AiBrain 
An increasingly visible line of research argues that LLMs have reached their limits, and that the real path to AGI lies in world models: systems trained to predict the dynamics of the physical world rather than text. LeCun has made this his crusade for years, with a complete architecture comprising perception, a world model, a cost module, short-term memory, an actor, and a configurator, rather than a single isolated model. The intuition is compelling: a system that “understands” physics, object permanence, and mechanical causality seems closer to what we intuitively call intelligence than a statistical parrot that predicts tokens.
I think this intuition confuses two different things. And that confusion is costly if it guides research.
AGI is not about prediction, but general adaptation
A world model may be better than an LLM at predicting a local dynamic: the trajectory of a falling object, the behavior of a robotic arm, or the evolution of a video scene. On that specific axis, physical grounding wins, and that is normal. It is the terrain for which it was built.
But AGI is not about “predicting a dynamic better”. It is about adapting behavior to unseen problems by relying on structures learned elsewhere. On that axis, cross-domain transfer, composition, and the integration of new rules in context, systems built around LLMs currently have the strongest available profile. Not because the model is intelligent in any strong sense on its own. Because, once equipped with tools, it can translate goals into plans, call tools, converse with humans, manipulate abstractions, and connect heterogeneous domains. It offers an interface and composition profile that no current world model has demonstrated at comparable breadth.
The question is therefore not “which one is more intelligent?” It is: on which axis does the generality we are looking for actually lie?
Text is not a poor subsample of the world
The classic objection to LLMs is that the text corpus is too small and too impoverished compared with the richness of the real world. The objection is not false, but it is incomplete.
Text is not a raw sample of the world. It is a refined human artifact, selected and compressed by millennia of collective cognition. It is poor in perception, spatiality, fine-grained physics, and temporal continuity. Agreed. But it is extraordinarily dense along other axes: abstraction, narrated causality, categories, procedures, social norms, law, code, strategies, documented errors, reasoning, debates, and explanatory models. Text is not the raw world. It is the world already digested by humans, and much of the generalization work has already been done upstream by us, before the model sees its first sentence.
This argument still needs a qualification to hold. That compression is not neutral. It is structurally biased toward what can be verbalized. By construction, it excludes tacit knowledge, gestures, and the expert practitioner’s intuition. As Polanyi put it, “we know more than we can tell”. It also mostly encodes validated results, not the thousands of failed attempts or correct intuitions that were never formalized: a massive survivorship bias. Text is therefore not an impoverished subsample of the world. It is an oversample of the fraction of the world that passes through language.
And that is precisely what makes the argument robust rather than fragile. It explains why LLMs are strong exactly on layers of abstraction and weak exactly on the tacit and sensorimotor layers. It is the same fault line that Moravec’s paradox described forty years ago: what is easy for humans is hard for machines, and vice versa. Text is not merely a window onto LLM capabilities. It is an especially clear window onto their limits because it draws the boundary of the tacit by construction, not by accident.
Why world models will be worse positioned for transfer
This is the core of the argument, and it needs to be stated carefully because it is not merely a question of confinement. It is a question of the nature of validity.
There is a useful distinction here, borrowed from Pearl’s causal hierarchy. A model can be valid in two different ways. Observational validity means that it has seen correlations, trajectories, or descriptions without ever intervening on the system it models. Interventional validity means that it knows not only what usually follows a situation, but what will happen if the agent acts. The second is much stronger, and much rarer.
Not all world models achieve it. Many still learn passively from video or observed trajectories. They remain largely observational, like LLMs, just on another modality. Even a model that predicts a consequence conditional on an action does not automatically learn a causal effect. In recorded trajectories, the action often depends on the state and on the policy that selected it. Learning P(Y|A) is not enough to learn P(Y|do(A)). Real interventional validity requires action-conditioned data, controlled experiments, or sufficiently faithful simulators.
The engineering consequence is concrete. A purely observational model fails as soon as the agent departs from the policy that generated its training data. It has never seen the states its own errors lead it into, so it has no idea how to escape them. This is the classic distribution shift, and it is exactly what training under intervention corrects: by acting, the model visits the states its policy produces, including the bad ones.
This is where the structural trade-off appears, and it is subtler than “world models are inherently confined”. It is relatively easy to accumulate broad observations of the world. It is much harder to perform varied, controlled, and safe interventions on “the world in general”. The stronger the interventional validity we seek, the narrower the scope in which it can be acquired at reasonable cost tends to become. We do not necessarily trade all generality for all causality. But we often trade easily scalable observational breadth for validity under action, which is much more expensive to generalize. Borrowed generality on one side, tested causality on the other.
