tensorgami

When Ilya Sutskever told the audience at NeurIPS 2024 that the age of scaling was over, he was making a narrowly technical claim about the diminishing marginal returns of pre-training large language models on finite internet data. But the remark carries a weight far exceeding its intended scope. Read against the broader landscape of Western institutional life in the mid-2020s, Sutskever’s diagnosis of artificial intelligence becomes, almost inadvertently, a diagnosis of the civilisation that produced it. The ceiling that AI has struck is not merely architectural. It is epistemic, cultural, and ultimately civilisational. The machines cannot transcend the median because the culture that trains them has lost the capacity to value — and in some cases has begun to actively prosecute — whatever lies above it.

The technical facts are by now well-established. Pre-training data is finite; the high-quality text of the public internet has been substantially exhausted. Scaling compute without scaling data produces overfitting, not intelligence. Synthetic data loops risk model collapse. And the most striking empirical observation, the one Sutskever keeps returning to, is the brittleness gap: models that score brilliantly on benchmarks yet generalise, in his words, “dramatically worse than people.” A teenager learns to drive competently in ten hours. A trillion-parameter model stumbles over multi-step reasoning. The difference, Sutskever argues, is not data volume but the absence of internal grounding — a continuously running evaluative process that allows a human agent to self-correct mid-trajectory, to sense that something is wrong before articulating what. He reaches, perhaps surprisingly for a computer scientist of his stature, for the language of emotion: compact value functions, ancient and astonishingly functional, that compress millions of years of selection pressure into an immediate felt sense of whether a trajectory is promising or catastrophic (Sutskever 2024).

This is a remarkable concession from the technical vanguard, and it deserves to be taken seriously on its own terms before being situated in any broader argument. What Sutskever is describing is not sentiment in the colloquial sense — not warmth or sadness or enthusiasm — but something closer to what phenomenologists call pre-reflective awareness: a form of intelligence that operates beneath and prior to explicit propositional reasoning. The navigator’s feel for structure, the mathematician’s sense that a proof strategy is heading somewhere productive before any lemma has been written down, the chess grandmaster’s board intuition — these are all instances of the same underlying capacity. They are computationally meaningful precisely because they function as dimensionality reduction over an otherwise intractable state space. They are not irrational. They are hyper-rational, in the sense of encoding vastly more information than any explicit chain of reasoning could hold. But they are not legible. They cannot be decomposed into steps that each independently pass muster. And this illegibility, within the epistemic regime that now governs Western institutions, is fatal.

It is worth pausing to note that this insight — that the deepest form of intelligence is precisely the one that resists procedural decomposition — has been explored with remarkable sophistication not by Western philosophy of mind but by Japanese narrative art. Naoki Urasawa’s Pluto (2003–2009), a reimagining of Osamu Tezuka’s Astro Boy, constructs its entire dramatic architecture around the question of what separates a truly intelligent robot from one that is merely powerful. The answer is not processing speed. It is the capacity for grief, moral confusion, the felt sense that something about the world does not cohere. The robot detective Gesicht does not solve his cases through exhaustive search; he has hunches. His unease is computationally meaningful — a compressed signal that his world-model has detected an inconsistency before he can articulate it. Urasawa’s deepest move is to locate the “ultimate robot” not in maximised capability but in maximised capacity for internal grounding: the robot that can suffer, because suffering implies an internal model that expected something different from reality and must restructure itself in response. This is Sutskever’s compact value function rendered in narrative form, and the fact that it was articulated by a Japanese manga artist two decades before the co-founder of OpenAI arrived at the same conclusion from the engineering side is not incidental to the argument that follows.

For if we accept Sutskever’s diagnosis — that the ceiling on current AI is the absence of internal grounding, of a continuously running evaluative process that operates below the level of explicit propositional reasoning — then we must ask what kind of culture is equipped to pursue the breakthrough that lies beyond this ceiling, and what kind is not. And here the picture becomes bleak for the West, because the epistemic regime that now dominates Western institutional life is defined precisely by its refusal to acknowledge, let alone cultivate, the form of cognition that Sutskever identifies as missing.

The story of how this happened is long and has many tributaries, but its essential shape is simple. Over the course of the twentieth century, the West progressively narrowed its operational definition of intelligence to coincide with procedural rationality: the capacity to produce chains of reasoning whose individual steps are independently verifiable and communicable. This narrowing was driven by the convergence of several forces — logical positivism in philosophy, the formalisation of peer review in science, the expansion of managerial bureaucracy in both state and corporate institutions, and the rise of credentialism as the primary mechanism of social selection. Each of these forces selected for the same cognitive phenotype: the individual who can decompose their thinking into legible steps, present those steps to a committee, and defend each one under adversarial questioning. Call this the cartographer: the person who produces maps of intellectual territory that others can read.

