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

Published on: 1 April 2026
Updated: 6 November 2026
Type: Essay
Reading time: 56 min read
Language: en

Translated from the Portuguese original 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.

I. Prologue: Astonishment and Sobriety

There is a productive tension running through the contemporary debate on artificial intelligence — a tension that is rarely named with precision, but that contaminates almost every public discussion of the subject. On one side are those who look at the linguistic capabilities of large language models (LLMs) and declare, with almost messianic enthusiasm, that artificial general intelligence is within reach, that language is thought and that, therefore, mastering language is mastering cognition. On the other side are the skeptics — many of them from cognitive biology, robotics and the philosophy of mind — who dismiss LLMs as “stochastic parrots,” sophisticated automata that simulate understanding without understanding anything at all.

Both positions are intellectually lazy.

The thesis of this essay is that the linguistic capability of modern language models is a genuine achievement of intelligence — not a trivial simulation, not a trick of statistical mirrors, but a computational accomplishment that deserves serious epistemological recognition. At the same time, this achievement illuminates, by contrast, the vastness of what general intelligence actually involves: capacities for action in the world, learning from direct experience, interventional causal reasoning, real-time adaptation to unpredictable physical environments — domains in which animals such as crows, chimpanzees, octopuses and even ants display a sophistication that LLMs lack and that no purely linguistic architecture, however many parameters it contains, can reach on its own.

General intelligence is not a one-dimensional spectrum with a “dumb” end and an “intelligent” end. It is a multidimensional topography — a space with independent axes that include symbolic manipulation, spatial reasoning, social cognition, motor planning, causal learning, metacognition, temporal adaptation and many others. Different species and different artificial systems occupy distinct regions of this topography. LLMs occupy an extraordinary region on the axis of language and abstract symbolic manipulation. Crows occupy an extraordinary region on the axes of causal planning, tool use and sequential cognition. Chimpanzees dominate social cognition, visual working memory and intragroup political negotiation. Octopuses are masters of distributed motor control and morphological adaptation.

None of these agents is “more intelligent” than the others in absolute terms. Each is brilliant in its region of the topography and limited in the rest. The question of artificial general intelligence is, precisely, the question of how to integrate these regions into a single coherent system — and that is the hardest question in cognitive science, in AI and, possibly, in human civilization.

II. What Language Models Really Do

The pattern machine that exceeded the sum of its parts

To understand why the linguistic capability of LLMs is an achievement of intelligence — and not a mere statistical trick — one must first understand what these systems really do, without mystification and without deflation.

A language model, in its most basic formulation, is a system that learns the conditional probability distribution of the next word (or token) given a context of previous words. The goal of training is to minimize the cross-entropy loss between the model’s predictions and the actual tokens observed in a massive corpus of text. Formally, for a sequence of tokens $x_1, x_2, \ldots, x_T$ , the model learns parameters $\theta$ that maximize:

$\mathcal{L}(\theta) = \sum_{t=1}^{T} \log P_\theta(x_t \mid x_1, \ldots, x_{t-1})$

Described this way, the task seems prosaic — almost banal. Predict the next word. Any n-gram model from the 1990s did this, with mediocre results. What changed?

What changed is that, when this task is carried out at sufficient scale — with corpora of trillions of tokens, hundreds of billions of parameters and the Transformer architecture with multi-head attention — the system develops internal representations that cannot be reduced to mere memorization or matching of superficial patterns. Research in mechanistic interpretability (Elhage et al., 2022; Olsson et al., 2022; Nanda et al., 2023) has shown that Transformer models form specialized internal circuits: induction heads that perform abstract pattern matching, compositional reasoning circuits that decompose multi-step problems, and representations of semantic relations that preserve relational structure in a way that mirrors, in many respects, human conceptual relations.

The work of Kenneth Li et al. (2023) on the Othello-GPT model is particularly revealing. A Transformer trained only to predict sequences of Othello moves — with no information about the board, the rules of the game, or the notion of “position” — spontaneously develops an internal representation of the board state. The model was not given this representation; it built it as a by-product of optimizing prediction. This is not memorization. This is the emergence of representational structure from predictive pressure — a phenomenon that deserves to be called understanding, even if in a provisional and qualified sense.

Beyond prediction: reasoning, analogy and composition

The question of whether LLMs “merely predict the next word” is, in a way, analogous to the question of whether the human brain “merely transmits electrochemical impulses between neurons.” The description is mechanistically correct and functionally empty. What matters is not the primitive mechanism, but the capabilities that emerge from orchestrating that mechanism at scale and with the right architecture.

Modern LLMs demonstrate capabilities that, if observed in any other system — biological or artificial — would immediately be recognized as intelligent:

Analogical reasoning. When an LLM correctly answers that “Tokyo is to Japan as Paris is to France,” it is not retrieving this information from a pre-stored table. It is computing relations in a high-dimensional semantic space, and this computation is structurally isomorphic to processes of analogical reasoning that cognitive psychologists have studied for decades.

Semantic composition. LLMs can interpret and generate sentences that have never existed in any training corpus — sentences with genuinely novel semantic combinations. “The digital twilight spilled pixels melancholically over the terminal window” exists in no training text, but an LLM can process this sentence, extract its metaphorical meaning and even continue the style. This implies productive composition — the ability to generate new meaning from the combination of smaller units — which is one of the central criteria of linguistic competence according to Chomsky and the generative tradition.

Translation between domains. LLMs can translate not only between languages, but between registers, styles, technical domains and even levels of formality. Explaining a concept of quantum physics in language accessible to a ten-year-old, or translating a legal argument into colloquial language, requires a form of understanding — the ability to extract the deep structure of a content and re-express it in a different surface structure.

Chain reasoning. Presented with problems that require multiple steps of reasoning, models equipped with chain-of-thought prompting (Wei et al., 2022) demonstrate the ability to decompose problems, hold intermediate state and reach conclusions that are not accessible through simple statistical shortcuts. Models with extended reasoning — such as OpenAI’s o1 and o3 families and Anthropic’s models with extended thinking — go further still, devoting explicit computation to deliberation before producing an answer.

None of these capabilities is perfect. LLMs make reasoning errors, suffer from hallucinations, fail at problems that require truly long reasoning, and lack direct perceptual grounding. But the imperfection of a cognitive capability does not invalidate it as a cognitive capability. Four-year-old children also make systematic reasoning errors, fail Piagetian conservation tasks and confuse appearance with reality — and no one argues that four-year-olds are not intelligent.

III. Language as a Dimension of Intelligence: The Affirmative Argument

What the philosophy of mind tells us

The question of whether linguistic competence constitutes “real intelligence” cannot be answered without first facing a more fundamental one: what is intelligence?

The history of psychometrics and of the philosophy of mind offers us a rich but fragmented tradition. Charles Spearman postulated the g factor — a general cognitive capacity underlying all specific intellectual performances. Louis Thurstone disputed this, proposing seven independent primary abilities. Howard Gardner expanded it to eight (or nine) multiple intelligences. Robert Sternberg offered his triarchic theory — analytical, creative and practical. More recently, François Chollet proposed, in the paper “On the Measure of Intelligence” (2019), a formal definition of intelligence as skill-acquisition efficiency, emphasizing generalization to new tasks with minimal prior experience — a definition that directly inspires the ARC (Abstraction and Reasoning Corpus) benchmark.

Whatever framework one adopts, practically everyone recognizes that linguistic competence is, at the very least, a significant component of human intelligence. Language is not merely a communication channel; it is a cognitive scaffold — a technology of thought that makes possible:

Abstraction and categorization. Language enables the creation of abstract categories (“justice,” “entropy,” “recursion”) that transcend immediate perceptual experience. Without language, thought would be tied to the concrete and the present — to what can be directly perceived and manipulated. Vygotsky argued extensively that language does not merely express thought, but constitutes it at its most abstract levels.
Memory and cultural transmission. Language enables the encoding of experience in transmissible and cumulative formats. Without it, each generation would start from scratch. Science, philosophy, law, engineering — the whole apparatus of civilization — depends on the capacity to encode, store, transmit and recombine linguistic knowledge.
Hypothetical and counterfactual reasoning. Language makes it possible to think about what is not — about possibilities, alternative futures, hypothetical scenarios. “What would happen if…” is a linguistic construction with no direct equivalent in non-verbal cognition, and it is the foundation of planning, experimental science and fiction.
Metacognition. Language makes it possible to think about one’s own thinking — to narrate cognitive processes internally, evaluate them, correct them. The “inner dialogue” that psychologists identify as crucial for cognitive self-regulation is, essentially, language directed at oneself.

