What are 'world models,' and why AI researchers think they're the next leap past chatbots

A robot arm trained to pick up a coffee cup in a tidy lab stalls the moment the chairs are moved, the lights dim, or a cat walks through the frame. A chatbot that can draft a polished email cannot tell you whether the sidewalk is wet. The gap is not a matter of scale or compute; it is a difference in what each system is, at root, trying to predict.

A world model is an AI system that builds an internal model of its operating environment so it can predict what will happen when it (or something else) acts. The "world" here is narrow by design. As the Live Science explainer puts it, the term does not mean the entire physical reality. It means the warehouse, the road, the video game, the chemistry bench, the kitchen, the room the robot is in. Inside that world, the model learns the rules well enough to simulate a few steps ahead.

That is the representational shift that separates a world model from a large language model. An LLM is trained to predict the next token in a sequence of text, which makes it a powerful language engine and a poor navigator. A world model is trained to predict the next state of an environment, which is a different job with different data and different failure modes. Once the model can imagine the consequences of an action, you can use it to plan, to train other agents inside a simulated copy of the world, or to test what would happen if conditions changed.

The same labs that built the LLM boom are now reorganizing around that shift. Within twelve months, DeepMind shipped Genie 2 and then Genie 3, large-scale foundation world models that generate interactive environments from a single prompt image. NVIDIA released the Cosmos World Foundation Model Platform for Physical AI, pitched at robotics and autonomous-driving developers who need a simulator they can fine-tune on their own sensor data. Meta released V-JEPA 2, a self-supervised video world model that learns physical intuition by watching unlabeled video, and Physical Intelligence published π0, a generalist policy for robots that the company describes as a step toward models that can operate across many embodiments rather than one. The names diverge, but the bet is the same: a model that can simulate what an environment does next is more useful for action than a model that can only simulate what a sentence does next.

In a robotics lab, a robot that has only ever seen a clean tabletop can be dropped into a kitchen it has never visited, asked to fold a towel it has never held, and the world model underneath it can imagine how the towel will deform as the gripper closes, what the counter will look like if the towel slips, and which path through the room keeps the arm away from the cat. Whether that picture is real or aspirational is one of the open questions the field is still working through.

The term "world model" reaches back to 1950s cognitive science, where it described the mental simulations humans use to navigate their environment, and to control theory in applied mathematics, where engineers have long built internal models of the systems they want to steer. The contemporary AI meaning inherits both: a learned, internal simulation good enough to act on. Yunzhu Li, an assistant professor of computer science at Columbia, told Live Science that the idea has strong connections with the intuitive models humans use to navigate their environment; in her framing, world models are the attempt to give machines something similar.

Some researchers insist the term should be reserved for systems that genuinely model the underlying physics of an environment; others use it for any generative model whose outputs are consistent enough to be treated as a simulator. That disagreement is not a marketing problem to be dismissed; it is the field working out what the object of study actually is. The original Genie paper and the V-JEPA 2 preprint are not yet peer-reviewed, and lab capability claims about real-time interactivity, long-horizon consistency, or sim-to-real transfer should be read as the labs' own framing of what they have shown, not as third-party validated results.

The next twelve months will tell. Genie, Cosmos, V-JEPA 2, and π0 share a vocabulary, not a definition. If a robotics team can use one of them to deploy a policy in a new building without retraining from scratch, the term will harden. If the releases keep diverging in what they take as input and what they guarantee as output, the wire will keep using "world model" as a synonym for whatever a frontier lab shipped that quarter.