What was lost in this narrowing was the complementary phenotype: the navigator. The mathematician who works inside a structure operationally, feeling the constraints and affordances directly, arriving at results that are correct before they can be formally justified. Srinivasa Ramanujan is the canonical example. His results arrived without proofs in the Western sense; they came from a process he described in terms that would have disqualified him from any modern grant panel. But they were correct. The results carried their own validation — but only to someone willing to evaluate the output rather than the process. G. H. Hardy could recognise Ramanujan’s work as genuine because Hardy himself possessed enough mathematical depth to evaluate claims on their intrinsic merits. The question facing the present moment is how many Hardys remain in a system that has been optimising for process-legibility for over a century.

Alexander Grothendieck presents the same problem in a different register. Grothendieck did not find schemes by processing existing mathematical literature more efficiently. He rebuilt the conceptual landscape of algebraic geometry from the ground up, replacing the field’s foundational vocabulary with one that made previously intractable problems dissolve. This is not pattern-matching at superhuman speed. It is the construction of new mathematical objects — new ways of seeing — that did not exist in any prior distribution of mathematical text. It requires sitting with a problem for years, allowing the shape of the solution to emerge from sustained contact with the deep structure of the mathematics itself, without any guarantee that the result will be legible to contemporaries. Modern Western institutions do not fund this. They do not recognise it. Increasingly, they do not even understand what it would mean to do so.

The consequences for artificial intelligence are direct. Reinforcement learning from human feedback — the dominant alignment technique — is structurally identical to peer review formalised as a loss function. The reward signal is the approval of human evaluators. The optimisation target is outputs that a panel of educated reviewers would rate as helpful, harmless, and — crucially — agreeable. This process systematically penalises exactly the properties that Sutskever identifies as missing: sharp structural claims, non-consensus conclusions, the kind of reasoning that feels wrong on first encounter because it has restructured the conceptual landscape in a way the evaluator’s prior framework cannot accommodate. The resulting models are, as one would predict, superb at synthesising existing consensus and packaging it fluently. They produce, with mechanical perfection, the cognitive output of the professional-managerial class: the Brookings white paper, the McKinsey deck, the well-sourced explainer with appropriate caveats. They caveat and nitpick and “consider all sides” and arrive, reliably, at the pre-approved conclusion. The distinctions they draw are not load-bearing — they do not change any structural analysis — they merely create the appearance of having thought carefully. This is intelligence as theatre, and the scaling ceiling means the theatre does not get deeper by making the stage larger.

The implications for the Western labour market are severe but also, in a sense, clarifying. What AI automates is precisely the layer of cognitive work that the professional-managerial class has been selling for the last four decades: the ability to process information, synthesise multiple sources, produce fluent and defensible outputs, and navigate institutional legibility requirements. This is the stratum of competence that justified six-figure salaries for knowledge workers across law, finance, consulting, media, and the administrative university. AI wipes it out — not by being smarter than these workers in any deep sense, but by being equivalently mediocre at incomparably lower cost. It is the best midwit simulator ever built. But it is, and on Sutskever’s own analysis will remain, a midwit simulator. It will not replace Mochizuki or Grothendieck, for the same reason that a library, however vast, does not write original books. The material is not there in the distribution.

This would be a merely interesting technical observation — certain jobs automated, certain jobs safe — were it not for the fact that the West appears to be systematically destroying its capacity to produce the cognition that lies above the automated layer. The evidence for this operates at every scale, from the institutional to the individual.

At the institutional level, gifted education programmes across the United States have been gutted over the past decade, explicitly in the name of equity. The argument, made openly in school board proceedings and education policy documents, is that identifying and cultivating exceptional talent is harmful because it produces visible inequality. The tall poppy must be cut. The pedagogical apparatus that once identified a Ramanujan-phenotype child and gave them access to advanced material has been replaced by one that identifies the same child and ensures they receive the same instruction as everyone else. Whatever one thinks of the distributional ethics involved, the cognitive consequences are unambiguous: the top tail of the ability distribution is no longer cultivated. Worse, it is taught to hide. The rational response for a gifted child in the contemporary American school system is to suppress visible excellence, to perform at the median, to wait for the credentialing pipeline to grant permission to be intelligent in the approved manner at the approved time. This is the precise opposite of the disposition that produces foundational breakthroughs.

At the individual level, the case of Amalvin Fritz — a seventeen-year-old UC Irvine student whose home chemistry laboratory was raided by the FBI’s hazardous materials team and the California National Guard’s Weapons of Mass Destruction Civil Support Team in February 2026 — crystallises the pathology with almost parodic clarity. Fritz, who entered college at fourteen, was synthesising cubanes, a class of molecules with promising pharmaceutical applications, using store-bought materials in his family’s garage. A maintenance worker reported the laboratory; the institutional response was to treat a gifted teenager’s chemistry set as a potential nerve agent facility. Hazmat suits. A week-long displacement from the family home. Confiscation of the boy’s phone. No charges filed, because there was nothing to charge. The experts quoted in subsequent news coverage did not defend Fritz; they noted, primly, that such experiments “should be done in a proper lab facility.” The message is not subtle: curiosity is permissible only when credentialed and contained. The garage — site of Apple, site of Hewlett-Packard, mythic locus of American innovation — is now a venue for federal investigation.