If language is all this for human intelligence, then a system that masters language with the proficiency of a modern LLM possesses, at the very least, a significant fraction of what we call intelligence. Not all intelligence. Not general intelligence. But a non-trivial slice — an achievement that should not be diminished by the fact that other slices are still absent.

The stochastic parrot argument and its limitations

The “stochastic parrot” metaphor, popularized by Emily Bender and colleagues in 2021, argues that LLMs do not understand the language they produce — that they merely reproduce statistical patterns without underlying meaning. It is a powerful metaphor, but one that suffers from at least three serious problems.

First, the problem of the criterion. If we demand “genuine understanding” as the criterion for attributing linguistic intelligence, we need an operational criterion for what constitutes genuine understanding — and no criterion proposed to date manages to distinguish, in a non-arbitrary way, between human understanding and the performance of the best LLMs on highly complex linguistic tasks. The Turing test is not sufficient (it is easily gamed). Reading-comprehension tests are insufficient (LLMs already surpass them). The objection always retreats to something like “but the model doesn’t really understand” — and that “really” remains undefined. We are facing what Daniel Dennett would call an intuition pump with no empirical content: a strong intuition that something is missing, with no ability to specify what.

Second, the problem of reverse projection. If human understanding of language is implemented by neurons firing patterns of electrochemical activation, it is not immediately clear why activation patterns in a Transformer could not implement something functionally analogous. The objection that “it’s just statistics” projects a dichotomy between mechanism and meaning that may not be justified. After all, it is possible that understanding is, at the implementational level, a very sophisticated form of statistical pattern processing — in brains as in artificial neural networks.

Third, the empirical problem. The predictions of the stochastic parrot model are empirically testable, and many of them have failed. If LLMs were mere reproducers of superficial patterns, they should not be able to solve new problems that require composing concepts never combined in training, to extract abstract principles from a few examples (in-context learning), or to exhibit emergent behavior that was neither programmed nor predicted. But they do all these things — imperfectly, unevenly, but demonstrably.

This does not mean the stochastic parrot argument has no merit. It does. It points to something real: the linguistic competence of LLMs, however impressive, lacks anchoring in the world — grounding — and this lack has profound cognitive consequences. But the argument errs in concluding, from this lack, that no intelligence is present. The correct conclusion is subtler: what is present is a form of intelligence that is partial, specialized and, in certain important respects, ungrounded.

IV. The Intelligence That Does Not Speak: Action, Body and World

The gap between saying and doing

Here we come to the heart of the problem. LLMs are extraordinarily competent at symbolic manipulation — but general intelligence, as observed in the most cognitively sophisticated biological organisms, involves far more than symbols. It involves action in the physical world: perceiving dynamic environments, planning motor sequences, manipulating objects, adapting to unexpected outcomes, learning causally from interventions — and doing all of this in real time, under energetic and temporal pressure.

Consider the following asymmetry:

An LLM can generate a ten-thousand-word essay on the tool-use strategies of New Caledonian crows, correctly citing the studies of Hunt (1996), Weir et al. (2002) and Taylor et al. (2007). It can describe in detail how these crows bend wire to make hooks, how they fabricate compound tools from multiple materials, how they plan sequences of tool use of up to three steps.

But the LLM itself cannot bend a wire.

Not because of a hardware limitation — after all, it has no body. The point is deeper: even if coupled to a robotic body with precise actuators and high-fidelity sensors, the LLM, qua language model, does not possess the internal representations needed to guide fluid physical action in the world. It has no dynamic model of the physical world, no representation of forces and torques, no sensorimotor loop that allows a crow to feel the resistance of a twig and adjust its beak within milliseconds.

This asymmetry is not a technical curiosity. It points to a fundamental difference in the nature of the intelligence involved. Linguistic intelligence operates in the space of symbols — a discrete, compositional, temporally slow, deliberative space. The intelligence of action operates in physical space — a continuous, dynamic, temporally urgent, sensorimotor space. Both are forms of intelligence. Both involve representation, inference and adaptation. But the computational substrates and the selective pressures that shaped them are radically different.

The crow and the meta-tool

New Caledonian crows (Corvus moneduloides) are possibly the most sophisticated tool engineers in the non-human animal world. What makes their behavior particularly notable for the discussion of intelligence is the presence of several capabilities that LLMs lack:

Projective fabrication. Crows do not merely use tools they find — they make them. More impressively, they shape tools in ways that anticipate future use: they bend wire into hook shapes, they cut pandanus leaves into stepped forms that serve as efficient probes. This implies an internal representation of the goal (extracting a grub from a hole), of the means (a tool of a specific shape), and of the causal relation between the shape of the tool and its effectiveness at the task. The crow does not need language for this. It operates with what psychologists call imagistic representation — a mental model of the physical world that allows simulation and planning without linguistic symbolism.

Sequential planning. In controlled experiments (Taylor et al., 2007; Wimpenny et al., 2009), crows demonstrated the ability to solve meta-tool problems — using one tool to obtain another tool to obtain food. In some cases, sequences of three tools were used. This is genuine prospective planning: the animal must mentally represent a causal chain of three links, in which the result of each action is a precondition for the next. This capacity is analogous to what in AI we call hierarchical task planning — and crows carry it out without language, without formal training, often on the first attempt.

Innovation. Betty, the crow made famous in the studies of Weir et al. (2002), spontaneously bent a straight wire into a hook to retrieve a bucket of food from a vertical tube — a behavior that had never been observed and could not have been learned by imitation, since wire does not exist in the species’ natural habitat. This is genuine innovation: the generation of a novel behavioral solution to a problem never before encountered, from reasoning about the physical properties of the available materials.

Causal learning by intervention. Crows learn about the world not only by observation, but by active intervention — manipulating objects, testing implicit hypotheses, observing results and updating their behavior. This is, functionally, the experimental method — the same logic that underlies science, operating in an eight-gram brain without a single word.

When we compare these capabilities with those of an LLM, the asymmetry is striking. The LLM can describe all of these capabilities in sophisticated language. But it cannot execute them. It cannot plan a sequence of physical actions, anticipate the mechanical properties of a material, improvise a solution to a new physical problem, or learn from direct interaction with the world. Its intelligence is deeply descriptive and declarative; the crow’s is deeply operative and procedural.

The chimpanzee and politics without words

If crows are the engineers of the animal world, chimpanzees (Pan troglodytes) are the politicians. Their social and cognitive capabilities offer an equally instructive contrast with LLMs.

Social cognition and theory of mind. Chimpanzees demonstrate sophisticated forms of understanding of the mental states of other individuals. Studies by Hare et al. (2001) and Call & Tomasello (2008) showed that subordinate chimpanzees track what a dominant individual can or cannot see, adjusting their foraging behavior accordingly. They practice tactical deception — hiding food from rivals, suppressing excitement vocalizations in the presence of dominants, using distractions to divert attention. This requires modeling the perceptual and intentional states of other agents — what philosophers call theory of mind and what, in AI, remains an open problem of formidable difficulty.

It is true that LLMs demonstrate a certain capacity to reason about mental states in linguistically described scenarios — for example, solving variants of the Sally-Anne test when presented in text. But there is a crucial difference: the LLM reasons about mental states described in text; the chimpanzee perceives and models mental states in real time, in dynamic social interactions, with immediate consequences for its survival and status. The computational, temporal and ecological pressure is of a completely different nature.

