At the recent AI Action Summit held in Paris, Yann LeCun, a key figure in the development of artificial intelligence and a pioneer in neural networks, launched a forceful criticism of large-scale language models (LLMs), such as those that dominate the AI landscape today.
According to LeCun, these models, although impressive in linguistic tasks, do not possess the necessary capabilities to achieve intelligence comparable to that of humans.
Specifically, he highlighted four main deficiencies: they do not understand the physical world, they lack persistent memory, they do not reason effectively and they are incapable of planning complex actions in a hierarchical manner.
To overcome these limitations, it proposes the development of World Models, systems that simulate the real world more complete.
These statements not only reflect a critical vision of the current state of AI, but also set a course towards a new paradigm in research. But what leads LeCun to say this?
Limitations of Large Scale Language Models (LLMs)
LLMs, like OpenAI’s GPT-4o or Meta’s Llama 4, are AI models based on neural networks, especially in the transformer architecture (that’s the “T” in GPT).
Trained with enormous amounts of text, these systems stand out for their ability to generate coherent text, answer questions or translate languages. However, its exclusive dependence on written language reveals serious limitations.
Understanding the physical world
Although an LLM may describe “an object falling due to gravity,” it lacks direct sensory experience. It does not perceive the texture, weight or resistance of the air; it simply associates words and statistical patterns. This lack distances it from the basic intuition that even an animal without formal training applies when interacting with its environment.
Persistent memory
Each application to an LLM is essentially treated as an isolated interaction. Although they can maintain context during a single session, they do not incorporate learning from previous conversations on an ongoing basis. A human, on the other hand, builds a wealth of experiences that guides his behavior and allows him to adapt in the long term.
Limited Reasoning
LLMs are based on learned correlations to generate “logical” answers, but they do not carry out an authentic deductive process or understand cause and effect relationships. They may associate the sun with sunrise because they have seen the pattern in thousands of texts, but they do not reason why it occurs, nor do they extract universal principles that transcend the examples seen.
Inability to plan
Faced with complex tasks (such as planning a trip, designing a business strategy or coordinating the actions of a robot), LLMs do not decompose problems into hierarchical steps. Their approach is essentially reactive: they respond to immediate indications without drawing up a global plan or anticipating medium and long-term consequences.
These limitations reflect the Moravec paradox, which highlights how easy it is for current AI to tackle abstract tasks (language) and how difficult it finds the perceptual and motor functions that humans and animals perform almost automatically.
The World Models: A Promising Alternative
Faced with the limitations of LLMs, LeCun advocates World Models, systems designed to create internal representations of the real world.
World Models are systems designed to reconstruct internally the dynamics of the real environment. Instead of limiting themselves to language, they integrate heterogeneous data (images, audio, video, physical sensors) to learn the laws that govern the world: gravity, friction, causality and spatial transformations.
Key Advantages
- Physical understanding: By processing direct sensory information, a World Model can predict how a ball will roll, how an object deforms when collided, or how a liquid flows, simulating physical laws in a “virtual laboratory.”
- Long-term memory: They maintain an internal state that evolves with each interaction, allowing continuous learning and the refinement of models according to new experiences.
- Causal Reasoning: When constructing representations of cause and effect, they are not content with correlations; They can explain why a change in the environment produces a certain result.
- Strategic planning: This internal simulation capability allows them to decompose complex tasks into subtasks and evaluate different routes of action, crucial in robotics, automation and autonomous vehicles.
A notable example is Cosmos from NVIDIA, which uses World Models to create varied physical scenarios in which robots and autonomous vehicles rehearse maneuvers, anticipate obstacles, and learn to react to new situations.
How do LLMs and World Models compare?
To clarify the differences, LLMs rely exclusively on text and focus on linguistic patterns, while World Models integrate sensory data to model the physical world.
The former lack persistent memory and deep reasoning, limiting themselves to tasks such as chatbots or translations. On the other hand, the latter offer a solid foundation for systems that need to interact with the environment, such as robots or simulators.
This distinction underlines why LeCun sees the World Models as a necessary advancement.
Challenges and future of World Models
Although promising, the World Models face challenges. They require large volumes of sensory data and significant computational power, as well as ensuring they can be generalized to new situations.
Some experts question whether this approach is truly innovative, but advances in hardware and data make it viable today. Projects such as Meta’s V-JEPA or DeepMind’s simulators show that this direction is gaining ground.
Yann LeCun’s words in Paris capture a turning point in AI: LLMs, while extraordinary in language processing, fall short when it comes to emulating complete intelligence, with physical sense, long-lasting memory, deep reasoning and strategic planning.
The World Models are emerging as the way to overcome these barriers, offering rich simulations that connect statistical theory with empirical experience.
This post is also available in: