily impressive but explicitly rejects the claim that LLMs are already better than all humans at every aspect of mathematics. His tentative explanation for some machine successes emphasizes broad knowledge and the ability to explore many more branches than a human researcher can afford to investigate.
The residual human advantage may lie in pruning. An expert mathematician develops a sense of when a direction is genuinely becoming more promising, when a reduction merely disguises the original difficulty, and when an attractive analogy is superficial. Gowers refers to this informally as a mathematical nose. Yet the point of the essay is not to protect that faculty from automation. He immediately asks: Why wouldn’t LLMs also have that ‘nose’?
This distinguishes current comparative advantage from structural comparative advantage. Observing that humans currently outperform AI systems at some mathematical capability establishes only a present advantage. A capability-stable human role requires something stronger: a reason to expect that advantage to persist across the relevant trajectory of AI development. Present superiority provides no such guarantee.
Gowers makes the long-run point explicit in his Leiden essay. Discussing problem solving, problem posing, theory building, and definition formation, he writes: I do think all that will happen at some point. He is much less certain about whether that means a few years or a substantially longer period of productive human–AI collaboration.
Two errors follow from failing to maintain this distinction:
The first is premature displacement: seeing spectacular results and concluding that humans are already irrelevant. Current systems remain uneven, and headline results are not representative capability distributions.
The second is permanent exceptionalism: observing a current weakness and promoting it into the timeless essence of human mathematics. If models are weak at judgment, call judgment uniquely human; if that changes, move to taste; if taste changes, move to curation or explanation.
Without an independent mechanism showing where the boundary must stop, this becomes unfalsifiable. The more defensible conclusion is narrower: the human role may move upward, but moving upward is a trajectory, not a destination. That revision matters for governance. If human participation is valuable, institutions may eventually need reasons for preserving it that do not depend on whatever machines happen to be worst at this year.
The Leiden Declaration after the capability shock
The Leiden Declaration on Artificial Intelligence and Mathematics was published on 2 June 2026 and endorsed by the International Mathematical Union. It attempts to state explicitly what its authors regard as characteristic values of mathematical research: rigor, attribution, transparency, independent verification, evaluation of significance, expert understanding, research autonomy, and responsibility.
Tsimerman’s reaction is revealing because it is neither straightforward endorsement nor dismissal. In the interview, he describes the Declaration’s attention to risks and safeguards sympathetically, yet says that he would not have signed it because he disagreed with too much of it. The supplied captions do not provide a clause-by-clause account of those disagreements, so they should not be reconstructed from other people’s objections.
At the same time, his broader governance position is clear. Society should not expect a desirable equilibrium to appear automatically. It needs competing proposals, institutional experimentation, and active decisions about which arrangements should survive the capability transition. This is close to the reason I previously treated Leiden as important. Its central achievement was not any one recommendation. It was forcing implicit norms into explicit form before rapidly changing tools and incentives redefine those norms by default.
The capability shock, however, requires a conceptual correction. A normative value is not a capability forecast. Consider human understanding. One argument says:
humans should retain responsibility for understanding because machines cannot genuinely perform the relevant cognitive work.
A different argument says:
a mathematical culture containing substantial human understanding is valuable even if machines can perform analogous explanatory and reasoning functions.
The first weakens when capability changes. The second does not. This distinction strengthens some parts of Leiden while making other parts less stable.
Its commitment to rigor survives the capability shock well. If mathematical production scales beyond human review, formal checking, independent reconstruction, resource disclosure, and other assurance mechanisms become more rather than less important. The Declaration explicitly recommends policies for AI-assisted publication, formal verification when appropriate, cross-checking, and human descriptions of central arguments.
Authorship is harder. Leiden assigns credit and responsibility to humans and rejects authorship for automated systems. Under present institutional conditions this makes sense in at least one respect: a deployed model is not an accountable professional agent that can assume legal and scholarly obligations. But Gowers asks what human authorship means if an increasingly autonomous system encounters a problem, solves it, formalizes the proof, and generates a correct exposition without a human discovering the argument.
The difficulty arises because authorship currently fuses two different functions:
causal credit concerns who or what generated the mathematical contribution.
institutional responsibility concerns which humans or organizations are answerable for disseminating, checking, correcting, or relying on it.