Visual working memory. The study by Inoue & Matsuzawa (2007) is one of the most surprising results in comparative cognition. Young chimpanzees, trained on a sequential numerical memory task (in which digits are briefly displayed and then masked, and the subject must touch the positions in numerical order), consistently outperformed adult humans — with exposure times of only 210 milliseconds. This suggests a visuospatial working-memory capacity that, in at least one specific dimension, exceeds the human one. LLMs, operating exclusively on sequential text, have nothing analogous.

Culture and social transmission. Different chimpanzee populations possess distinct behavioral traditions — different ways of using tools, different hunting techniques, different hygiene and greeting practices — that are socially transmitted across generations. Whiten et al. (1999) documented at least 39 cultural behaviors in wild chimpanzees. This is culture — cumulative social learning — operating without articulated language.

Negotiation and coalition. De Waal (1982, 2007) extensively documented the power dynamics of chimpanzee communities: alliance formation, post-conflict reconciliation, exchange of favors, punishment of betrayers, manipulation of third parties. These interactions involve long-term strategic planning, modeling of multiple simultaneous agents, and evaluation of expected utility in complex social scenarios — capabilities that map directly onto problems of multi-agent game theory that remain challenging in AI.

The octopus: intelligence without centralization

The octopus (Octopus vulgaris and related species) deserves special mention because it illustrates a form of intelligence so radically different that it challenges our intuitions about what cognition can be. With two thirds of its neurons located in the arms — not in the central brain — the octopus implements a distributed cognitive architecture that has no parallel in vertebrates. Each arm has significant sensorimotor autonomy: it can explore, manipulate objects, solve local problems, and coordinate with other arms without constant central supervision.

Octopuses demonstrate observational learning (Fiorito & Scotto, 1992), tool use (Finn et al., 2009 — carrying coconut shells for future use as shelter), mechanical puzzle solving, complex spatial navigation, and even behaviors suggestive of play — a classic indicator of cognitive complexity.

What the octopus teaches us about intelligence is that it does not require a single architectural blueprint. Intelligence can be centralized (as in primates) or distributed (as in octopuses). It can operate in symbolic spaces (as in humans and LLMs) or in purely sensorimotor spaces (as in many invertebrates). Intelligence is a design space, not a single blueprint.

V. The Topography of Intelligence: A Conceptual Framework

Axes, not ladders

The dominant metaphor for intelligence in popular culture — and, unfortunately, in much technical discussion — is the ladder: a linear progression from “less intelligent” to “more intelligent,” with human beings comfortably at the top and AI systems somewhere below, climbing. This metaphor is deeply misleading.

A more faithful representation of reality is that of a multidimensional space — what I am calling the topography of intelligence. In this space, each axis represents a relatively independent cognitive capability, and each agent (biological or artificial) is a point (or, more precisely, a probabilistic cloud) in that space. Without claiming exhaustiveness, we can identify at least the following axes:

1. Symbolic and linguistic manipulation. The ability to operate with discrete symbols — words, numbers, formulas, code — composing complex meanings from simple units. LLMs are exceptional on this axis. So are humans, but with a much lower processing bandwidth (we cannot process millions of tokens per hour). Most non-human animals operate with limited forms of symbolic communication: bee dances, alarm calls in vervets (Seyfarth et al., 1980), dialects in cetaceans.

2. Causal and interventional reasoning. The ability to understand cause-and-effect relations and, crucially, to intervene in the world to test and explore those relations. Crows, chimpanzees and humans are strong on this axis. LLMs have some capacity for verbal causal reasoning, but cannot perform causal interventions in the world and therefore cannot learn causally from direct experience. Judea Pearl distinguishes three levels of causal cognition — association, intervention and counterfactuals — and has argued that LLMs operate predominantly at the first level, with limited forays into the third (via language), but without genuine access to the second.

3. Spatial cognition and navigation. The ability to represent three-dimensional space, navigate environments, mentally rotate objects, plan routes. Many animals — from ants to pigeons to cetaceans — possess extraordinary spatial navigation capabilities. Homing pigeons navigate thousands of kilometers using a combination of the magnetic field, solar position and path integration. Bees communicate distances and directions with precision via the waggle dance. LLMs have no native spatial representation — their forays into this domain depend on the linguistic encoding of spatial information, which is inherently poorer than direct visuospatial representation.

4. Social cognition and theory of mind. The ability to model the mental states, intentions and knowledge of other agents. Chimpanzees, dolphins, elephants and corvids demonstrate sophisticated forms of social cognition. LLMs can simulate reasoning about mental states when these are described linguistically, but they lack the real-time social perception that allows a chimpanzee to read a rival’s body posture and adjust its strategy within milliseconds.

5. Motor control and dexterity. The ability to plan and execute complex motor sequences, coordinate multiple effectors, and adapt movements in real time to perturbations. This dimension is where animals shine most unambiguously: the flight of a hummingbird, the leap of a jumping spider, a chimpanzee’s manipulation of tools. LLMs lack this capability (by definition, being purely computational systems). Even state-of-the-art robotic systems with end-to-end learning (such as those from Chelsea Finn’s lab at Stanford, or the work of Sergey Levine at Berkeley) remain far short of animal dexterity and adaptability.

6. Few-shot learning and generalization. The ability to acquire new skills from a few examples — what Chollet emphasizes as central to intelligence. Here the picture is mixed: LLMs demonstrate impressive in-context learning in linguistic domains (Brown et al., 2020), but their ability to generalize to truly new domains is limited. Animals demonstrate one-shot learning in ecological contexts — a rat that eats a novel food and then falls ill will avoid that food forever (Garcia & Koelling, 1966), and crows that observe a human using an object as a tool can replicate the behavior within minutes.

7. Metacognition and uncertainty. The ability to assess one’s own confidence, recognize the limits of one’s knowledge, and adjust behavior accordingly. There is evidence of metacognition in rhesus monkeys (Smith et al., 2003), dolphins (Smith et al., 1995) and crows. Well-calibrated modern LLMs can, to some degree, modulate their expression of confidence, but they still suffer from systematic overconfidence and hallucination — indicating that their metacognition is, at best, rudimentary.

8. Temporal adaptation and continual learning. The ability to learn continuously over time, integrating new experiences with prior knowledge without forgetting the past (the problem of catastrophic forgetting). Animals do this naturally — it is the basic function of biological memory. LLMs, once trained, are static: they do not learn from new interactions (except in limited scenarios of fine-tuning and in-context learning within a session).

Where each agent stands

If we visualize this topography as a radar diagram (with each axis as a spoke), the profiles of the different agents would be dramatically distinct:

Modern LLMs would have very high values on axis 1 (symbolic manipulation), moderate on axes 2 (verbal causal reasoning), 4 (simulated social cognition) and 6 (linguistic in-context learning), and low to null on axes 3, 5, 7 and 8.

Crows would have high values on axes 2, 3, 5 and 6, moderate on axis 4, and low on axis 1.

Chimpanzees would have high values on axes 2, 4 and 5, moderate on axes 3, 6 and 7, and low on axis 1 (with some rudimentary symbolic capacity).

Humans would have high to very high values on all axes — which is probably the most precise definition of “general intelligence” we can offer: not the maximum on any individual axis, but an exceptionally broad and reasonably high coverage of all of them.

This visualization makes immediately apparent why it is so hard to argue that LLMs “are” or “are not” intelligent. The answer depends entirely on which axis is being evaluated. On the linguistic axis, they are at the human level or above on many tasks. On the axis of physical action, they are below an ant.

VI. The Grounding Problem: The Gap Between Symbol and World

Harnad and the symbolic Chinese room

The “symbol grounding problem,” articulated by Stevan Harnad (1990), is perhaps the most serious critique that can be directed at the claim of general intelligence for LLMs. The problem is this: if all of a system’s representations are symbolic — if “red” is merely a relation to other symbols such as “color,” “blood,” “traffic light” — then the system has no genuine semantic content. Its symbols mean nothing “in themselves”; they merely point to one another in a self-referential carousel.

