答一
Dear friend, Yann LeCun’s argument in “World Models: Enabling the Next AI Revolution” is essentially this:
The present AI revolution taught machines to manipulate language. The next one must teach them how the world works.
He thinks large language models are impressive but structurally insufficient for human-level intelligence. They predict sequences of tokens; they do not reliably possess the kind of grounded, intuitive understanding that even cats, infants, and young children acquire through observing and acting in the physical world. His alternative is an AI built around an internal world model: a learned representation that can predict what will happen, imagine the consequences of possible actions, and use those predictions to plan.
1. LLMs know language, but not necessarily the world
LeCun’s criticism is not that LLMs are useless. It is that language is an impoverished and indirect record of reality.
Text contains descriptions of falling glasses, moving cars and opening doors, but it does not contain the continuous sensorimotor information from which animals learn:
- object permanence;
- gravity and inertia;
- three-dimensional space;
- cause and effect;
- what actions are physically possible;
- what is likely to happen next.
A model may eloquently explain that an unsupported object will fall, yet still fail at elementary physical prediction in an unfamiliar visual scene. LeCun invokes Moravec’s paradox: tasks that are intellectually difficult for humans, such as symbolic calculation, may be easy for computers, while sensorimotor abilities mastered by young children remain extremely difficult for AI.
So the difference is:
An LLM predicts what word follows a sentence.
A world model predicts what state of the world follows an event or an action.
That is the movement from word models to world models.
2. A world model is an internal simulator—but not a photorealistic one
The term can be misleading. LeCun does not mean that an AI must reproduce every atom or pixel of reality.
Its world model should represent the aspects of a situation that matter for prediction and action. For example, a driver needs to know that another vehicle may enter the intersection; it does not need to predict the exact motion of every leaf beside the road.
The model performs two main functions:
- It infers information that is missing from immediate perception.
- It predicts plausible future states, including what may happen after a proposed action.
Because the world is uncertain, it must represent several possible futures rather than a single predetermined outcome. The model is thus less like a perfect film of the future than a task-sensitive field of possibilities.
This is crucial. Intelligence, for LeCun, is not merely recognizing what is present. It is being able to say:
“Given the present situation, these futures are possible; if I do this, that is likely to happen.”
3. Prediction must occur in an abstract latent space
LeCun believes conventional generative models waste enormous effort predicting details that are intrinsically unpredictable or irrelevant.
Imagine predicting the next frame of a video. The exact pattern of smoke, the precise movement of leaves, or every reflection on water may be impossible to predict. A pixel-generating system is penalized for failing to reproduce these details, even when they do not matter for understanding the event.
His proposed Joint-Embedding Predictive Architecture, or JEPA, instead predicts the abstract representation of a missing or future event.
It does not ask:
“What will every pixel look like?”
It asks something closer to:
“What kind of event or state is likely to come next?”
For example, at a high level, a model may predict that a cook will mix ingredients and fry a crêpe, without predicting the precise millisecond-by-millisecond trajectory of every finger. Prediction at higher levels of abstraction also permits longer time horizons.
This is perhaps the most original technical element of LeCun’s position:
Understanding requires learning what not to predict.
The model should preserve structure, causality and affordances while discarding unpredictable noise.
4. Intelligence requires planning, not merely producing an answer
A language model usually maps an input to a response. LeCun’s autonomous architecture is different: it considers possible action sequences, runs them through the world model, estimates their consequences and chooses an action that best satisfies an objective.
In simplified terms:
- Perception constructs a representation of the current situation.
- The world model predicts possible future states.
- An actor proposes possible actions.
- A cost or objective module evaluates their expected consequences.
- The system selects the action expected to minimize cost or accomplish the goal.
This resembles model-predictive control: repeatedly imagine several futures, act on the best current plan, observe what actually happened, and revise.
Thus, LeCun’s “reasoning” is not principally an internal verbal monologue. It is:
simulation under an objective.
A system reasons by manipulating an internal model of possible states and actions.
5. AI must learn as animals do: mainly by observation
Humans do not learn basic physics from millions of labelled examples. Infants watch, move, touch, drop things and gradually form expectations about the world.
