AI Researchers Shift from Chatbots to World Models

A growing contingent of elite AI researchers, including "Godmother of AI" Fei-Fei Li and pioneer Yann LeCun, are shifting their focus from language models to what they call "world models"—AI systems designed to understand and interact with the physical world rather than just predicting words. The movement represents one of the most significant pivots in AI research since the rise of large language models.

For the past several years, the AI industry has been dominated by large language models (LLMs) like ChatGPT and Claude—systems trained to predict the next word in a sequence with remarkable accuracy. These models have transformed office work, creative fields, and countless other domains. But a growing number of researchers believe that this approach has fundamental limits. Language, they argue, is an abstraction of the world. True intelligence requires understanding the world itself.

"Where language models learn the statistical structure of text, world models learn the statistical structure of space and time: how light falls on a surface, how a garden looks from an angle no camera has captured, how objects respond to force and follow the laws of physics," Li wrote in a recent essay.

The shift is not merely academic. Over the past 18 months, more than $10 billion has flowed into world model and robotics AI companies. LeCun left his position as Meta's chief AI scientist to start Advanced Machine Intelligence (AMI) Labs, which has already secured over $1 billion in seed funding. Li's World Labs has raised $230 million from top-tier investors including Andreessen Horowitz. OpenAI is reallocating resources to "longer-term world simulation research." A new AI frontier is being defined.

The Limits of Language: Why Words Aren't Enough

Large language models have achieved remarkable feats. They can write essays, generate code, and engage in sophisticated conversations. But they have a fundamental limitation: they are trained on text, and text is an abstraction of the world. A chatbot can describe what happens when you drop a glass, but it doesn't truly understand the physics involved. It has learned the statistical patterns of language, not the causal dynamics of the physical world.

This limitation becomes apparent in real-world applications. A language model trained on a database of simulated taxi trips in New York City can provide effective directions—unless it is forced to take a detour, in which case it fails completely. It doesn't have an internal map of the city; it has memorized patterns from the data. This brittleness is a fundamental consequence of training on language rather than on the world.

Chatbots also can't interact with the physical world. As Carnegie Mellon's Martial Hebert put it: "Chatbots can't pick up a coffee mug." The geometry of the world, the dynamics of movement, the physical interaction with objects—all of this is far more complex than predicting the next word in a sentence.

For many researchers, these limitations suggest that language alone is insufficient for true artificial intelligence. "Humans not only do we survive, live, and work, but we build civilization beyond language," Li said. The next frontier, they argue, is building AI that can understand, navigate, and interact with the physical world.

What Are World Models?

The concept of a "world model" has become one of the most important—and most overused—terms in AI today. Definitions vary, but they all center on a core idea: an internal representation of the external world that enables an AI system to simulate what will happen next.

The term has its roots in neuroscience. In 1943, psychologist Kenneth Craik proposed that the mind works by running "small-scale models" of reality. We use these mental world models to navigate our surroundings and guide our actions. Our brains simulate our environments with enough fidelity to predict what will happen if we push a mug off a table or cross a busy street.

In AI, a world model is a system that learns the dynamics of an environment—how objects move, how they interact, how they respond to forces. Given a current state and an action, a world model predicts the next state. This prediction can be at multiple levels: visual (what will it look like?), structural (how will objects move?), or causal (what will be the consequences?).

Mathematically, a world model learns the transformation function p(s′ | s, a) that maps a current state and action to a future state.

The promise of world models is that they will enable AI systems to reason about the world in ways that language models cannot. Rather than merely predicting the next word, world models can simulate the future, plan actions, and adapt to new situations. This is the kind of intelligence that humans take for granted.

The Pyramid of World Models: Renderers, Simulators, and Planners

To bring clarity to the confusion, Fei-Fei Li and her team at World Labs published a "taxonomy of world models" that divides the field into three distinct categories.

This taxonomy helps distinguish between different approaches and their use cases. A renderer like Marble creates beautiful, interactive 3D environments—but it doesn't understand physics. It can show you what a building looks like, but it can't tell you if the building will collapse under its own weight. A simulator, in contrast, must respect physical laws. It can be used to train robots because it captures the underlying dynamics of the world.

Li's ultimate vision is a "unified world model": a foundation model that can render photo-realistic views, generate physically accurate structures, plan action sequences, and switch between different output modes according to downstream needs. She argues that "when their boundaries disappear, they will jointly reshape something more grand: the relationship between machine intelligence and the physical world."

Fei-Fei Li's Vision: Spatial Intelligence

For Fei-Fei Li, the path to world models runs through "spatial intelligence"—the ability to understand, reason about, interact with, and generate 3D worlds. The world is fundamentally three-dimensional, and AI must learn to operate in 3D space.

Li's World Labs is building world models that can generate persistent, downloadable 3D environments. The company's first commercial model, Marble, can create editable 3D worlds from text or 2D images. Users can export these worlds as Gaussian splats, meshes, or videos, and import them into game engines like Unity.

The technical innovation behind Marble is its use of Gaussian splatting—a technique that represents a scene as thousands of colorful, fuzzy points in 3D space. This approach is extremely fast and easy to use, though it prioritizes visual fidelity over physical accuracy. Marble captures "what the world looks like" rather than "how the world works."

This choice reflects Li's pragmatic approach. She sees immediate commercial applications in gaming, filmmaking, and creative industries. Her view is that building world models requires starting with what's achievable and scaling from there. "We aim to lift AI models from the 2D plane of pixels to full 3D worlds—both virtual and real—endowing them with spatial intelligence as rich as our own."