The most famous version of this argument is John Searle’s Chinese Room (1980): an English speaker, locked in a room, follows rules for manipulating Chinese characters to produce answers that seem fluent to an outside observer who reads Chinese. The English speaker does not understand Chinese; he merely manipulates symbols. Analogously, Searle argues, a computer that processes language does not understand language — it merely manipulates symbols.

Searle’s argument has been extensively debated and, in my view, is not conclusive — the systems reply points out that understanding may reside in the system as a whole, not in any individual component, just as consciousness resides in the brain as a whole, not in any individual neuron. But Harnad’s argument is more incisive because it does not depend on intuitions about consciousness or “genuine” understanding. It simply observes that purely symbolic systems lack perceptual anchoring — and that this anchoring is a constitutive part of meaning.

For a human, “red” is not just a node in a semantic network. It is a perceptual experience — a visual quale — that anchors the concept to a category of sensory stimuli. For a crow, “flexible twig suitable as a probe” is not a linguistic concept; it is a perceptual-motor category that integrates visual information (length, thickness, curvature), tactile information (flexibility, resistance) and functional information (suitability for extracting grubs). This anchoring allows the concept to guide action in ways that purely symbolic concepts cannot.

LLMs lack this anchoring. Their concepts are entirely relational — defined by their relations to other concepts in embedding space. This does not prevent them from manipulating concepts with extraordinary competence in purely linguistic domains. But it prevents them from using concepts to guide action in the world in ways that require understanding the physical, causal and dynamic properties of the referents.

The pragmatist objection and its implications

There is, however, an important pragmatist objection to this line of argument. Pragmatist philosophers such as William James and John Dewey would argue that the meaning of a concept is constituted by its practical consequences — by the set of actions and inferences it makes possible. If an LLM can use the concept of “red” to make correct inferences in complex linguistic contexts — “the red light means stop,” “redshift indicates galactic recession,” “a red herring is a rhetorical distraction” — then the concept has meaning for the system, even if that meaning is narrower than human meaning (which also includes the perceptual dimension).

This pragmatist position is, in my view, essentially correct — but with a crucial qualification. The meaning of an LLM’s concepts is functionally complete within the linguistic domain and functionally incomplete when it comes to guiding action in the physical world. An LLM can reason correctly that “if the light is red, one must stop” — but it cannot perceive a red light, distinguish it from a red billboard, estimate its distance, and initiate a braking sequence. Inferential competence is present; perceptual-motor competence is absent.

This does not invalidate the inferential competence. But it illuminates its limits.

VII. What Animals Know Without Saying: Non-Verbal Cognition and Its Challenges for AI

Tacit knowledge and embodied intelligence

Michael Polanyi coined the concept of “tacit knowledge” — the knowledge we possess but cannot articulate. “We know more than we can tell,” he said. We can recognize faces without articulating the rules of recognition. We can ride a bicycle without explaining the gyroscopic dynamics. We can catch a ball in flight without solving the equations of Newtonian mechanics.

Animals are masters of tacit knowledge. Almost all animal cognition is tacit — not because animals choose not to articulate what they know, but because their knowledge is implemented in representational systems that are not linguistic. A crow that makes a tool possesses knowledge about the mechanical properties of different materials, about the geometry of the problem it faces, about the sequence of actions needed to reach the goal. But this knowledge is not encoded in propositions; it is encoded in sensorimotor representations, in perceptual maps, in patterns of neural activation that guide action without passing through the bottleneck of language.

LLMs, by construction, operate exclusively in the domain of explicit knowledge — knowledge that can be linguistically encoded. This is at once their great strength and their great limitation. They can access, process and generate colossal quantities of explicit knowledge with a speed and breadth that no biological brain can match. But they are blind to the vast ocean of tacit knowledge that constitutes most of animal and human cognition.

Affordances and the world as a field of possibilities

James Gibson, the ecological psychologist, introduced the concept of affordance — the possibilities for action that an environment offers an agent. A chair affords sitting (for a human). A thin twig affords probing (for a crow). A horizontal surface affords walking (for a quadruped). Affordances are neither objective properties of the world nor subjective projections of the agent — they are relations between agent and world that depend on the agent’s capabilities.

The perception of affordances is a crucial component of the intelligence of action — and it is something LLMs lack. When a crow examines an environment in search of tools, it perceives affordances: “this twig is rigid enough to serve as a lever”; “this leaf is flexible enough to be shaped into a probe”; “this wire can be bent into a hook.” This perception integrates visual, tactile, proprioceptive and memory information into a holistic and immediate assessment of the possibilities for action.

LLMs can reason about affordances when these are described linguistically — “a wire can be bent into a hook shape.” But they do not perceive affordances in real environments. The difference is profound: the perception of affordances is a skill that makes it possible to innovate — to find new uses for familiar objects, to see possibilities no one has described before. The crow Betty did not read anywhere that wire can be bent into hooks; she perceived the affordance directly, by interacting with the material.

Learning by interaction: the missing loop

Perhaps the most fundamental difference between animal cognition and LLM cognition is the interactive learning loop. Animals learn by doing: trying, failing, observing results, adjusting behavior. This loop — perception → action → observation of outcome → model update → new action — is the fundamental mechanism of causal and adaptive learning.

In AI, this loop is formalized by reinforcement learning (RL): an agent interacts with an environment, receives rewards or punishments, and updates its action policy to maximize expected cumulative reward. RL is powerful — it has produced agents that defeated world champions in Go, chess and StarCraft — but it remains extraordinarily difficult to apply in real-world physical environments, where the state space is continuous, rewards are sparse, and errors can be costly or irreversible.

Animals have solved this problem. Hundreds of millions of years of evolution produced brains that implement forms of biological RL that are extremely sample-efficient, robust to noise, and capable of transfer between domains. A crow can learn to use a new tool in minutes. A young chimpanzee can learn a new foraging technique by watching an adult for a few sessions. A rat can learn to navigate a maze in a few trials.

LLMs do not learn by interaction (in their default inference mode). They were trained once, on a static corpus, and then operate as fixed systems. In-context learning — the ability to adapt behavior within a session from examples provided in the prompt — is a limited form of adaptation, but it is not learning by interaction in the full sense: the model does not act in the world, does not observe results, does not permanently update its parameters.

This gap is recognized by the AI community, and there are multiple lines of research trying to close it: LLM agents with tools (such as ReAct, Toolformer and agentic architectures), multimodal models with visual and tactile inputs, integration of LLMs with RL frameworks (as in NVIDIA’s Voyager and Eureka). But the gap remains large, and the reason is fundamentally architectural: the Transformer was designed to process static sequences, not to interact dynamically with environments.

VIII. Animal Communication: Language Without Grammar, Meaning Without Syntax

The spectrum of communication

Animal communication is often dismissed as “instinctive” or “simple” — an assessment that reflects human ignorance more than the poverty of animal communication. In fact, communication in non-human species exhibits a variety and complexity that, although it does not reach the productivity and recursion of human language, constitutes a very rich spectrum of communicative capabilities.

Referential communication. The famous alarm calls of vervets (Seyfarth, Cheney & Marler, 1980) showed that these monkeys produce distinct acoustic calls for different kinds of predators — eagles, leopards and snakes — and that other members of the group respond appropriately (looking up, climbing a tree, or looking at the ground, respectively). This is referential communication: signals that refer to specific external entities or categories, not merely to internal states of the sender.

Combinatorial communication. Some species combine elementary calls into sequences with compound meanings. Campbell’s monkeys (Cercopithecus campbelli) use acoustic suffixes that modify the meaning of base calls — a primitive form of morphology. Japanese tits (Parus minor) combine notes in sequences whose order matters — a rudimentary form of syntax (Suzuki et al., 2016).

Cultural communication. The song dialects of songbirds (and of humpback whales) are learned socially and vary between populations — demonstrating cultural transmission of communicative patterns. Birdsong is often described as “language” in a metaphorical sense, but the parallels with human language run deeper than the metaphor suggests: both involve vocal learning during a sensitive period, production of structured sequences from discrete units, and cultural variation between populations.

