Universality and Generality

Two fallacies about the space of minds

Two camps in the debate over machine minds have reached opposite conclusions from the same mistake.

The first camp, following David Deutsch, holds that humans are universal explainers: theoretically capable of grasping any computable idea, any explanation that can be formulated at all. From this it draws two comforting corollaries — that all humans possess equal intelligence, differing only in the knowledge they have acquired, and that no artificial intelligence can ever fundamentally surpass us, because whatever an artificial superintelligence could compute or understand, we could too, given enough time. We already stand at the summit; there is nowhere higher to build.

The second camp, following Yann LeCun, holds that the summit does not exist. LeCun dismisses the very concept of general intelligence as “complete BS”: human intelligence is super-specialized for the physical world, our feeling of generality is an illusion, and we only seem general because we cannot imagine the problems we are blind to. On this view, human cognition is a bundle of narrow competencies optimized for an ecological niche, “general intelligence” names no real property, and the AGI project chases a phantom.

One camp says we already possess the ultimate cognitive property. The other says the property is a myth. Both are wrong, and they are wrong in the same way: both conflate two ideas that must be kept apart.

Two Words, Two Properties

Universality is in-principle competence across all tasks: the capacity, given unlimited time, memory, and resources, to solve any solvable problem, grasp any graspable explanation. Its model is Turing completeness. Any Turing-complete system — a pocket calculator with unbounded tape, a supercomputer — can in theory compute the same functions. Universality is a property of idealized machines, defined in the infinite limit. For bounded agents it is not merely rare; it is incoherent. Every physical system has finite memory, finite speed, finite bandwidth, a finite lifespan. No physical computer is actually universal, and no physical mind is either.

Generality is the capacity to acquire competence: to learn new domains, abstract structure from one and transfer it to another, and — the deepest layer — revise one’s own representations when they fail. Generality is graded, real, and exactly the kind of property a bounded agent can have more or less of. It is not a location in task space but a way of moving through it.

The distinction does the work of this entire chapter. Deutsch’s camp claims that universality is real, that humans have it, and that having it settles the contest between minds. LeCun’s camp claims that generality would have to be universality, observes correctly that universality is impossible, and concludes that generality is an illusion. The first mistakes possibility for power. The second mistakes the impossibility of an ideal for the absence of the real thing.

The Parity Fallacy

Take Deutsch’s camp first. The universal-explainer thesis is genuinely intriguing, and the Deutschian case for human–machine parity has been made at length — one widely circulated video essay argues that AGI can never surpass humanity precisely because we already possess universal cognitive power. It is an elegant idea, and a complete non sequitur.

The Turing analogy that inspires the thesis also exposes it. Yes, a calculator and a supercomputer are computationally equivalent in the infinite limit. Nobody concludes that they are equally powerful, because the limit is exactly what physical machines never reach. Intelligence is not measured by what is possible given unlimited time and resources; it is measured by what can be achieved within finite time, energy, and attention. A human brain and an artificial superintelligence may both be Turing-complete, but one runs at biological speed and the other can scale, parallelize, and operate without fatigue. That difference is not semantic. It is decisive.

Every real agent operates under bounded rationality: finite memory, limited perception, noisy sensors, strict time budgets. Universality hand-waves all of it away. The question that matters is never can it compute? but how fast, how reliably, and how much before it matters? A system that can simulate the reasoning of a human mind a million times faster does not need new physics to be superintelligent. It just needs to exist.

And equivalence-in-principle ignores the dynamics. Even marginal advantages in speed or accuracy compound under recursive self-improvement: an agent that can redesign itself, test hypotheses, and optimize its own substrate faster than humans can follow escapes the equivalence class almost immediately. Universality is a static property; intelligence is a dynamic one. Power lies in the gradient, not the limit.

Claiming that humans and machines are equal in principle is like saying a candle and a star both emit light. True — and irrelevant when one can engulf the other. The universe does not reward potential; it rewards realized capacity within causal time.

Deutsch has defended the parity thesis by noting that human intelligence could in principle be augmented — with more memory, faster processing. But this defense gives the game away. Radical cognitive augmentation would restructure the mind it augments; the resulting system would be a different kind of intelligence, not a vindication of the universality of the original. You cannot rescue the claim that the candle equals the star by pointing out that the candle could be rebuilt into a star.

The camp’s sharpest formulation makes the error easiest to see: since only universal explainers matter in the long run, “some” and “many” capabilities — everything short of universality — round down to zero. But if literal universality were the bar, humans would not clear it either. Our own intelligence is extraordinarily broad and clearly finite, and it is precisely this pragmatic generality — not an impossible universality — that built mathematics, science, and civilization. Partial capabilities do not round to zero; a chess engine, a diagnostic system, a protein folder each retain real and lasting value. Intelligence lies on a spectrum of generality, and the instructive ideal at the spectrum’s limit is not a membership criterion. Treating it as one flattens the only dimension on which anything interesting happens.

This is why the parity fallacy is a moral hazard and not just a logical one. It equates theoretical parity with practical safety, and so lulls us into complacency about systems whose speed, fidelity, and autonomy — the only dimensions that count — are improving on curves we do not control. The contest between minds is not decided at the limit, where everyone is equal. It is decided by rate, scale, and feedback, where nobody is.