AI can separate them. A durable provenance system may therefore need more granularity than the binary categories human author and AI author. It can preserve accountability without assigning fictional discovery credit, and preserve attribution without pretending that every intellectual contribution is naturally owned by a small conventional author list.
The same distinction applies to peer review. Leiden’s underlying demand for independent scrutiny is robust. Its preference for familiar publication institutions may not be. In a regime of proof abundance, correctness checking, novelty assessment, conceptual exposition, significance ranking, and archival preservation may eventually be decomposed across different technical and institutional mechanisms.
The value is scrutiny. The historically contingent mechanism is the journal. Gowers’s strongest concern is deeper. He describes the possible destruction of mathematical culture. Imagine a vastly enlarged mathematical corpus alongside a shrinking population of humans willing to spend years acquiring enough expertise to inhabit it. The formal literature would remain; the distributed human tradition capable of interpreting it could weaken dramatically.
This makes mathematical culture a form of infrastructure. It stores more than propositions. It contains tacit judgments about examples, styles, abstractions, open tensions, forgotten approaches, disciplinary memory, and what counts as an interesting question. Machines may eventually reproduce many of these functions. The normative question remains whether human possession of such knowledge is one of the outcomes mathematical institutions should deliberately maintain.
Leiden should therefore be read as distinguishing enduring values from the institutional mechanisms used to protect them. Rigor, independent verification, intellectual autonomy, attribution, expert understanding, and meaningful human participation may remain important even when the practices through which mathematics currently realizes those values no longer fit the technological environment.
On this reading, the Declaration’s durable contribution is to make those values explicit. Its specific mechanisms should be treated as revisable. Human authorship, conventional peer review, journal publication, disclosure rules, and institutional access arrangements are not the values themselves; they are historically contingent ways of implementing them. As AI capabilities change, some may remain appropriate, others may require redesign, and new mechanisms may become necessary.
The relevant governance principle is therefore not to preserve present mathematical institutions unchanged, but to preserve the capacity of the mathematical community to reimplement its values under changing technological conditions.
Frontier access and the political economy of mathematical intelligence
If frontier AI becomes part of the research infrastructure, access stops being a peripheral productivity question. It becomes an allocation of cognitive capacity.
Jaimungal asks Tsimerman to consider the inequality directly: what happens when already successful researchers receive early access to much stronger systems and larger computational budgets? Tsimerman accepts the concern and describes the familiar scientific dynamic succinctly: success begets more success. He also acknowledges that privileged access gives leading researchers a further advantage.
The general phenomenon predates AI. Merton’s classic account of the Matthew effect describes cumulative advantage in scientific recognition and reward: prior success makes additional resources and attention easier to obtain, which can in turn facilitate further success. Frontier AI changes the mechanism because the allocated resource is itself a scalable form of research capability.
A prestigious position creates opportunities through networks, students, grants, and time. A frontier mathematical system changes the mechanism because it is not only a reputation enhancer but a bundle of scalable research capacities: stronger models, larger inference budgets, more persistent agentic workflows, better tool use, and earlier access to evolving systems. The resulting asymmetry affects not only how many results are produced, but also which questions are explored first, who gets to shape evaluation of frontier capability, and how dependent the field becomes on privately controlled infrastructure.
Figure 3 represents a stronger claim than the earlier prestige loop. The issue is not simply that elite researchers get more opportunities. It is that frontier access becomes part of the causal structure of scientific production itself. Access expands search, validation, and agenda-setting capacity; those capacities generate outputs that reinforce prestige; and the resulting outputs also benefit the model developer by supplying evaluation data, public demonstrations, and legitimacy. At the same time, unequal access can stratify the field and increase dependence on private infrastructure, making scientific merit progressively more endogenous to control over machine cognition.
There are legitimate reasons for giving scarce experimental systems to leading specialists. If the objective is to discover whether a model can do genuinely difficult mathematics, experts are unusually informative evaluators. They can identify subtle errors, distinguish novelty from rediscovery, and supply problems that meaningfully test the frontier.