Trained symbolic communication. The experiments with great apes and language — Washoe (Gardner & Gardner, 1969), Kanzi (Savage-Rumbaugh & Lewin, 1994), Koko — showed that chimpanzees and bonobos can learn to use hundreds of symbols (lexigrams or ASL signs) to communicate desires, describe objects, comment on events and, occasionally, combine symbols into novel constructions. Kanzi, the bonobo, demonstrated comprehension of English sentences with complex syntactic structure, including relative clauses and embedded constructions — although his linguistic production remains more limited.

What animal communication reveals about the language of LLMs

The comparison between animal communication and LLM language is instructive in both directions.

On the one hand, animal communication shows that intelligence does not require sophisticated language. Crows make complex tools with zero linguistic communication about the process. Chimpanzees negotiate political coalitions with vocalizations, gestures and facial expressions — but without grammar, without recursion, without subordinate clauses. The intelligence of action, of social life and of causal reasoning can operate without the scaffolding of language.

On the other hand, the language of LLMs shows that linguistic competence can exist without the perceptual, social and motor anchoring that accompanies language in biological organisms. For a human, saying “I see a red apple” is anchored in a rich perceptual experience. For an LLM, the same sentence is a relation between tokens in an embedding space. Both productions are linguistically competent. But the human’s is connected to a perceptual, motor and affective world; the LLM’s is connected to a statistically learned representational space.

Animal communication also reminds us that language, however powerful, is one evolutionary adaptation among many — not the crowning of cognition, but a cognitive tool that complements (and partially replaces) other forms of representation and reasoning. Animals that did not develop complex language are not “less intelligent” in some absolute sense; they developed other cognitive tools — affordance perception, implicit causal reasoning, non-verbal social cognition, spatial memory — that serve analogous functions in their ecological niches.

IX. General Intelligence: Convergence as Horizon

What “general” really means

The idea of “artificial general intelligence” (AGI) is often treated as a clear concept — but it is in fact deeply ambiguous. What does it mean to be “general”?

If “general” means “capable of solving any computable problem given enough time and resources,” then a universal Turing machine is already general intelligence — and the concept is trivial. If “general” means “capable of matching or surpassing humans at any cognitive task,” then we are defining general intelligence by reference to a specific benchmark (the human), which is anthropocentric and potentially arbitrary.

The most productive definition, in my view, is the one that derives from Chollet’s tradition: general intelligence is the capacity to efficiently acquire new skills in previously unencountered domains — the ability to generalize, adapt and transfer knowledge between contexts with minimal prior experience. This definition is scalar (it admits degrees) and dimensional (it can be assessed separately in different domains).

By this definition, no existing system — biological or artificial — is fully “general” in all domains. Humans are extraordinarily general, but not perfectly so: we are terrible at calculating large numbers in our heads, at perceiving magnetic fields, at echolocating, at processing multiple simultaneous conversations. LLMs are extraordinarily general within the linguistic domain — they can adapt to new topics, styles, languages and linguistic tasks with a few examples — but are narrow outside it. Crows are surprisingly general in the domain of physical manipulation and mechanical problem solving, but limited in complex symbolic communication.

Convergence as a project

If general intelligence is a multidimensional topography, and different existing systems cover different regions of that topography, then the project of building AGI is, essentially, a project of convergence — of integrating capabilities that today are dispersed across separate systems.

In biology, the human brain is the most successful example of convergence: it integrates language, causal reasoning, spatial cognition, social cognition, motor control, metacognition and continual learning in a single biological substrate of 1.4 kg. But this integration took hundreds of millions of years of evolution to achieve, and involved the superposition of multiple neural systems — the reptilian brainstem, the mammalian limbic system, the primate neocortex, Broca’s and Wernicke’s areas, the cerebellum — each optimized for different aspects of cognition.

In AI, we are at the beginning of this process of convergence. LLMs are an extraordinary point on the linguistic axis. RL systems are significant points on the axis of learning by interaction. Computer-vision models cover aspects of the perceptual axis. Robotic control systems address aspects of the motor axis. But the integration of these capabilities into a coherent system — one that can reason linguistically, perceive visually, act physically, learn by interaction, plan causally and adapt continually — remains an open problem of formidable difficulty.

This is, incidentally, precisely the motivation behind the architectures that have come to be called Composite AI or compositional AI: systems that combine multiple AI paradigms — foundation models, reinforcement learning, symbolic systems, world models, episodic memory — in integrated architectures. The intuition is that no single architecture will suffice for AGI; what is needed is the intelligent composition of multiple specialized competences.

X. Embodied Cognition and the Hypothesis of the Necessary Body

Varela, Thompson and enactivism

The tradition of embodied cognition, developed by Varela, Thompson and Rosch in The Embodied Mind (1991), and extended by Clark and Chalmers in the “extended mind,” argues that cognition is not a purely computational phenomenon that happens “inside” the brain — it is constitutively dependent on the body, on action and on the environment. The mind does not process internal representations of the world; it enacts — co-constitutes — the world through sensorimotor interaction.

If enactivism is correct (even partially), then general intelligence is intrinsically tied to embodiment — and no purely computational system, however powerful at symbolic manipulation, can achieve it without some form of bodily interaction with the world. This does not mean that LLMs are not intelligent; it means they are intelligent in a way that, by definition, lacks a dimension that embodied intelligence possesses.

There is a strong version and a weak version of this thesis. The strong version — that all genuine cognition requires physical embodiment — is possibly excessive. Mathematicians operating in abstract spaces, logicians proving theorems, programmers writing code — all these are examples of deep cognition that operates predominantly in the symbolic domain, with minimal dependence on real-time sensorimotor interaction. The weak version — that general intelligence, as we understand it in biological organisms, necessarily includes an embodied dimension that cannot be captured by pure symbolic processing — is, in my view, defensible and empirically well supported.

Simulation vs. experience

There is an intermediate approach that promises to partially bridge the gap between linguistic intelligence and embodied intelligence: simulation. If an LLM can be connected to a physically realistic simulated environment — with Newtonian physics (or better), visual rendering and action interfaces — then it can, in principle, develop representations of the physical world that approach embodied understanding.

This is the logic behind projects such as Habitat (Meta) and SAPIEN (Stanford), and the efforts in foundation models for robotics by groups such as RT-2 (Google DeepMind), which train multimodal models to translate linguistic instructions into robotic actions. The results are promising, but still far short of animal fluency. A robot trained with RT-2 can follow instructions such as “pick up the plastic bottle and put it in the trash can” in a controlled office environment — but fails in unfamiliar environments, with objects never seen, under variable lighting conditions. A crow would solve these challenges without hesitation.

The philosophical question is whether simulation can, in principle, replace real embodied experience. Enactivists would say no — that simulation always lacks something that real interaction with the world possesses. Functionalists would say yes — that if the simulation is faithful enough, it produces the same computational states as real experience. This is an open question that will probably not be resolved by philosophical arguments, but by empirical experimentation: we will build the systems, see what they can do, and judge from the results.

XI. The Intrinsic Value of the Linguistic Achievement

Against deflationism

There is a tendency in critical discourse about AI to deflate the achievements of LLMs — to treat them as impressive but essentially empty tricks, as automation without understanding, as imitation without substance. This tendency is understandable as a reaction to the exaggerated hype surrounding AI, but it is intellectually unfair.

Consider what has been achieved. For the first time in history, we have artificial systems that can:

Take part in natural-language conversations on virtually any topic, with coherence, relevance and depth that are often surprising.
Translate between hundreds of languages with quality that, in many language pairs, rivals professional human translators.
Generate working code in dozens of programming languages from natural-language specifications.
Summarize complex documents while preserving the essential points.
Reason about abstract multi-step problems.
Explain technical concepts at different levels of complexity for different audiences.
Produce creative text — poetry, fiction, essays — that, although it does not replace the best human writers, demonstrates sophisticated manipulation of style, structure and content.

