Dropping the Purism: What Language Models Have Conquered, What Still Eludes Them, and Why General Intelligence Is a Continent No Species Has Fully Explored

This is the English translation of the original Portuguese essay ‘Largando o purismo: o que modelos de linguagem conquistaram, o que ainda lhes escapa, e por que a inteligência geral é um continente que nenhuma espécie — biológica ou digital — explorou inteiramente.’

There is a particular kind of intellectual dishonesty that afflicts both the critics and the enthusiasts of large language models: the tendency to evaluate a phenomenon against a standard it was never designed to meet. The critics point to failures of spatial reasoning, of persistent memory, of causal counterfactual thinking, and announce that therefore the systems are “mere” statistical pattern matchers. The enthusiasts celebrate emergent capabilities and claim these represent the dawn of artificial general intelligence. Both are looking at the same thing and seeing something it is not.

This essay attempts a more precise accounting: what have language models actually achieved, in terms that do not require either denigration or inflation? Where do they genuinely fail, and why? And what does an honest engagement with their limitations and capabilities suggest about the nature of intelligence itself?

I. What the Models Got Right (That Everyone Got Wrong)

The original scaling skeptics made a coherent argument: more parameters trained on more text would produce more elaborate next-token prediction, but not qualitatively new capabilities. Language, on this view, was a surface phenomenon — a statistical fingerprint of cognition, not cognition itself. Training on text was training on shadows, not on the things casting shadows.

This argument was, in retrospect, substantially wrong — and being honest about why it was wrong matters for understanding what language models actually are.

The error was to underestimate the cognitive density of language. Language is not merely a communication medium; it is a compression format for human thought accumulated over centuries. The structure of a well-formed argument encodes not just its conclusion but the logical relations between premises. A carefully explained concept carries within it the conditions that make it applicable and the conditions that would falsify it. A narrative contains, embedded in its temporal and causal structure, a model of human motivation and consequence that is extraordinarily difficult to acquire any other way.

When a language model is trained on sufficient quantities of this material at sufficient scale, something happens that was not fully anticipated: the model doesn’t just learn correlations between words. It learns the structural relationships that generate those correlations. It learns something about the shape of arguments, the structure of explanations, the logical constraints that make certain propositions follow from others. This is not nothing. In important respects, it is extraordinary.

II. The Topology of Failure

The failures of current language models are real and important — but they are not random. They have a topology that reveals something about the architecture of intelligence.

Models fail most systematically at tasks that require: (1) novel generalization from few examples, (2) persistent tracking of state across a long reasoning chain, (3) simulation of physical dynamics that depend on embodied intuition rather than verbal description, and (4) genuine counterfactual reasoning about situations that diverge significantly from the training distribution.

The common thread in these failures is instructive. They are all cases where performance requires a kind of representation that cannot be acquired by observing the verbal description of experience — only by having, or having something functionally analogous to, the experience itself. The model that correctly describes the physics of a bouncing ball has not simulated a bouncing ball; it has retrieved and composed textual descriptions of bouncing balls. For most purposes, these are equivalent. For some purposes — the ones where the failures cluster — they are not.

This is not a reason to dismiss the models. It is a reason to be precise about what they are and what they are not. They are extraordinarily capable text engines that have internalized a surprising amount of human knowledge and reasoning structure through the statistical patterns of language. They are not cognitive systems in the fullest sense — systems that have built their representations through active interaction with a world.

III. The Intelligence Question

The question of whether language models are “truly” intelligent is, in its most common forms, not a scientific question but a definitional one. If intelligence means efficient acquisition of new skills from minimal data (Chollet’s definition), current models are demonstrably less intelligent than a preschool child in several important domains, despite outperforming most humans in many linguistic and knowledge-retrieval tasks.

If intelligence means the capacity to produce useful outputs across a diverse range of tasks, current models are genuinely impressive — and the practical threshold for “useful” keeps falling as the capabilities of the systems increase.

The more interesting question is not “are these systems intelligent?” but “what do the patterns of their competence and incompetence reveal about the structure of intelligence?” And here the evidence is suggestive. The domains where models excel are the domains where human knowledge is densest in language — where thinking about a problem and writing about it tend to coincide. The domains where models fail are the domains where the knowledge that matters is not primarily linguistic — where the crucial representations are sensorimotor, temporal, or spatial in ways that don’t reduce to description.

IV. General Intelligence as Unexplored Continent

The subtitle of this essay describes general intelligence as “a continent no species has fully explored.” This is not rhetorical flourish; it is a literal claim about the epistemic situation we are in.

Humans are genuinely generally intelligent in many respects — we can acquire new skills in new domains, we can transfer knowledge across radically different contexts, we can improvise in genuinely novel situations. But we are also deeply domain-specific in ways we rarely acknowledge. Human intuition about probability is notoriously unreliable. Our reasoning about large numbers, about deep time, about systemic causation in complex systems — these are areas where unaided human cognition performs very poorly. We have built institutions (mathematics, statistics, scientific methodology) precisely to compensate for these limitations.

Current AI systems are genuinely generally capable in some respects — they operate across domains in ways that narrow AI systems do not. But they are specifically limited in ways that humans are not. The honest picture is not “humans have general intelligence and AI systems don’t.” It is “both humans and current AI systems have partial, domain-skewed profiles of capability, and neither has reached anything like the ceiling of what general intelligence could be.”

General intelligence is not a destination that has been reached; it is a direction that different systems are approaching from different angles. The fact that language models have gotten quite far in one dimension — the linguistic dimension — while remaining limited in others is not a failure of the project. It is a data point about the topology of that space.

V. What Comes Next

The most honest answer to the question of what comes next is: we don’t know. The history of AI has been repeatedly characterized by confident predictions from both optimists and pessimists that have been repeatedly falsified in both directions. Confident predictions about the imminence of AGI have been wrong since Minsky in the 1960s. Confident predictions about the impossibility of certain capabilities have been wrong with almost equal frequency — the claim that machines could never learn to play Go at the grandmaster level was confidently made by researchers who should have known better, and it was falsified within a decade.

What seems clear is that the current architecture — transformer-based language models trained on text — has produced genuine capabilities and genuine limitations, and that the limitations are not trivially fixable by more data and more parameters. Something structurally different is needed for the domains where current models fail most systematically. What that something is remains an open research question.

The appropriate response to this uncertainty is not the breathless enthusiasm that treats every GPT release as the dawn of a new era, nor the reflexive skepticism that treats every demonstration of capability as mere statistical mimicry. It is the kind of careful, cumulative attention to what systems can and cannot do, and why, that allows knowledge to actually accumulate — which is, not coincidentally, the same attitude that makes science work.

---

Originally published in Portuguese. Read the original here: Largando o purismo