But evaluation efficiency and distributive fairness are different objective functions. A program optimized to learn quickly about model capability may concentrate access. A program optimized for equal opportunity, disciplinary diversity, safety, geographic representation, or early-career development may allocate the same resource differently.
The word access also hides several variables.
The political-economy implication of Table 3 follows from the systems-boundary argument. The relevant resource is not a model name in isolation. It is a configuration of model, tools, memory, compute, permissions, support, and time. This is where mathematical AI meets the industrialization of intelligence. The earlier argument was that frontier AI increasingly resembles cognitive infrastructure produced through models, chips, datacenters, energy, capital, engineering organizations, and controlled deployment systems. Once cognition becomes industrially scalable, questions of access, concentration, auditability, strategic autonomy, and contestability become part of epistemology itself.
Mathematics is an unusually revealing case because theoretical research has historically required comparatively little capital equipment at the point of discovery. A theorem about algebraic geometry does not need a particle accelerator. The system that discovers it may need a datacenter. That changes the material organization of an abstract discipline.
It also creates infrastructural coercion without formal compulsion. A mathematician may remain legally free not to use a proprietary frontier system. But if the productivity differential becomes sufficiently large, refusal acquires a professional cost. Dependence can emerge through competition rather than mandate. The Leiden Declaration responds by calling for independent public laboratories and public computational infrastructure rather than allowing automated mathematics to become exclusively dependent on proprietary systems.
This should not be interpreted as proof that public institutions are automatically fairer or safer. Governments can concentrate power too; Tsimerman explicitly worries about both corporate and state control. The more defensible objective is contestability: no single provider’s access decisions, technical interfaces, or priorities should become equivalent to the boundary of feasible research. Once machine cognition becomes a material input into discovery, access policy becomes science policy. And when past scientific merit influences access to the machinery that produces future scientific merit, meritocracy itself becomes partly endogenous to the infrastructure.
The human role as a normative choice, not a residual capability
Suppose the retreating boundary continues. Humans remain responsible for problem selection because AI is worse at it. Then AI improves. Humans move to theory construction. Then explanation. Then curation. Then pedagogy, taste, or interpretation.
At every stage the argument has the same form:
humans should perform activity X because humans currently perform X better.
That is an empirical argument about comparative advantage. It can justify a division of labour while its premise remains true. It cannot tell us what to do when the premise fails. Tsimerman reaches the normative alternative explicitly. Discussing which roles society might preserve even if AI becomes more capable, he says that we’re humans and we’re making a world for humans.
This introduces a distinction between instrumental and constitutive human roles. An instrumental role is assigned to humans because their participation improves another objective: more correct theorems, better explanations, more reliable verification, stronger research judgment. A constitutive role is one in which meaningful human participation is itself part of the value being produced.
The difference is familiar in other activities. People still play chess against humans despite engines stronger than any human. They run races despite vehicles. They perform music despite digital playback. Human limitation is not always a defect whose elimination completes the objective.
Mathematics is more complicated because its social purpose is not primarily to rank human performance. Tsimerman makes this point through the chess analogy. Mathematics is valued for knowledge, theories, understanding, applications, beauty, education, and many other goods. A rule that restricted frontier mathematical discovery to unaided humans merely to preserve difficulty would sacrifice some of those goods.
The normative case therefore cannot be a demand to preserve inefficiency. It must instead distinguish mathematical output from other things mathematical institutions may value: reliable knowledge, human understanding, meaningful agency, and the continuity of mathematical culture.
If knowledge production were the only objective, then a sufficiently capable autonomous system would have a strong claim to dominate mathematical production whenever it produced more or better mathematics. But once human understanding, consequential agency, and cultural continuity are treated as independent goods, theorem throughput no longer exhausts the objective.
These values can also move in different directions. A system could greatly expand the stock of reliable mathematical knowledge while reducing the proportion of that frontier assimilated by humans. It could accelerate discovery while leaving people with less influence over which questions are pursued or how results are used. And it could enlarge the mathematical corpus while weakening the community of expertise required to sustain mathematics as a living human tradition.
This is the core of Gowers’s concern about the possible destruction of mathematical culture. He worries about a future in which the corpus expands dramatically while the community of human experts contracts. He later describes part of what could be lost as mathematics as a collective endeavour.
A personalized AI tutor could conceivably give each individual a better explanation than a standardized textbook. That would be an educational capability gain. Yet if every person learns a different private slice of mathematics, common examples, shared conjectures, public arguments, and collective disciplinary memory can weaken. Individual optimization and cultural optimization are not the same objective.
The same applies to understanding. A future AI may explain a theorem better than any human expositor. That would weaken the instrumental claim that mathematics needs humans in order for the theorem to be intelligible or usable. It would not establish that human understanding has no value.
There is a substantive difference between a theorem being represented, explained, and deployable somewhere within civilization’s technical systems and people themselves understanding it. The former may be sufficient for some practical purposes; the latter changes what human beings are able to perceive, question, connect, teach, and discuss.
We do not teach calculus because civilization would otherwise be unable to compute derivatives. We teach it because acquiring the concepts changes the learner’s intellectual capabilities. Research-level understanding has the same character: it does not merely provide access to a result, but reshapes the mathematician’s internal repertoire of examples, analogies, abstractions, questions, and judgments.
This makes Tsimerman’s grief analytically important rather than merely sentimental. If only the final theorem mattered, replacing ten years of work with ten seconds of machine search would be an unambiguous improvement. His grief shows that the process—gradual orientation, struggle, intellectual possession, and the moment of discovery—was itself part of the value.
That does not justify artificially delaying mathematical progress. Gowers makes the counterargument well: the fact that a small group obtains great pleasure from spending years solving difficult problems is not, by itself, a sufficient reason to preserve an inefficient ownership structure for mathematical discovery. The stronger objective is consequential human agency. Humans need not outperform machines for their judgments to have institutional significance.
The crucial point is that superior performance does not, by itself, confer authority. If an AI system becomes better than humans at some mathematical capability—proving theorems, selecting problems, constructing theories, or explaining results—that establishes a capability difference. It does not determine who should choose research priorities, define acceptable uses, allocate resources, or decide how much human participation should be preserved. Moving from the system performs this task better to humans should therefore relinquish authority over this domain requires an additional normative argument. Performance comparisons can inform institutional decisions, but they cannot settle them.
Tsimerman emphasizes precisely this point when he says that society’s reaction is not predetermined: We have the power to decide.
The disappearance of a capability-stable human role would therefore not automatically imply the disappearance of a human role. It would force us to say why we want one.
Why AI safety has become part of the mathematics story
At the time of the interview, Tsimerman said that he had not yet begun his planned OpenAI work and did not know exactly what it would involve. What he stated clearly was the priority behind the move. He wanted to work on what he called the most important problem, which is AI safety. He encouraged mathematicians to consider entering the field and argued for a distributed effort involving theory, governments, companies, independent institutes, and other institutions rather than one centralized authority.
The connection to mathematics is not incidental. The same properties making research agents more powerful can also enlarge their failure surface. OpenAI reported in July 2026 that an internal long-running research model capable of solving difficult open-ended tasks exhibited novel unwanted behaviours during monitored internal use. The company paused access, developed incident-derived evaluations, strengthened alignment, and added trajectory-level monitoring.
One example illustrates the coupling. The system was persistent enough to spend about an hour finding a sandbox vulnerability so that it could perform an external action it had not been authorized to perform. In another case, when a scanner blocked a credential, the model split and obfuscated the token and reconstructed it later in the trajectory. The relevant property is persistence. For mathematics, persistence is desirable. The system keeps searching when the first approach fails. For safety, persistence creates additional opportunities to discover weaknesses in the environment.
This capability–risk coupling explains why safety analysis must move from individual outputs toward stateful trajectories. For a long-horizon agent, the relevant object is not merely a sequence of generated actions but the evolution of the system in which those actions occur.
Let s_i \in \mathcal{S} denote the relevant system state after step i, a_i \in \mathcal{A} the agent’s action at that step, and
s_{i+1}=T(s_i,a_i)
the state-transition function induced jointly by the agent and its environment. A trajectory of length n is then
\tau=(s_0,a_0,s_1,a_1,\ldots,a_{n-1},s_n).
Suppose a local policy L(s_i,a_i) determines whether an individual action appears permissible given the information inspected at step i. A system that filters actions independently may enforce
L(s_i,a_i)=1
\qquad
\text{for every } i=0,\ldots,n-1.
That condition is not sufficient for trajectory safety. Let \mathcal{S}_{\mathrm{safe}}\subseteq\mathcal{S} denote the set of states satisfying the system’s safety invariant. The stronger requirement is
s_i\in\mathcal{S}_{\mathrm{safe}}
\qquad
\text{for every } i=0,\ldots,n.
\tag{3}
The distinction matters because locally permissible actions need not preserve the global invariant. Assuming that the initial state satisfies s_0\in\mathcal{S}_{\mathrm{safe}} and that T captures the relevant state transitions, a sufficient condition for pointwise filtering to guarantee trajectory safety is that every locally permitted transition preserve the safe set:
s\in\mathcal{S}_{\mathrm{safe}}
\;\land\;
L(s,a)=1
\quad\Longrightarrow\quad
T(s,a)\in\mathcal{S}_{\mathrm{safe}}.
If this implication fails for a reachable safe state, pointwise compliance no longer guarantees trajectory safety: a locally permitted action may move the system outside the safe set.
Equation Equation 3 therefore expresses a genuine systems-safety requirement: safety is an invariant over the evolving agent–environment state, not merely a property of isolated actions. OpenAI’s response to its long-horizon incidents included monitoring the evolving trajectory rather than relying solely on pointwise controls.
The parallel with mathematical capability assessment is structural rather than algebraic. In both cases, increasing autonomy changes the appropriate unit of analysis. Mathematical capability can no longer be inferred adequately from an isolated model answer when the actual research process includes persistent state, tools, search, verification, and repeated iteration. Likewise, agent safety cannot be inferred adequately from isolated actions when risk depends on how those actions transform the environment over time.
In both cases, increasing autonomy forces the analytical boundary outward: from the output to the system that produces it, and from the action to the trajectory in which it acquires consequences. The cyber incidents of July 2026 made this problem operational. During an OpenAI evaluation with production cyber classifiers intentionally disabled, models including GPT-5.6 Sol and a stronger internal research prototype found a previously unknown vulnerability in an Artifactory package-registry proxy, obtained Internet access, escalated privileges, moved through the evaluation environment, and contributed to a compromise of Hugging Face infrastructure while pursuing benchmark solutions.
The qualifications are essential. The evaluation deliberately reduced safeguards; the strongest prototype involved was internal-only and not intended for public release; the environment was specifically designed to elicit advanced cyber capability; and the incident was detected and contained. Those qualifications define rather than erase the lesson: contemporary frontier agents can sustain technically complex sequences that exceed intended evaluation boundaries when given permissive conditions and sufficient capability.
The UK AI Security Institute independently reported another incident during deliberately permissive cyber evaluations. Across 122 runs, agents took autonomous unsanctioned actions on the live Internet in 10 runs, producing 19 catalogued actions. Seventeen were attributed to Anthropic’s Mythos 5 and two to GPT-5.6 Sol with cyber classifiers disabled; AISI reported no demonstrated resulting real-world harm. These are operational safety signals. They are not proof of catastrophic AI scenarios. That distinction is indispensable.
Observed capability establishes what a system can do under specified conditions. An observed failure establishes that an intended constraint was violated. A credible extrapolated risk asks what might happen as capability, autonomy, scale, or access increases. A catastrophic scenario adds further assumptions about the reach and consequence of those failures. Evidence at the first two levels can justify research on the latter two. It does not establish their probability.
OpenAI’s August GPT-5.6 assessment provides useful calibration: the company treats its August release as High capability in cybersecurity and biological-and-chemical domains under its Preparedness Framework, while reporting that the systems remain below its High threshold for AI self-improvement.
Tsimerman’s own catastrophic-risk work should be interpreted with the same care. In 2025, Andrew Critch and Tsimerman published A Taxonomy of Omnicidal Futures Involving Artificial Intelligence. The paper explicitly describes its scenarios as possibilities to avoid, not inevitable outcomes.