Each of these capabilities, taken individually, would be a monumental achievement of engineering and cognitive science. Taken together, they constitute a qualitative leap in the capacity of human civilization to process, generate and communicate knowledge.

Deflating these achievements does not demonstrate intellectual sophistication; it demonstrates a lack of historical perspective. For centuries, language was considered the quintessence of human intelligence — the Rubicon that separated the human from the animal, the rational from the instinctive. “In the beginning was the Word,” says the Gospel of John. Descartes argued that linguistic ability was the definitive proof of the rational mind. The twentieth-century analytic tradition, from Frege to Wittgenstein to Davidson, made language the central object of philosophy precisely because it saw it as the mirror of cognition.

And now, when an artificial system demonstrates extraordinary linguistic competence, the response is: “oh, but it’s only language”?

The inconsistency is revealing. The same tradition that elevated language to the supreme criterion of intelligence now demotes it to a cheap trick when a machine masters it. This suggests that the true criterion was never language itself, but the substrate — that “real” intelligence requires biology, carbon, neurons. This is substratism — a form of biological chauvinism no more justifiable than the species chauvinism against which comparative cognitive biology has been fighting for decades.

What should not be claimed

That said, it is equally important not to inflate the achievements of LLMs. Linguistic competence is not general intelligence. The ability to describe the world is not the ability to act in it. The ability to reason about causal relations in text is not the ability to intervene causally in the world. The ability to simulate theory of mind in narrated scenarios is not the ability to perceive and model mental states in real time.

LLMs have achieved something real and significant. But what they have achieved is a form of specialized intelligence — brilliant in its region of the cognitive topography, but covering only a fraction of the total space. To claim that this fraction is “the” intelligence — or that it is an inevitable path toward general intelligence — is an inference that the data do not support.

The history of AI is full of moments when an impressive achievement in a specific domain was prematurely generalized. Deep Blue defeated Kasparov — and many thought general intelligence was near. Watson won at Jeopardy! — and many thought the same. AlphaGo defeated Lee Sedol — again. In each case, the achievement was real, but the generalization was unjustified. LLMs may be the fourth chapter of this story — or they may be something genuinely different, given that language is, arguably, the most “general” domain ever mastered by an artificial system. But even so, epistemological caution is wiser than premature enthusiasm.

XII. Paths to Convergence

Composite architectures and the future of AI

If general intelligence requires the integration of multiple cognitive capabilities, then the path to AGI does not run through scaling LLMs — through adding more parameters, more data, more computation to a fundamentally linguistic architecture. It runs through composite architectures that integrate different AI paradigms.

Some promising approaches include:

Multimodal models with action. Systems that combine language processing, vision, audio and, crucially, the ability to act in environments — whether real robotic ones or simulated. The work on RT-2 (Brohan et al., 2023) and PaLM-E (Driess et al., 2023) represents first steps in this direction: models that directly translate multimodal perception into robotic actions, with language serving as the instruction interface.

World models. Systems that maintain internal representations of the state of the world — including its physical, causal and dynamic properties — and that can mentally simulate the consequences of actions before executing them. Yann LeCun has advocated this approach under the name “Joint Embedding Predictive Architecture” (JEPA), arguing that world models are the missing piece between LLMs and general intelligence.

Neuro-symbolic integration. Systems that combine the generalization power of neural networks with the logical precision and compositionality of symbolic systems. The intuition is that formal reasoning, planning and verification may be better implemented by symbolic systems, while perception, natural language and learning may be better implemented by neural networks — and that the integration of the two may be more powerful than either alone.

Multi-agent systems. Architectures in which multiple specialized agents — each competent in a region of the cognitive topography — collaborate to solve problems none of them could solve alone. This approach echoes the modular organization of the human brain and can, in principle, capture the diversity of capabilities needed for general intelligence without requiring a single architecture to be good at everything.

Reinforcement learning with foundation models. The integration of LLMs with RL frameworks — where the LLM provides knowledge priors and RL provides adaptation through interaction — is an approach that can combine the vast declarative knowledge of LLMs with RL’s capacity for learning through action.

Each of these approaches faces formidable challenges. But the conceptual point is clear: if general intelligence is the integration of multiple capabilities, then the path to it is, necessarily, the integration of multiple approaches. There is no shortcut.

XIII. Final Reflections: Persistent Astonishment

Epistemic humility and calibrated admiration

Let us return to the tension with which we began. On one side, the enthusiasm of those who see in LLMs the harbinger of general intelligence. On the other, the skepticism of those who see in them nothing more than sophisticated automata. I have argued that both positions are inadequate — that the intellectually honest position is more nuanced and, at the same time, more interesting.

LLMs have achieved something that, for millennia, was considered the hallmark of human intelligence. They master language — not perfectly, not in all its aspects, but with a competence that surpasses that of any previous artificial system and that, on many specific tasks, rivals human competence. This is an achievement of intelligence. To call it “mere statistics” is as reductive as calling human cognition “mere electrochemistry.”

At the same time, general intelligence — the integrated capacity to perceive, act, learn, plan, communicate and adapt — is vastly broader than linguistic competence. Crows remind us that intelligence can mean bending a wire into a hook without ever having seen a hook. Chimpanzees remind us that intelligence can mean orchestrating a political revolution without uttering a word. Octopuses remind us that intelligence can exist in architectures radically different from our own.

General intelligence is a vast continent, with multiple regions, climates and ecosystems. LLMs have colonized one region — broad, rich and important — but have left most of the continent unexplored. Animals have colonized other regions, with completely different strategies. Humans are the only ones who, so far, have managed to inhabit the continent in all its extent — and even so, with gaps and limitations.

Astonishment as an epistemological stance

I close with a suggestion that is both methodological and existential. Faced with intelligence — in all its forms, biological and artificial — the most productive stance is calibrated astonishment: the capacity to marvel at what has been achieved without losing sight of what remains to be achieved.

Astonishment before an LLM that can discuss Kantian philosophy in six languages. Astonishment before a crow that invents a tool never seen before. Astonishment before a chimpanzee that negotiates a political alliance with a glance and a gesture. Astonishment before an octopus that solves a maze with semi-autonomous arms. Astonishment before a human brain that, somehow, does all of this at once.

And sobriety before the fact that we do not fully understand any of these forms of intelligence. We do not know, at the mechanistic level, how LLMs generate the representations they generate. We do not know how a crow’s brain plans three steps ahead. We do not know how consciousness emerges from neurons — or whether it emerges from Transformers. We do not know whether the integration of multiple forms of artificial intelligence will produce something analogous to human general intelligence, or something radically different.

What we do know is that intelligence is a phenomenon vaster, more varied and more mysterious than any single paradigm — linguistic, computational, biological or philosophical — can capture. And that each advance — whether a new model with trillions of parameters or a field study of crows in New Caledonia — illuminates one more facet of this multidimensional diamond.

The achievement of language models is real. The complexity of general intelligence is also real. Holding both truths simultaneously — without letting one diminish the other — is, perhaps, the most demanding cognitive challenge that the contemporary debate on artificial intelligence imposes on us. It is a challenge that requires, ironically, precisely that form of intelligence which no single system — human, animal or artificial — fully commands: the capacity to integrate radically different perspectives into a coherent understanding that honors the complexity of the real.

XIV. The Illusion of Scale and the Limits of Scaling

The scaling hypothesis and its promises

A dominant narrative in contemporary AI is the scaling hypothesis — the idea that, by increasing the number of parameters, the volume of training data and the amount of computation, LLMs will continue to acquire emergent capabilities that will progressively fill the gaps identified in this essay. This hypothesis, defended enthusiastically by researchers such as Ilya Sutskever, Dario Amodei (with important qualifications) and, more radically, by figures such as Sam Altman, is supported by real empirical observations: the scaling laws of Kaplan et al. (2020) and Hoffmann et al. (2022) demonstrated predictable relationships between computational scale and performance on linguistic tasks.

