What are world models?

World models are AI systems that learn a model of how the world works: by observing video, sensors and simulations, they learn to predict what happens next, what effect an action will have, how objects, spaces and forces behave. It is a different approach from LLMs, which predict the next word in a text: here the goal is predicting consequences, not completing sentences. The best-known advocate is Yann LeCun, who argues that text alone is not enough for human-level intelligence, since even a cat understands the physics of the world better than any LLM: his research line aims to make models learn representations of the world rather than pixels or words. The most concrete industrial hook is robotics, where a world model can act as a learned simulator in which a robot tries millions of times before touching anything real, though honestly today these remain a research frontier rather than a procurement line companies can buy into.

LeCun's bet

The best-known advocate is Yann LeCun, one of the fathers of deep learning, who has been repeating the same thesis for years: text alone is not enough for human-level intelligence, because a cat understands the physics of the world better than any LLM. His research line (the JEPA architectures developed at Meta) aims to make models learn representations of the world rather than pixels or words. In November 2025 LeCun left Meta after more than a decade to found AMI Labs, a company dedicated precisely to world models. It is the clearest signal that the bet has moved outside corporate research. He is not alone: the big labs are working on generative video models and learned simulators pointing the same way. It is a genuine bet, in the sense that the outcome is far from settled.

Robotics, and what it means for you

The most concrete industrial hook is robotics: training a robot in the real world is slow, expensive and risky, while a world model can act as a learned simulator where the robot tries millions of times before touching anything real. It is the training substrate the whole Physical AI wave is betting on, from warehouses to manufacturing. An honest note: today world models are a research frontier, not a procurement line. There is no world model for your company to buy, and anyone selling you one as a finished product is selling hype. What you can do now is more boring and more useful: curate the data that will one day fuel these systems, meaning the telemetry, video and sensor streams of your physical processes, with decent quality and governance.