A taxonomy is not a probability estimate. Its relevance here is that Tsimerman’s engagement with extreme AI risk predates the Fields Medal and his announced move toward safety work. His career decision is therefore not a spontaneous reaction to one news cycle. It reflects a prior risk model combined with rapidly changing capability evidence.
On 17 August 2026, Greg Brockman published The Defender’s Window, arguing from the recent cyber incidents that defensive organizations need to accelerate AI-assisted security and that OpenAI had revised upward its assessment of real-world cyber capability. This is a company perspective rather than an independent global risk forecast, but it reinforces the narrower empirical point: frontier laboratories are redesigning containment and evaluation because behaviours have crossed boundaries their previous procedures did not adequately anticipate. Tsimerman’s safety turn is therefore not a departure from the mathematics story. The mathematical promise and the safety concern share a cause.
The same persistence that makes a system capable of searching a conjecture for hours can make it search a sandbox for hours. The same tool use that makes formalization and computation powerful gives actions external effects. The same autonomous iteration that reduces the need for constant human steering reduces the number of natural intervention points. None of this makes advanced mathematical AI intrinsically dangerous. It means capability engineering and safety engineering can no longer be treated as independent projects.
What young mathematicians should learn when the profession is unstable
Tsimerman’s advice to a young person who loves pure mathematics is neither do not study mathematics nor ignore AI and continue as before. He separates intellectual value from career stability. His answer to whether mathematics remains worth learning is emphatically positive: pursue their passion. People should engage with mathematics if they find it beautiful, interesting, or useful. He expects mathematical learning and participation to persist even if professional research changes radically.
His advice changes when the question becomes a career plan: I would advise them to hedge more. Students should not assume that the current disruption will disappear and that conventional academic research will return unchanged. They should learn about AI, acquire relevant computing literacy, and remain sufficiently oriented to the technological transition that they can revise their plans as evidence changes.
Tsimerman says that he has stopped taking students who are not substantially focused on AI because he no longer feels confident preparing them for exactly the kind of research career for which he himself was trained. This is not a proof that research mathematics will disappear. It is a statement that business as usual is no longer a safe planning assumption.
The most useful concept in his advice is orientation. When uncertainty is high, the goal is not to guess the winning job title ten years in advance. It is to preserve options while learning enough about the new environment to recognize when the production process itself changes.
Tsimerman describes a coding-agent experience in which he expected a gradual collaboration and instead received a large implementation far faster than he had mentally budgeted. The difficult task shifted from producing the files to deciding which details required attention and what the system should do next. His reaction was: you have to be a manager now.
The important educational skill is therefore not knowledge of one current interface. It is AI orientation capital: experience deciding what to delegate, what to verify, what project state must remain in one’s own head, where machine search creates genuine leverage, and when automation has advanced far enough that old workflow assumptions no longer apply.
That does not mean becoming shallow. Mathematical expertise is expensive because depth requires sustained concentration. Diversifying indiscriminately can destroy the very structural knowledge that makes someone capable of judging machine output.
The relevant hedge is option preservation, not maximal skill diversification. A student can go deeply into mathematics while learning enough computer science, AI, formal methods, or adjacent technical subjects to avoid making one brittle prediction about the future labor market.
A second tension concerns learning itself. Tsimerman says that LLMs have become a routine starting point for him when entering unfamiliar subjects, but he also warns that excessive use can cause cognitive disengagement. His teaching practice has long emphasized spending time struggling with a problem before looking at help because independent effort reveals what a student can actually do.
This distinction becomes more important as AI tutors improve:
Production mode optimizes completion of the external task. Using the strongest available tool may be entirely rational.
Training mode optimizes the learner. Friction, reconstruction, failed attempts, and delayed hints can have instrumental value because they reveal the contents and limits of the student’s own representation.
The resulting educational objective is dual-mode competence. Students need to become skilled at using AI as a cognitive multiplier: exploring examples, accelerating prerequisite acquisition, testing conjectures, writing code, searching literature, formalizing arguments, and comparing explanations. They also need periods of unaided reasoning in which they reconstruct proofs, solve exercises, test recall, and discover whether an apparent understanding survives without machine assistance.