But the question is whether these scaling laws extend beyond the linguistic domain — whether more scale can endogenously produce capabilities such as interventional causal reasoning, learning by interaction, or adaptive motor control.

There are serious theoretical reasons for doubt. Wolpert and Macready’s (1997) No Free Lunch Theorem reminds us that no learning algorithm is universally superior: every advantage on one class of problems is offset by a disadvantage on another. The Transformer architecture, optimized for processing discrete sequences, may be structurally inadequate for representing and manipulating continuous, dynamic and spatial information — regardless of how much scale is applied.

Consider an analogy: if I take the world’s best supersonic aircraft and increase its power by a factor of a thousand, it does not become a submarine. More power in an architecture designed to fly produces faster flight — not the ability to submerge. By analogy, more scale in an architecture designed to process text produces more sophisticated text processing — not necessarily the ability to act in the physical world.

This does not mean that scaling is irrelevant. The emergent capabilities observed in LLMs — in-context learning, chain-of-thought reasoning, solving new problems — are real and surprising. But the question is whether these emergences converge toward general intelligence or reach asymptotes along each dimension of the cognitive topography that is not directly captured by the training objective (next-token prediction on text).

The ARC test and Chollet’s measure

The work of François Chollet deserves special attention in this context. His ARC benchmark (Abstraction and Reasoning Corpus) was explicitly designed to test general intelligence — not memorization, not linguistic processing, not matching of familiar patterns, but the ability to reason about abstract patterns never before encountered.

The results are instructive. Until recently, LLMs — even the most advanced — showed modest performance on ARC, significantly below average human performance. This is particularly revealing because ARC tasks are visual, not linguistic: patterns on colored grids that must be transformed according to abstract rules inferred from a few examples. The difficulty is not in linguistic processing (the grids can be described in text), but in the kind of reasoning required: abstraction of patterns, generalization from minimal examples, composition of transformations.

The most successful approaches on ARC combine LLMs with program search — generating transformation hypotheses as code and testing them against the examples. This is itself revealing: the LLM contributes prior knowledge and candidate generation, but the actual reasoning is delegated to a search-and-verification mechanism. The composition of paradigms outperforms the pure LLM — exactly what the thesis of multidimensional convergence would predict.

XV. Time, Memory and Continuity: The Neglected Dimension

Intelligence as a temporal process

There is a dimension of intelligence that has been systematically neglected in the discussion of LLMs: the temporal dimension. Biological intelligence is not a state — it is a process that unfolds in time. Animals learn, forget, relearn, consolidate, revise, integrate. Their intelligence is dynamic: today’s crow is more competent than yesterday’s crow, because yesterday it tried a new extraction technique and consolidated it during sleep. The young chimpanzee watches an adult cracking nuts for months before attempting it itself — and each observation modifies its internal representations, even before the first motor attempt.

LLMs, by contrast, exist in an eternal present. Each interaction session starts without memory of previous sessions (except for external memory mechanisms). Each inference is computationally independent of the previous ones. The model does not mature, does not develop, does not integrate experiences over time.

This is a profound limitation, because much of what we call intelligence in biological organisms is developmental — it emerges along developmental trajectories that extend over weeks, months, years. The intelligence of an adult crow is not innate; it is the product of a maturation process that includes juvenile exploratory play, observation of adults, trial and error, and consolidation during sleep. The intelligence of a chimpanzee is shaped by years of socialization, cultural learning, alliance formation and experiences of conflict.

Even in the linguistic domain, human capacity is deeply developmental. A child is not born with the linguistic competence of an adult; it builds it over years of interaction with speakers, of experimentation with grammatical structures, of error correction, of gradual expansion of vocabulary and syntactic complexity. The end result — adult linguistic competence — reflects this entire developmental history. The result of an LLM — its post-training competence — reflects an optimization process over static data, without interaction, without development, without trajectory.

Episodic memory and the narrative self

Endel Tulving distinguished between semantic memory (general knowledge about the world) and episodic memory (recollections of specific events experienced by the subject). LLMs possess something analogous to semantic memory — vast general knowledge about the world, encoded in the network’s parameters. But they do not possess episodic memory: they do not “remember” specific interactions, do not experience events, do not construct a narrative of themselves over time.

Episodic memory is crucial for several forms of intelligence. It makes it possible to learn from specific events (not only from statistical regularities). It enables planning based on analogy with past experiences. It enables the construction of a model of the self as an agent with a history — which some philosophers consider a precondition for genuine rationality.

Animals such as corvids, chimpanzees and dolphins demonstrate sophisticated forms of episodic memory. Scrub jays (Aphelocoma californica) remember not only where they hid food, but when — and prioritize retrieving perishable items over non-perishable ones based on elapsed time (Clayton & Dickinson, 1998). This is episodic memory: recollection of a specific event (the act of caching), indexed by time and place, and used to guide future behavior adaptively.

The absence of episodic memory in LLMs does not make them “not intelligent,” but it deprives them of a dimension of intelligence that is central to biological cognition. An LLM can discourse on episodic memory with academic precision; but it does not possess one.

XVI. Evolution, Selective Pressure and the Design of Intelligence

Intelligence as an evolutionary product

All biological intelligence is the product of evolution — of hundreds of millions of years of variation, selection and inheritance. This has profound implications for how we understand the relationship between language, action and intelligence.

The intelligence of action — the ability to perceive, navigate, manipulate and act in physical environments — is evolutionarily much older than linguistic intelligence. The first nervous systems arose more than 500 million years ago, under selective pressure to control movement in aquatic environments. The first true brains evolved in fish, some 400 million years ago, to coordinate complex motor behaviors. Causal cognition, tool use and sequential planning emerged in birds and mammals over the last 100 million years.

Articulated language, by contrast, is extraordinarily recent: probably no more than 100,000 to 300,000 years — a minuscule fraction of the evolutionary history of intelligence. This means that language is, literally, the last layer added to a cognitive edifice that was already extraordinarily complex and multifaceted.

This evolutionary perspective illuminates the relationship between LLMs and animal intelligence in a revealing way. LLMs have replicated the most recent layer of cognition — language — without possessing any of the ancestral layers on which language was evolutionarily built: motor control, spatial perception, causal cognition, social cognition. It is as if we had built the roof of a building without the foundations, the pillars and the walls. The roof is impressive — elegant, functional, well finished — but it floats without structural support.

This does not invalidate the roof. But it explains why it is not, on its own, a complete building.

The evolutionary argument against pure scaling

There is a subtle evolutionary argument that deserves attention. If general intelligence could emerge solely from language processing at sufficient scale, it would be surprising that biological evolution invested hundreds of millions of years building perceptual, motor, emotional and social systems before adding language. Evolution is a relentless optimizer: if language alone were sufficient for general intelligence, why the massive investment in all the other capabilities?

The obvious answer is that language is not sufficient — that it is an extraordinarily powerful amplifier of a cognitive base that was already rich and multifaceted. Language potentiates cognition by enabling abstraction, cultural transmission and hypothetical reasoning. But it potentiates a base that includes perception, action, emotion, sociality and temporality — and without that base, it is an amplification tool with no signal to amplify.

Or rather: it amplifies the only signal available, which is the corpus of text on which it was trained — the linguistic distillation of humanity’s collective experience. LLMs are, in this sense, benign epistemic parasites: they feed on knowledge encoded linguistically by billions of human brains that possess the capacities for action, perception and experience that LLMs themselves lack. The model’s linguistic competence is extraordinary, but it is derived — not original. It reflects humanity’s embodied intelligence, filtered and distilled through the medium of language.

XVII. A Map for the Future: What Is Missing and How to Get There

The five critical gaps

If we want to move from today’s LLMs toward something that genuinely deserves the name of artificial general intelligence, five gaps need to be addressed:

1. Perceptual grounding. Connecting linguistic representations to multimodal perception of the real world — vision, hearing, touch, proprioception. Multimodal models (GPT-4o, Gemini, Claude with vision) are first steps, but they still process images as static inputs, not as the continuous perceptual stream of an agent situated in the world.