LeCun therefore places enormous importance on self-supervised learning. The system learns from raw video and sensory streams by predicting missing or future representations, without requiring humans to label every object or event.
This reverses the classical supervised-learning formula. Instead of humans telling the machine the answer repeatedly, the structure of the world supplies the training signal.
Hence one of his recurring slogans:
“The revolution will not be supervised.”
His point is that labelled data are far too sparse and expensive to produce the breadth of commonsense knowledge required by autonomous intelligence.
Interaction remains necessary, especially for learning the consequences of the system’s own actions, but passive observation supplies much of the initial background knowledge.
6. Why video matters
Text compresses only a tiny fraction of human experience. Video contains vast information about space, time, motion, objects and interactions.
LeCun argues that a child acquires more information about the physical world through a relatively brief period of vision than even an enormous language model obtains from all its textual training. This is not simply because video files are numerically larger; it is because sensory data expose regularities that language often leaves unstated.
No one must repeatedly write:
- solid objects do not normally pass through one another;
- unsupported objects fall;
- occluded objects continue to exist;
- pushing an object may cause it to move.
Children absorb these regularities by observation.
Meta’s V-JEPA research was intended as an early implementation of this idea. V-JEPA predicts masked or future content in an abstract representation space rather than reconstructing every video pixel. Its successor, V-JEPA 2, extended this toward prediction, planning and robot control, including tasks involving unfamiliar objects and environments.
7. The destination is embodied AI
The practical target is not merely a better chatbot. It is AI capable of operating in open, changing environments:
- household robots;
- industrial robots;
- autonomous vehicles;
- wearable assistants;
- systems that learn new physical tasks;
- agents that can plan without being retrained for every situation.
Current robots often perform well only in highly structured settings. Slightly changing an object, room or task can cause failure. A world-model-based system should be able to understand the new configuration, predict the effects of possible actions and adapt.
This is why LeCun regards world models as potentially revolutionary. They could move AI from:
producing representations of the world
to
competently inhabiting and acting within the world.
8. Memory and hierarchy are indispensable
LeCun’s envisioned system is modular rather than a single giant feed-forward model.
It requires short-term memory to retain recent states and actions. It may also require longer-term memory for reusable knowledge. Its representations must be hierarchical because the world unfolds at multiple temporal and spatial scales.
Consider making tea:
- at a high level: make tea;
- beneath that: boil water, prepare the cup, steep the leaves;
- beneath that: reach for the kettle;
- beneath that: coordinate fingers, wrist and arm.
The lower the level, the more precise and short-term the prediction. The higher the level, the more abstract and long-term it becomes.
A hierarchical JEPA would ideally predict across these levels. This is how a system might connect low-level motor control with long-term purpose.
9. Objectives replace unconstrained generation
LeCun’s architecture is often called objective-driven AI.
Rather than merely generating outputs that statistically resemble training data, the agent acts to satisfy objectives while respecting constraints. Some objectives may be supplied by the task, while others may be built in—for example, avoiding damage, conserving energy or obeying safety restrictions.
The world model predicts consequences; the objective function distinguishes preferable from undesirable consequences.
This is also central to LeCun’s view of AI safety. He tends to reject the idea that intelligence itself necessarily produces dangerous goals. Instead, he argues that safe behaviour should be engineered through objectives and guardrails that constrain planning. Whether this will be adequate remains highly contested, but it is structurally important to his proposal.
10. Why he calls this the “next revolution”
LeCun sees several historical waves:
- supervised deep learning allowed machines to recognize patterns;
- generative AI and LLMs allowed machines to manipulate language and other symbolic sequences;
- world models may allow machines to understand dynamics, anticipate consequences and plan actions.
The next revolution is therefore not simply another scaling step. It is an architectural transition:
from correlation to consequence;
from recognition to prediction;
from answering to planning;
from language to embodied interaction.
In his account, scaling existing autoregressive LLMs will continue to improve them, but will not automatically produce the grounded common sense, persistent memory and reliable planning needed for advanced autonomous intelligence. World models are meant to supply precisely these missing capacities.
What has actually been demonstrated?