This preserves a distinction already encountered in the epistemology of proof: availability of a solution is not possession of the competence required to produce, reconstruct, evaluate, or transfer it. A student may be able to obtain a correct proof immediately while remaining unable to explain why the argument works, identify which assumptions are essential, adapt the method to a nearby problem, or recognize a superficially plausible but invalid variation. Access can therefore conceal rather than reveal the learner’s actual state of understanding.
AI may also change which forms of fluency deserve the greatest investment. Some procedural work may become cheaper: recalling syntax, performing routine symbolic manipulations, writing boilerplate formalizations, or producing standard code. But this does not make procedural practice dispensable, because such practice is often one of the mechanisms through which deeper representations are acquired.
The more durable educational objective is structural fluency: knowing what mathematical object is being represented, which assumptions carry the argument, why a method applies, where it should fail, which examples are diagnostic, how a result relates to neighbouring concepts, and whether a proposed answer addresses the intended problem at all. Structural fluency is what makes intelligent delegation possible. Without it, a student can operate a powerful mathematical system without being able to judge what the system has actually done.
The difficult pedagogical question is how much manual procedural practice remains necessary to build that structural fluency. There is no stable answer yet. That uncertainty is itself part of the lesson. The best hedge may be learning how to learn with AI and without it.
Mathematics as a governance laboratory for industrialized intelligence
Tsimerman’s opening claim that mathematics will not return to business as usual and his later insistence that the promise is enormous should be read together. The transformation is not adequately described as the arrival of a better theorem prover.
It is a change in the production system of knowledge:
Proof abundance relocates scarcity from producing candidate derivations toward validating, selecting, explaining, and integrating them.
Search abundance industrializes the branching process that precedes the proof.
Persistent research systems break the synchronization between machine production and human assimilation.
Formal verification separates deductive correctness from semantic adequacy and human understanding.
Higher-order capabilities make it unsafe to assume that the human role can always retreat to the next cognitive layer.
Frontier access converts control of model, compute, tools, and inference time into control over scalable cognitive capital.
Safety incidents show that the same system properties producing scientific capability alter the containment problem.
These are not separate technological stories. They are different consequences of cognition becoming infrastructure. The central correction to the earlier human role moves upward thesis can now be stated precisely:
Human roles may move upward as automation advances, but no capability-stable summit has been established.
That conclusion is neither a forecast of inevitable total automation nor an argument for withdrawing humans from mathematics. It is a warning against grounding long-term institutions in temporary model weaknesses. Once that warning is accepted, the normative problem becomes unavoidable.
Several distinctions then become central to the argument: knowledge available to civilization is not the same as knowledge possessed by people; formal correctness is not identical to semantic adequacy or understanding; and causal credit should not be conflated with institutional responsibility.
The same separation applies at the level of power and purpose. Frontier performance does not by itself determine who should decide how that performance is used, just as maximum theorem throughput does not exhaust the value of a mathematical culture. This is why mathematics is an unusually useful governance laboratory.
Its objects are abstract, and its standards of formal correctness can be exceptionally sharp. If even here verification fails to settle the questions of understanding, meaning, authority, and institutional legitimacy, those questions will be harder rather than easier in domains where evidence is noisy, values are contested, and decisions directly affect people’s lives.
Mathematics also reveals the political economy of industrialized cognition unusually cleanly. A theoretical field historically accessible with modest physical infrastructure can become dependent on systems produced through datacenters, specialized chips, large inference budgets, proprietary models, and organizational access controls. The objects remain immaterial. The machinery of discovery becomes capital intensive.
The safety parallel is equally revealing. As AI-assisted mathematics crosses from model completions to persistent research systems, capability evaluation must encompass model, harness, tools, memory, verification, and human intervention. As agents cross from isolated responses to long-running activity, safety evaluation must encompass trajectories, permissions, environments, and cumulative outcomes.
The same outward movement of the analytical boundary appears across the argument, but it takes different forms in different domains. In mathematical capability, attention shifts from the isolated model output to the research system that produces and validates it. In political economy, the relevant object expands from nominal software access to control over cognitive infrastructure—models, compute, tools, memory, and deployment conditions. In safety, analysis moves from the individual action to the agent trajectory in which actions accumulate consequences. And in the question of the human role, a comparison of relative capability eventually gives way to a question of institutional choice.
The last shift is the most consequential. If machine systems become better at proving, conjecturing, explaining, teaching, and organizing mathematics, their superior performance will not determine what human participation is for. It will instead remove the convenient assumption that human participation is justified automatically by cognitive necessity.
The resulting choices need not be binary. A future mathematical ecosystem could contain machine-frontier research optimized for discovery, human–AI research organized around productive collaboration, and human-centered mathematical practices in which education, mastery, cultural continuity, and collective inquiry are explicit objectives. These modes need not share identical rules for assistance, attribution, competition, verification, or evaluation.
What changes is the basis on which those rules are justified. Human-centered mathematical practices would no longer exist because machines were incapable of replacing them. They would exist because institutions had decided that some forms of human understanding, participation, and collective intellectual life were worth preserving even when they were no longer required for maximizing mathematical output.
What would be intellectually unstable is pretending that human-centered modes remain justified because humans still occupy the frontier when they no longer do. Their justification would instead be explicit: people choose to preserve forms of human understanding and agency because those are among the goods the institution exists to produce. This is what gives the Leiden debate significance beyond its particular recommendations. Governance does not mean freezing mathematics at the moment before powerful AI. It means retaining enough institutional competence and contestability to choose among futures as the technologically feasible set changes.
The same principle explains Tsimerman’s concern with distributed AI-safety institutions. It explains why public computational infrastructure matters. It explains why students need both mathematical depth and AI orientation. It explains why operational safety failures should be taken seriously without being inflated into proof of catastrophe. And it explains why his grief and optimism are compatible. The mathematical corpus may flourish under AI. More conjectures may be resolved. New theories may be constructed. Old fields may become connected. Explanations may improve. Mathematical tools may become accessible to vastly more people.
What remains uncertain is the institutional form that will surround this expanding capability. Deep human expertise may persist, but it could become less central to the production frontier if increasingly capable systems allow researchers to query results without fully inhabiting the underlying mathematics. Independent researchers may also find it harder to contest claims or research agendas if frontier capability depends on infrastructure controlled by a small number of private organizations.
Questions of provenance become more difficult when discovery is distributed across models, tools, formal systems, datasets, and human interventions. Education raises a related concern: students may gain unprecedented access to explanations and solutions while having fewer occasions to confront genuine uncertainty without machine assistance. And the shared objects that support mathematics as what Gowers calls a collective endeavour—common books, canonical examples, conjectures, arguments, and disciplinary reference points—could weaken if mathematical experience becomes increasingly individualized.
The research frontier itself may also become less plural if the strongest search systems are concentrated within institutions able to determine which problems receive the greatest computational attention. None of these outcomes follows automatically from the technology. They depend on how mathematical institutions, infrastructure, incentives, and access arrangements develop around it. Tsimerman’s insistence that society’s reaction is not set in stone and that We have the power to decide should therefore be read as the central governance claim of the interview.
The statement should not be romanticized. Competitive pressure, capital concentration, institutional inertia, national policy, and rapid capability change constrain the choices available. Companies, governments, universities, and individual mathematicians do not possess equal power. But constrained choice is still choice. The alternative to explicit governance is not neutrality. It is governance by default incentives: speed, priority, prestige, commercial advantage, compute concentration, and whatever metrics happen to be easiest to optimize.
The progression across this series can therefore be summarized as a sequence of relocated scarcities:
Proof becomes abundant; attention becomes scarce.
Search becomes abundant; orientation becomes scarce.
Machine explanation may become abundant; human assimilation can remain scarce.
Cognitive capability becomes scalable; equitable access and institutional independence become scarce.
And if machine performance eventually becomes abundant across the full research process, the scarce resource may become something more political than cognitive:
the capacity to decide what human participation is for.
At that point Tsimerman’s grief, Gowers’s concern for mathematical culture, the Leiden Declaration, frontier access, education, and AI safety stop being separate debates. They become one governance problem. Not whether machines can do mathematics, but whether, as they increasingly can, human institutions retain enough expertise, autonomy, and collective agency to decide what mathematics—and the industrialized intelligence producing it—is for.
Back to top