2. Agency and interaction. The capacity to act in the world, observe the consequences of actions, and learn from that interaction. This requires not only robotic (or simulated) bodies, but a deep integration of perception, planning and action that preserves the interventional causal-learning loop that animals perform so naturally.

3. Continual learning. The ability to integrate new experiences without losing prior knowledge — the opposite of catastrophic forgetting. This probably requires architectures with explicit memory, consolidation during periods of low activity (analogous to sleep), and selection mechanisms that prioritize informative experiences.

4. Temporality and development. Systems that not only process sequences, but exist in time — that mature, develop, and build increasingly sophisticated representations along developmental trajectories that extend over significant periods.

5. Robust metacognition. The ability to monitor one’s own confidence, recognize the limits of one’s knowledge, and adjust behavior accordingly. This includes the ability to say “I don’t know” in a calibrated way, to actively seek information when uncertain, and to distinguish between what is known by experience and what is known by inference.

None of these gaps is trivial. Each represents decades of research challenge. But clarity about what is missing is itself progress: it makes it possible to direct efforts, allocate resources and evaluate results more precisely than the vague aspiration to “more scale.”

The beauty of incompleteness

There is something deeply human — and perhaps universally intelligent — in the capacity to recognize one’s own incompleteness. A crow does not know that it does not speak; an LLM does not know that it does not act. But we — humans reflecting on the topography of intelligence — can see both the mountains that each system has climbed and the valleys that remain unexplored.

This panoramic view is itself an exercise of general intelligence: the integration of perspectives from computer science, evolutionary biology, cognitive psychology, philosophy of mind, robotics and ethology into a synoptic understanding that none of these disciplines, in isolation, could produce.

And in this integration lies, perhaps, the deepest lesson: that general intelligence — whether in biological organisms, in artificial systems, or in civilization as a whole — is, fundamentally, the capacity to integrate. To integrate perception with action. Language with experience. Past with future. What is known with what is not known. The familiar with the strange.

LLMs have integrated an unprecedented quantity of human linguistic knowledge into extraordinarily powerful systems of symbolic manipulation. Crows have integrated perception, causality and action into problem-solving systems that put our best robots to shame. Chimpanzees have integrated social cognition, memory and strategy into political systems that would rival many human organizations.

Artificial general intelligence, when it comes — if it comes — will be the integration of all these integrations. And the astonishment we feel before it will be, predictably, proportional to the astonishment we already feel before each of its parts — multiplied by the irresolvable mystery of a totality that exceeds the sum.

Selected References

The references below are indicative, not exhaustive, and serve as entry points for readers interested in going deeper into the themes discussed. They are organized by subject area for easier navigation.

Language Models and Interpretability

Bender, E.M., Gebru, T., McMillan-Major, A., & Shmitchell, S. (2021). On the Dangers of Stochastic Parrots. FAccT ‘21.
Brown, T., et al. (2020). Language Models are Few-Shot Learners. NeurIPS.
Elhage, N., et al. (2022). Toy Models of Superposition. Anthropic Research.
Hoffmann, J., et al. (2022). Training Compute-Optimal Large Language Models (Chinchilla). arXiv:2203.15556.
Kaplan, J., et al. (2020). Scaling Laws for Neural Language Models. arXiv:2001.08361.
Li, K., et al. (2023). Emergent World Representations: Exploring a Sequence Model Trained on a Synthetic Task. ICLR.
Nanda, N., et al. (2023). Progress measures for grokking via mechanistic interpretability. ICLR.
Olsson, C., et al. (2022). In-context Learning and Induction Heads. Anthropic Research.
Wei, J., et al. (2022). Chain-of-Thought Prompting Elicits Reasoning in Large Language Models. NeurIPS.

Animal Cognition and Comparative Biology

Call, J., & Tomasello, M. (2008). Does the chimpanzee have a theory of mind? 30 years later. Trends in Cognitive Sciences.
Clayton, N.S., & Dickinson, A. (1998). Episodic-like memory during cache recovery by scrub jays. Nature.
De Waal, F. (1982). Chimpanzee Politics: Power and Sex Among Apes. Johns Hopkins UP.
Finn, J.K., Tregenza, T., & Norman, M.D. (2009). Defensive tool use in a coconut-carrying octopus. Current Biology.
Fiorito, G., & Scotto, P. (1992). Observational Learning in Octopus vulgaris. Science.
Garcia, J., & Koelling, R.A. (1966). Relation of cue to consequence in avoidance learning. Psychonomic Science.
Hare, B., Call, J., & Tomasello, M. (2001). Do chimpanzees know what conspecifics know? Animal Behaviour.
Hunt, G.R. (1996). Manufacture and use of hook-tools by New Caledonian crows. Nature.
Inoue, S., & Matsuzawa, T. (2007). Working memory of numerals in chimpanzees. Current Biology.
Savage-Rumbaugh, E.S., & Lewin, R. (1994). Kanzi: The Ape at the Brink of the Human Mind. Wiley.
Seyfarth, R.M., Cheney, D.L., & Marler, P. (1980). Vervet monkey alarm calls: Semantic communication in a free-ranging primate. Animal Behaviour.
Smith, J.D., Shields, W.E., & Washburn, D.A. (2003). The comparative psychology of uncertainty monitoring and metacognition. Behavioral and Brain Sciences.
Suzuki, T.N., Wheatcroft, D., & Griesser, M. (2016). Experimental evidence for compositional syntax in bird calls. Nature Communications.
Taylor, A.H., et al. (2007). Spontaneous metatool use by New Caledonian crows. Current Biology.
Weir, A.A.S., Chappell, J., & Kacelnik, A. (2002). Shaping of hooks in New Caledonian crows. Science.
Whiten, A., et al. (1999). Cultures in chimpanzees. Nature.

Philosophy of Mind and Embodied Cognition

Chollet, F. (2019). On the Measure of Intelligence. arXiv:1911.01547.
Clark, A., & Chalmers, D. (1998). The Extended Mind. Analysis.
Gibson, J.J. (1979). The Ecological Approach to Visual Perception. Houghton Mifflin.
Harnad, S. (1990). The Symbol Grounding Problem. Physica D.
Pearl, J. (2009). Causality: Models, Reasoning, and Inference. 2nd ed. Cambridge UP.
Polanyi, M. (1966). The Tacit Dimension. University of Chicago Press.
Searle, J. (1980). Minds, Brains, and Programs. Behavioral and Brain Sciences.
Tulving, E. (1985). Memory and consciousness. Canadian Psychology.
Varela, F., Thompson, E., & Rosch, E. (1991). The Embodied Mind: Cognitive Science and Human Experience. MIT Press.
Vygotsky, L. (1934/1986). Thought and Language. MIT Press.

Robotics and Action Models

Brohan, A., et al. (2023). RT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic Control. arXiv:2307.15818.
Driess, D., et al. (2023). PaLM-E: An Embodied Multimodal Language Model. ICML.
Wolpert, D.H., & Macready, W.G. (1997). No free lunch theorems for optimization. IEEE Transactions on Evolutionary Computation.

Final Note

This essay was written with the conviction that intellectual clarity about what language models are and are not is a prerequisite for progress — both in the technical development of AI and in the public debate about its impact. The temptation to deflate genuine achievements is as epistemologically dangerous as the temptation to inflate them: both distort the map and disorient us in the territory.

The position defended here — that the linguistic competence of LLMs is real and partial intelligence, that general intelligence is multidimensional and no system possesses it completely, that animals demonstrate forms of intelligence that LLMs lack and that these forms are as legitimate as linguistic competence — is not a comfortable position. It allows neither the triumphalism of the AI industry nor the cynicism of its harshest critics. It demands that we hold multiple truths in productive tension, without prematurely resolving them into a simple narrative.

But perhaps this is itself the highest form of intelligence available to us at this historical moment: the capacity to inhabit complexity without simplifying it, to admire without idolizing, to criticize without diminishing. It is a capacity that is in the parameters of no model and, at the same time, is the best that the human intellectual tradition has to offer.

— comments