LeCun’s programme is not merely philosophical. I-JEPA demonstrated latent-space representation learning from still images, and V-JEPA extended it to video. V-JEPA 2 reported stronger physical prediction and limited zero-shot robotic planning after video pretraining plus a comparatively small amount of robot-action data. More recent work has explored which JEPA design choices improve navigation and manipulation and how latent representations can be made more suitable for planning.
But these are still early systems. They do not yet possess a child’s broad physical common sense, open-ended autonomy or robust long-horizon planning.
So LeCun has presented:
- a serious research programme;
- some promising components;
- suggestive robotic demonstrations;
but not yet a completed alternative to the contemporary LLM paradigm.
The central vulnerability in his argument
World models face a severe difficulty: the world is not simply observed; it is selected according to relevance.
To predict well, a system must decide:
- which objects matter;
- which possible futures deserve attention;
- at what level of abstraction to represent events;
- what counts as success, danger or error;
- when an unexpected event requires revising the model itself.
LeCun assigns much of this work to learned representations, objectives and configurable modules. But that means the system requires not merely a world model; it requires a way of determining which world is presently at stake.
A cup can be:
- an obstacle;
- a tool;
- a fragile possession;
- evidence at a crime scene;
- a gift from someone deceased.
Its physical properties may be identical, but its significance changes the appropriate action.
Thus, a physical world model may supply causality and affordances without yet supplying a lived world.
My reading
I think LeCun is largely correct that contemporary AI possesses vastly more linguistic competence than sensorimotor common sense. He is also right that an intelligent agent cannot survive through next-token prediction alone: it must anticipate consequences, distinguish relevant from irrelevant detail and revise actions in response to reality.
But his phrase “world model” contains two different meanings:
- a model of physical regularities—objects, motion, causality, possible actions;
- a world in the phenomenological sense—a meaningful field in which something matters to someone.
His architecture is making progress on the first. It does not yet explain the second.
So I would state the limitation this way:
LeCun’s world model may enable a machine to predict what will happen if the cup falls.
It does not yet explain how the machine could understand why this particular cup must not be broken.
And this returns to our recent discussion. A world model may allow AI to enter physical reality. But an analogical self inhabits not merely physical reality; it inhabits a world of residues, histories, attachments, accidents and singular significance.
LeCun’s next AI revolution may therefore produce machines that can successfully move among things. The unanswered question is whether those things will ever become, for the machine, irreplaceable.
Dear friend, Yann LeCun’s argument is that the current generative-AI boom—especially large language models—is not yet the decisive AI revolution. LLMs manipulate language impressively, but they do not possess the kind of grounded, predictive understanding that even cats, infants, or experienced drivers display.
His proposed next revolution is the passage:
from word models to world models;
from generating plausible descriptions to predicting the consequences of actions.
1. Why LLMs are not enough
LeCun regards autoregressive LLMs as sophisticated next-token predictors. They absorb enormous amounts of human text and learn statistical regularities in language, but text is only a compressed and imperfect record of the world.
An LLM can explain that an unsupported cup will fall, but that does not mean it has acquired the intuitive physical understanding through which an animal anticipates the fall, adjusts its body and acts accordingly.
LeCun emphasizes a striking disparity: a child acquires vastly more sensory information through a few years of vision than the amount of linguistic information contained in all the text a person could read. Much of human intelligence therefore cannot originate primarily in language. It comes from observing a spatial, temporal and causally structured world.
In his view, current systems remain weak at:
- physical common sense;
- persistent understanding of objects and situations;
- reliable multistep planning;
- learning from relatively few examples;
- acting competently in unfamiliar environments;
- understanding how their own actions change what happens next.
This is why fluent language alone will not readily produce domestic robots, fully autonomous vehicles or genuinely autonomous intelligent agents.
2. What LeCun means by a world model
A world model is an internal model that represents:
- the present state of the environment;
- relevant information that is currently hidden or missing;
- how the environment is likely to evolve;
- what might happen under different possible actions.
It is therefore not merely a stored picture of the world. It is a counterfactual simulator.
An agent with a world model can ask internally:
What will probably happen if I do this rather than that?
A driver approaching an intersection, for example, does not merely recognize “car,” “road” and “stop sign.” The driver anticipates several possible futures: the other car may stop, fail to stop, accelerate or turn. The driver slows down because the internal model represents those possibilities before they become actual.