But Li also has her eye on the longer horizon. She has written about how world models could facilitate the development of robots that explore the deep sea, assist healthcare providers, and navigate unstructured environments. The challenge is to build the data and engineering infrastructure needed to make that vision a reality.

Yann LeCun's JEPA: The Cognitive Framework

If Li's approach is to build the interface, Yann LeCun's approach is to build the brain. LeCun has long been a critic of large language models, arguing that they lack the fundamental intelligence needed to understand the world.

"Continue stacking LLMs, feeding them more data, hiring thousands of people to hand-hold the system... in my view, this is completely nonsense," he has said.

LeCun's alternative is the Joint-Embedding Predictive Architecture (JEPA). JEPA doesn't predict pixels. It predicts abstract representations of the world in a latent space. The goal is to capture the causal structure of the world without wasting compute on generating pixels. It learns "how the world works" rather than "what the world looks like."

This approach is grounded in control theory and cognitive science. JEPA is designed to enable an AI agent "to predict the consequences of its own actions." It doesn't need to generate beautiful images—it needs to understand what will happen if a robot tries to pick up a cup.

In a May 2026 paper, LeCun and his collaborators provided mathematical proof that JEPA can learn the true underlying physical structure of the world. The paper showed that when the world's latent variables (like object position and velocity) follow a Gaussian distribution, the representations learned by JEPA form a linear mapping to the real world.

This is a foundational result. It means that JEPA is not just learning a convenient encoding—it is learning the actual physical structure of the world. This gives LeCun's approach a theoretical grounding that many other world model efforts lack.

LeCun's startup, AMI Labs, has already raised over $1 billion in seed funding, the largest seed round in AI history. The company is building on JEPA to create "systems that understand the physical world, have persistent memory, can reason, and can plan complex action sequences."

The Race: Who's Building World Models?

Li and LeCun are the most prominent figures in the world model movement, but they are far from alone. A rapidly growing ecosystem of companies, research labs, and startups is building world models across multiple approaches.

Google DeepMind has released Genie 3, a world model that generates interactive video environments. Unlike Li's approach, Genie 3 focuses on video-generation that allows users to explore environments freely. It solves the problem of long-term consistency—"no turning around and finding the whole building disappeared." However, it is still based on video logic rather than physics, making it more of a simulator than a planner.

Overworld, founded by Louis Castricato, is building interactive video game worlds where scenes adapt as virtual characters move through them. The company optimizes for interaction above all else, enabling users to walk through doors and interact with detailed environments.

Moonvalley, founded by former DeepMind researchers, is developing world models that emphasize "world simulation, adherence to physical reality, long-term consistency, and action-conditioned generation." The company is already training robots on its models.

In China, the Beijing Academy of Artificial Intelligence has unveiled Physis-v0.1, which it describes as the world's first general world foundation model. Unlike LLMs that learn from text, Physis is designed to learn and predict how the real world behaves. "Where language models learn the statistical structure of text, world models learn the statistical structure of space and time: how light falls on a surface, how a garden looks from an angle no camera has captured, how objects respond to force and follow the laws of physics," the researchers said.

NVIDIA is investing heavily in world models for robotics and autonomous driving. The company's DreamZero and DreamDojo projects provide virtual training grounds where robots can practice without risk.

The investment in world models has been staggering. MoE Capital estimates that over $100 billion has flowed into world model and robotics AI companies in the past 18 months. Venture capital is betting that world models are the next frontier in AI.

The Frontier: What Comes Next

The shift from chatbots to world models represents a fundamental reorientation of AI research. Several trends are likely to shape the next phase of development.

Simulation is the Key: Among the three types of world models, simulators are the bridge that connects renderers to planners. A model that masters simulation can project its understanding into pixels (for humans) and into actions (for robots). The most difficult open problems in the field also concentrate here. 3D data with clear geometric, material, and physical annotations is much scarcer than the internet videos used by renderers. And the "simulation-to-reality" gap remains a fundamental challenge.

The Rise of WAMs: World Action Models (WAMs) represent the next step in world model evolution. Rather than viewing VLA (Vision-Language-Action) models and world models as alternatives, researchers are merging the two. A WAM can "understand" (through VLM capabilities) and "predict" (through world model capabilities) within a unified architecture. Physical Intelligence's π0.7, for example, embeds a lightweight world model component to help the VLA "foresee" the state after completing a task.

Self-Evolution: To maintain coherence over extended horizons, world models must move beyond static prediction toward active calibration. This "self-evolution" involves continuous refinement through the interplay of generation, imagination, and environmental feedback. RoboGen exemplifies this by enabling agents to autonomously propose tasks and synthesize new environments, forcing the internal model to adapt continuously.

Physical Anchoring: Without grounding in physics, adaptability leads to unbounded drift. World models must be anchored to reality through physics-informed constraints—explicit differentiable dynamics or implicit intuitive physics. This ensures that self-evolution remains tied to the real world.

The transition from chatbots to world models is not about abandoning language models. It is about recognizing that language is one piece of a much larger puzzle. As LeCun has said, "The basic idea is that you don't predict at the pixel level." Instead, you train a system to run an abstract representation of the world so that you can make predictions in that abstract representation.

For the AI industry, this shift represents both a challenge and an opportunity. The challenge is technical: building world models requires different data, different architectures, and different evaluation metrics than language models. The opportunity is vast: world models could enable AI systems to finally understand and interact with the physical world, unlocking applications in robotics, healthcare, manufacturing, and beyond.

Fei-Fei Li's words capture the moment: "Language enables machines to talk about the world. And world models will enable machines to finally understand, imagine, reason, and interact with the world." The journey from language to the world has begun.