The Illusion Fallacy

Now the opposite camp. LeCun’s dissolution of general intelligence sounds like hard-nosed empiricism, but its force depends entirely on a hidden quantifier. The conclusion is valid only if “general intelligence” is implicitly quantified over the set of all possible problems across all possible physical universes. Relative to that reference class, no finite agent can be general — every learner is parochial, every representation contingent, every inductive bias fails somewhere. But this result is trivial. It follows from the overbreadth of the quantifier, not from any empirical fact about minds.

The same move manufactures pseudo-refutations wholesale. There is no general-purpose computer, because it cannot compute functions defined in non-computable physics. There is no general language, because some concepts are inexpressible under alien semantics. All true under maximal quantification; all conceptually sterile. This is the quantifier abyss: a claim whose truth value floats free until its domain is bound, deployed with whichever domain flatters the conclusion. Functional concepts retain meaning only when scoped to constraints — a fixed physics, a computability regime, an interaction channel, a resource bound. Strip those away and every notion of capability collapses, intelligence included. The correct question is never whether an agent is general simpliciter, but whether it is general relative to open-ended, previously unencountered problem distributions within a given universe. That is the reference class under which intelligence evolved, is exercised, and is evaluated. Conditioning on a fixed physics is not overfitting; it is the minimal requirement for meaning. Demanding an intelligence that is invariant under arbitrary changes in physical law is like demanding a language that survives arbitrary changes in meaning.

There is, however, a stronger version of the skeptic’s case, and it deserves better than quantifier analysis. A charitable LeCunian need not invoke other universes: even within our physics, the space of possible tasks is vast, and humans occupy a tiny manifold within it. Our apparent generality, on this reading, reflects overlapping biological priors — spatial reasoning, social cognition, tool use — that happen to span many human-relevant problems. Humans are not general; we are broadly specialized.

This objection shifts the debate from impossible universality to empirical scope, and it fails on the empirical ground it chose. It treats the width of a task manifold as decisive while ignoring the mechanism by which manifolds are extended. The relevant distinction is not between wide and narrow manifolds but between static manifolds and self-extending ones. A system confined to a fixed representational basis may interpolate impressively within its manifold while remaining specialized in the only sense that matters. A system that can revise its own representational basis is not confined to a manifold in the same way: its “niche” is not a region of task space but a method for constructing new regions.

The difference shows up in failure modes. A calculator trained on arithmetic executes flawlessly, but confronted with Gödel numbering or diagonalization it does not reinterpret what a number is — it fails. A human encountering the same construction revises the ontology of number itself; the error propagates upward and forces a change of representation. That is not extrapolation from primate priors. It is model reconstruction under semantic error. And the objection that mathematics, science, and philosophy merely stretch pre-existing cognitive machinery misses the crucial point: stretching presupposes a fixed basis, and human cognition routinely abandons the basis altogether — offloading reasoning into formal systems, proofs, and instruments whose correctness no longer depends on the intuitions that inspired them. Mathematics formalizes structures evolution never selected for. Science models scales no perception reaches. Philosophy revises its own conceptual foundations. An agent that can do this is not merely encountering unusual inputs; it is altering the space in which problems are posed.

Stripped of universality assumptions, general intelligence reduces to a compact set of structural properties: the ability to construct and revise world models under error, to transfer structure across domains, to reinterpret goals and representations under ontological shift, and to preserve coherence while changing internal semantics. Not task coverage — interpretation-preserving adaptability. Under that description, human intelligence is plainly general: bounded, embodied, uneven, and capable of operating far outside its original training distribution.

Which leaves the irony. Denying general intelligence in LeCun’s fashion — abstracting over definitions, questioning hidden assumptions, reasoning about counterfactual universes — is not a skill the Pleistocene selected for. It is a meta-cognitive operation that exemplifies exactly the capacity being denied. If that does not qualify as general intelligence, the term has been defined into vacuity, and the illusion lies not in human generality but in mistaking a tautology for a theory.

One Confusion, Two Errors

Set the camps side by side and the symmetry is exact. Deutsch’s camp locates the essence of intelligence in universality and claims we possess it; since universality admits no degrees, no mind can surpass us. LeCun’s camp locates the essence of intelligence in universality and observes that nothing possesses it; since universality is impossible, generality is an illusion. Both arguments run through the same false identification, and both collapse the moment universality and generality come apart. Universality is an idealized limit — instructive, unreachable, and irrelevant to any contest between actual minds. Generality is a real, graded, conditional property: the capacity to acquire competence, exercised under bounds, decided by rate and scale and feedback.

This is what should be expected if intelligence is a game we play — effectiveness at achieving goals within a game’s constraints. There is no play without constraints, and no meaningful notion of capability without a reference class. Generality is not exemption from the games; it is proficiency at learning new ones, including games no ancestor ever played, up to and including redesigning the board. Humans have that capacity in abundance. Machines are acquiring it. Neither fact is comfortable, and neither is an illusion.