LeCun describes the world model as the most complex component in an intelligent agent. It must both reconstruct relevant hidden aspects of the present and predict multiple plausible future states, including futures caused by the agent’s contemplated actions.
3. Common sense is accumulated predictive knowledge
For LeCun, common sense is not primarily a database of verbal propositions such as:
- unsupported things fall;
- solid objects do not normally pass through one another;
- an object continues to exist when temporarily hidden;
- pushing something may make it move.
Common sense is a layered collection of predictive models acquired through observation and limited interaction.
An infant learns object permanence, intuitive physics, spatial relations and elementary causality without receiving millions of linguistic instructions. Animals do something similar. A cat need not verbalize gravitational theory to predict where it can land.
LeCun’s wager is:
If machines can learn comparable internal models through observation, they may acquire the substrate from which reasoning and planning arise.
That would be fundamentally different from accumulating more verbal facts.
4. Why predicting pixels is the wrong objective
A naïve world model might try to generate every pixel of the next video frame. LeCun thinks this is inefficient and often impossible because the sensory world contains enormous quantities of unpredictable and irrelevant detail:
- the exact movement of individual leaves;
- subtle textures;
- tiny changes in lighting;
- every hair or reflection;
- the precise trajectory of background objects.
A useful agent does not need to predict all of these. It needs to preserve what matters for action.
This is why LeCun advocates Joint-Embedding Predictive Architectures, or JEPAs. A JEPA does not necessarily reconstruct raw input. It predicts an abstract representation of one part or future state of the world from another representation.
Instead of asking:
What will every future pixel look like?
it asks something closer to:
What meaningful state is likely to come next?
V-JEPA, for instance, predicts masked portions of video in an abstract representation space. Because it can disregard inherently unpredictable details, Meta reported substantially better training efficiency than pixel-generating approaches.
This is one of LeCun’s sharpest departures from contemporary generative AI:
Intelligence does not require generating everything. It requires retaining what is relevant and predicting at the right level of abstraction.
5. Hierarchy makes long-range prediction possible
The future is predictable at different levels over different timescales.
At the level of individual hand movements, a cook’s exact trajectory may be predictable for only a fraction of a second. At a higher level, however, one can anticipate that the cook will gather ingredients, mix batter, heat a pan and make crêpes.
LeCun therefore proposes hierarchical world models:
- lower layers represent fine, rapidly changing physical details;
- higher layers represent objects, events, intentions and extended action sequences;
- increasingly abstract layers permit prediction over increasingly long intervals.
This matters because long-term planning cannot be performed efficiently by simulating every microscopic detail. An agent planning to cross a room should reason in terms of obstacles, routes and goals—not calculate every future pixel and motor-neuron-like movement.
A hierarchical JEPA would therefore compress the world into several usable temporal and conceptual scales.
6. From prediction to reasoning
LeCun uses “reasoning” in a relatively operational sense. Reasoning is not necessarily an invisible chain of linguistic propositions. It can be understood as searching through possible latent states inside a world model.
The system:
- represents its current state;
- proposes one or more possible action sequences;
- uses the world model to predict their consequences;
- evaluates the predicted outcomes;
- selects an action expected to minimize cost or reach the goal;
- acts, observes the result and updates the model.
Thus, reasoning becomes a form of inference or optimization through an internal simulation.
Recent research associated with LeCun is attempting to make these latent spaces more suitable for planning. Work on “temporal straightening,” for example, tries to organize internal trajectories so feasible futures become easier to predict and navigate mathematically.
So the revolution is not only better perception. It is the coupling:
perception → world model → imagined futures → evaluation → action.
7. From passive prediction to autonomous learning
LeCun distinguishes learning by observation from learning through action.
A system first watches the world and acquires broad regularities: objects move, people interact, tools have effects, events unfold in recognizable patterns. It can then act experimentally and learn how its own interventions alter outcomes.
A recent paper co-authored by LeCun describes this as combining:
- a system for observational learning;
- a system for active behavioral learning;
- a meta-control mechanism that decides when to observe, explore or act.
The aim is an agent that continues learning autonomously rather than undergoing one enormous training phase and then remaining largely fixed.