A recent example illustrates the same trade-off, even though it is not strictly a world model. It is instead a specialized vision model, a visual foundation model without dynamics or predictions of change. NeuroVFM, pretrained on 5.24 million volumes from 566,915 examinations within a single large hospital system, outperformed a GPT-5-based pipeline by 21.4 balanced-accuracy points in critical-case detection and triage, in a one-week prospective silent study (Kondepudi et al., Nature Medicine, 2026). This is not a pure model-versus-model duel: NeuroVFM processes 3D volumes natively, while GPT-5 receives 2D slices constrained by API limits. The authors note, however, that increasing the number of series provided to GPT-5 changes little. What the example demonstrates is precise: a specialized perception module integrated into a pipeline suited to its modality can substantially beat a generalist used directly on that modality. It develops real breadth within neuroimaging, and is explicitly a generalist model for its domain, but it does not demonstrate cross-domain breadth comparable to an LLM. Maximum density in a restricted space, rather than breadth across heterogeneous spaces.
A specialized system, whether an interventional world model or a dense representation model such as NeuroVFM, can therefore develop real generalization, but mainly within its domain. This is not an accidental weakness to fix. It is the direct consequence of how it gained its reliability, through intervention or density. Both will therefore probably be better at local transfer, close to their training distribution, without yet having demonstrated robust transfer across strongly heterogeneous domains. That is at least one of the central axes of the generality we seek.
The argument must also be turned back onto the LLM itself, or it becomes too convenient. Saying “the LLM is broad because it is observational” is true of its pretraining, not of the complete system. An LLM agent that executes code, runs an experiment, calls a simulator, or acts in an environment can already produce and exploit interventional observations. Locally, inside the loop, it has a form of validity that was not present in its weights.
But we must not confuse two things: observing the result of an intervention is not the same as learning durably from that intervention. Without memorization, verification, and reuse, the system merely exploits a result it will forget on the next turn. Only when those experiments are captured does temporary validity become operational knowledge. The question becomes less “how do we replace the LLM with a world model?” and more “how do we accumulate, verify, and reuse these experiences without losing the initial breadth of transfer?” That moves the problem toward architecture, not toward choosing a single model.
Those digging into this path, not merely theorizing about it
This position is not isolated, and it is not merely philosophical. Teams are building on it.
François Chollet (Ndea) is the clearest case. He created ARC-AGI precisely to measure this axis: not performance on a known task, but the ability to generalize from a few examples to novel problems. His thesis for Ndea, the startup he co-founded with Mike Knoop, formerly of Zapier, is that neither raw LLM scaling nor physical grounding alone solves compositional generalization. What is needed is deep-learning-guided program synthesis: combining intuitive pattern recognition with explicit formal reasoning. Rephrased in this article’s vocabulary, that means building a bridge between the two regimes: using deep learning’s pattern-recognition strength to generate, for each new task, a small explicit and verifiable symbolic model. Not a universal world model learned once and for all, but a disposable model built on demand and tested for validity.
This has not remained theoretical. Launched in March 2026, ARC-AGI-3 pushes this definition toward fully interactive environments with no rules or explicit objective provided. The agent must explore, discover what “winning” means, and then transfer what it learns from one level to another. Every selected environment could be solved by humans, while the tested frontier systems remained below 1%. That is a harsh reminder of the gap between executing an instruction and discovering through interaction what must be accomplished. A clear improvement on this benchmark by an architecture that does not rely on LLM scaling alone would be a serious signal in favor of this thesis.
Two other voices mark out the same territory, more briefly. Josh Tenenbaum (MIT) represents the older academic lineage: learning concepts through probabilistic program induction, the idea that a child generalizes from three or four examples because they manipulate explicit compositional structures, not because they have seen a massive corpus. And Gary Marcus, more polemicist than builder, summarizes the position in one sentence: there is no AGI without a hybrid architecture and explicit symbolic manipulation.
The common point among all three is that none of them says LLMs are useless. They all identify explicit symbolic composition as the missing axis: not corpus size or raw sensorimotor richness, but the ability to manipulate and recombine formal structures in a verifiable way.
The thesis this leads to
The debate is not LLM versus world model. The debate is: which architecture can compose several regimes of representation, learning, and validation, textual, physical, symbolic, and interactive, into a system capable of adapting to new problems?
LLMs are not AGI. World models are not AGI either. The former inherit a horizontal generality: that of language, the medium in which humanity has compressed its abstractions, procedures, institutions, and explanatory models. Action-conditioned world models validated through intervention may provide a more vertical generality: a more robust but localized understanding of the dynamics specific to an environment.
Borrowed generality against tested causality. Easily extensible observational breadth against validity under action, which is expensive to generalize. These are not two versions of the same object. They are two regimes of validity that cannot be obtained with the same data or at the same cost.
The world model is a building block, useful wherever tested dynamics matter more than described dynamics. The tool-equipped LLM is currently the best available composition layer because its pretraining inherits the structure humanity has already built into language, without having to rediscover it experimentally, and because, once placed in an action loop, it can locally produce the validity it lacks. But the future system will probably not be a static LLM controlling a collection of static models. It will look more like a loop: abstract hypothesis → action or simulation → observation → update of a local model → planning → verification. The LLM would be less the “brain” than the translation protocol between heterogeneous representation spaces.
That shifts the real question. It is not “LLM or world model”. It is: how can we convert broad but borrowed linguistic generality into tested causality that is expensive to extend, and the reverse, within the same system? That is the open architecture problem.