Quick answer
A world model is an AI system trained to predict how an environment — physical or simulated — changes over time in response to actions, rather than predicting the next word in text. Instead of "what word comes next," it learns "what happens next": if a ball rolls off a table, where does it land; if a robot arm moves this way, what does the scene look like afterward. World models underpin robotics training, self-driving simulation, and video generators like Sora 2, which implicitly learn a rough model of physics to produce coherent footage. Genuinely robust world models remain an open research problem.
Large language models are extraordinarily good at one specific thing: predicting the next piece of text given everything before it. A world model is aimed at a different, arguably harder problem — predicting how the physical or simulated world itself evolves, not just what words describe it.
Why this is a different, harder problem
Text is a compressed, tidy representation humans invented for communication. The physical world is not tidy — objects occlude each other, physics has to stay consistent frame to frame, and there are infinitely more ways for a video of a bouncing ball to look wrong than for a sentence to look wrong. A model that has to predict "what does this scene look like one second from now" is solving a much messier problem than "what word comes next."
- Robotics: training a robot in a simulated world model is far cheaper and safer than trial-and-error on real hardware
- Self-driving: predicting how traffic, pedestrians, and other vehicles will move a few seconds ahead
- Video generation: Sora 2 and similar models implicitly learn a rough world model of physics and object permanence to keep generated video coherent
- Game AI: simulating a playable environment without needing a hand-built game engine
A genuinely robust world model — one that generalises to situations very different from its training data, the way humans handle novel physical situations easily — is widely seen as a missing piece on the road toward more general AI.
Where the frontier actually is
Today's best world models are impressive within familiar situations (a ball rolling, a car turning, a hand picking up an object) and still fail in ways that reveal they haven't truly learned physics the way a person understands it — they've learned strong statistical patterns that resemble physics most of the time. Closing that gap is one of the most closely watched research directions across every major AI lab in 2026.
Bottom line
World models are the quiet research bet behind a lot of 2026's progress in video generation and robotics. They are a genuinely different kind of AI problem than language modelling, and one that is nowhere near as solved.