This contrasts with the present LLM pattern:
train on a huge static corpus → freeze most capacities → prompt the model.
LeCun’s imagined agent instead resembles an organism:
observe → predict → act → encounter error → revise its world model.
8. The architecture is larger than the world model alone
LeCun’s proposed autonomous intelligence architecture contains several interacting components.
The perception system estimates the current state of the world from sensory input.
The world model fills in missing information and predicts possible futures.
The actor proposes action sequences.
A cost or objective module evaluates predicted outcomes. It includes intrinsic constraints or drives—such as avoiding damage or unnecessary energy expenditure—and learned estimates of future cost.
A short-term memory maintains relevant recent states, actions and predictions.
A configurator adjusts the system according to the current task.
Planning occurs when the actor proposes actions, the world model simulates their effects and the cost module evaluates the predicted outcomes.
This is why LeCun does not expect intelligence to emerge simply by scaling a single undifferentiated transformer. He expects a modular architecture containing perception, memory, prediction, goals and action.
9. Why robotics is the decisive test
A language model can conceal misunderstanding behind fluent sentences. A robot cannot.
A robot that misunderstands gravity, object permanence, friction or causal consequences will drop the glass, collide with the furniture or fail to complete the task. Embodiment therefore places intelligence under the discipline of reality.
Meta’s V-JEPA 2 is presented as a foundation world model capable of visual understanding, predicting how physical situations develop and supporting planning for robot control, including some zero-shot control in unfamiliar environments.
These remain early results, not evidence that general-purpose household robots have arrived. But they illustrate the direction:
moving from recognizing what is visible
to anticipating what will happen
to selecting an action that changes it.
For LeCun, this transition is far closer to genuine intelligence than producing increasingly polished text or images.
10. Why he calls it the next revolution
The current generative revolution gives machines extraordinary access to the products of human intelligence: language, pictures, code and recorded knowledge.
World models could give machines something more fundamental:
a mechanism for acquiring knowledge directly from the world.
That could enable systems to:
- learn with far fewer labels;
- develop elementary physical common sense;
- adapt to unfamiliar environments;
- plan several steps ahead;
- operate robots and autonomous machines;
- continue learning after deployment;
- reason without translating every process into language.
LeCun has gone so far as to suggest that learning useful world models may be the central AI challenge of the coming decades.
11. What is still unresolved
LeCun’s vision remains a research program rather than a completed theory of intelligence.
Several difficult problems remain:
- representing several genuinely different possible futures;
- deciding what information is relevant and what may be discarded;
- planning over long periods;
- learning stable objectives and intrinsic motivations;
- preventing accumulated prediction errors;
- integrating language with physical and social world models;
- transferring learned models across radically different settings;
- creating persistent memory without catastrophic interference.
Even a strong physical world model would not automatically understand human institutions, irony, loyalty, shame, history or another person’s private world. The physical world is only one stratum of the human world.
That distinction is important. A machine may learn that a cup will fall from a table before it learns why a particular cup cannot be discarded because it belonged to someone who died.
My reading: world models restore the world to AI
LeCun’s disagreement with the dominant LLM paradigm is ultimately philosophical.
An LLM begins with signs: words referring to other words, produced from human archives.
A world-model agent begins with events: something moves, disappears, resists, falls, obstructs, returns or changes because the agent acted.
The first system inhabits a corpus.
The second is intended to inhabit a world.
This connects directly with our distinction between the algorithmic and analogical self. A token-predictive model is primarily trained upon regularities already sedimented in language. A world model must encounter uncertainty, resistance and consequences. Its predictions can fail because reality does something else.
But LeCun’s “world” remains predominantly a world of actionable physical regularities. It is the world required by an animal or robot: what exists, what may happen and what actions are possible.
Your use of “world” is wider. A world also contains:
- things that have returned to their proper places;
- historical residues;
- personal associations;
- the strangeness of another being;
- objects whose meaning exceeds their function;
- a niche in which a life can dwell.
Therefore, world models might produce the next revolution in machine agency without yet producing a world in the Heideggerian—or your analogical—sense.
I would put the distinction this way:
LeCun’s world model enables the machine to predict what the world will do.
A fuller intelligence would also have to discover why this particular world matters.