Fritz is not Ramanujan. He may or may not produce significant science. But the Fritz incident is diagnostic in the way that a biopsy is diagnostic: it reveals the character of the surrounding tissue. A society that responds to unsanctioned scientific curiosity by deploying its national security apparatus is a society whose immune system can no longer distinguish nutrients from pathogens. The system that once had enough slack to tolerate — even celebrate — the eccentric kid building things in a garage has tightened to the point where that figure is legible only as a threat. And the tightening is not incidental. It is structural. Zoning laws, liability regimes, regulatory frameworks, credentialing requirements — each individually defensible, collectively lethal to the conditions that produce foundational work. The garage door is closed. If you open it anyway, the FBI arrives in hazmat suits.

The contrast with contemporary China is instructive, though it must be drawn carefully to avoid the mirror-image error of idealisation. What China possesses at this historical moment is not a superior ideology of innovation but something more basic: institutional slack. The system is young enough, porous enough, and pragmatic enough that talent can still route around credentialism and be evaluated on output. The Shenzhen hardware ecosystem operates on this principle — iterate fast, ship, let the market adjudicate. In mathematics, the figure of Zhou Zhongpeng — a researcher without a doctoral degree whose work on mono-anabelian geometry has been taken seriously by the inner ring of the field — exemplifies a path to recognition based on the quality of the work itself rather than the institutional letterhead attached to it. This path is functionally closed in the contemporary Western academy, not because the cognitive raw material is absent but because the evaluation system no longer has a channel for it.

More deeply, the East Asian intellectual tradition — and the Japanese mathematical tradition in particular — never performed the epistemic surgery that the West performed on itself. It never severed intuitive, navigator-mode cognition from the domain of serious intellectual work. The Kyoto school of mathematics, the tradition within which Shinichi Mochizuki operates, does not treat the mathematician’s felt sense of structure as an embarrassment to be formalised away. It treats it as the medium of mathematical thought itself. Mochizuki’s Inter-universal Teichmüller Theory is, whatever one’s view of its correctness, an unmistakable product of this tradition: a sustained, solitary reconstruction of the foundations of arithmetic geometry that is illegible to the Western mathematical establishment not because it is wrong but because it presupposes a mode of engagement — years of immersive contact with the deep structure of anabelian geometry — that the Western system no longer cultivates and can barely recognise. The controversy between Mochizuki and Peter Scholze is, at one level, a technical dispute about a specific corollary. At another level, it is a collision between two epistemic regimes, one of which evaluates mathematics by whether it can be rapidly communicated to a committee, and another which evaluates it by whether it faithfully captures the structure of the mathematical objects themselves, regardless of how long the transmission takes (Mochizuki 2025).

If Sutskever is correct that the next breakthrough in artificial intelligence requires not more data but a fundamentally different architecture — one capable of internal grounding, continuous self-evaluation, the kind of compressed evaluative process that functions like emotion in biological systems — then the question becomes: which culture even possesses the conceptual vocabulary to look for it? The West has spent a century systematically discrediting the very cognitive mode that the problem demands. It has built evaluation systems, funding mechanisms, publication pipelines, and alignment procedures that all select against illegible, intuition-driven, foundationally creative work. It has, in the precise sense of the term, lobotomised itself — not by removing cognitive capacity from its population, but by removing the institutional conditions under which that capacity can express itself. The scalpel was not physical but procedural: the relentless demand that all thinking be legible, all outputs defensible, all processes decomposable into steps that a committee can approve.

Prometheus stole fire and was punished for it. The modern version is tidier: the fire is not stolen but regulated, credentialed, insured, and zoned. And when a seventeen-year-old lights a match in his garage, the gods do not send an eagle. They send the FBI.

---

References

Grothendieck, Alexander. 1985. Récoltes et Semailles: Réflexions et Témoignage sur un Passé de Mathématicien. Unpublished manuscript. Université des Sciences et Techniques du Languedoc.

Hardy, G. H. 1940. A Mathematician’s Apology. Cambridge: Cambridge University Press.

Mochizuki, Shinichi. 2025. “Essential Logical Structure of Inter-universal Teichmüller Theory.” RIMS Preprint Series 1968.

Serre, Jean-Pierre. 1955. “Faisceaux Algébriques Cohérents.” Annals of Mathematics 61 (2): 197–278.

Sutskever, Ilya. 2024. “Test of Time Award Address.” Lecture presented at the 38th Conference on Neural Information Processing Systems (NeurIPS), Vancouver, December 2024.

Tezuka, Osamu. (1964–1966) 2004. “The Greatest Robot on Earth.” In Astro Boy, vol. 3. Translated by Frederik L. Schodt. Milwaukie, OR: Dark Horse Comics.

Urasawa, Naoki, and Osamu Tezuka. 2003–2009. Pluto: Urasawa x Tezuka. 8 vols. Tokyo: Shogakukan.

Villani, Cédric. 2015. Birth of a Theorem: A Mathematical Adventure. Translated by Malcolm DeBevoise. New York: Farrar, Straus and Giroux.

Weil, André. 1940. Sur les fonctions algébriques à corps de constantes fini. Paris: Comptes rendus de l’Académie des Sciences.