Most of the discussion about AI centres on LLMs, but this is just one branch of AI. We have machine learning, reinforcement learning, symbolic AI, expert systems and some more exotic ideas such as energy-based models. However, there is another important flavour of AI that deserves attention, and that is what is known as a world model. Large language models are trained on huge quantities of text and predict a likely next word (strictly, the next token, a small word or chunk of a word) in their output, one token at a time.
There are limitations to this approach. LLMs are static, so once trained they do not improve their understanding without retraining. Their understanding of physical reality is generally indirect, rather than grounded in physical interaction. The way that objects in the real world behave, such as how objects fall or break, or where things will be in the future if they roll along a surface, is not something that they learn directly. LLMs with multimodal training may gain knowledge about physical reality, but this understanding may be incomplete, unreliable or insufficiently grounded in direct interaction. This means that LLMs are of limited use in the physical world, such as steering autonomous vehicles or controlling robots.
A world model is an internal representation that lets a system predict how its environment will change in response to actions. It is effectively a simulator, predicting what will happen next. It works by taking raw sensory input such as sensor readings or pixels, encoding these and predicting a transition in a chain of states. A planning system can use such a model to test possible actions, select one and then act. A self-driving car might predict that a pedestrian is likely to step off the kerb into the road, for example. To consider another example, a robot might internally simulate three ways to grasp a glass, select the approach with the lowest predicted risk of spilling water, execute it and use the outcome to improve future predictions. A world model does not have to simulate every detail of reality; it just needs to capture enough of the environment to make useful predictions.
One of the most important applications is generating realistic driving scenarios that are rare or dangerous to capture in real life, such as a truck shedding its load or a child running into the road, so that driving systems can be trained and tested against them. In robotics, world models can generate scenarios such as boxes falling from warehouse shelves, and provide simulated environments in which robots learn to navigate and manipulate objects. They are also used for providing practice environments for reinforcement learning agents.
A summary of the differences between an LLM and a world model is shown in the following diagram.
Examples of world models are Cosmos by NVIDIA and Genie 3 from Google DeepMind, and there are other developments from AMI Labs (founded by AI pioneer Yann LeCun). Another notable example is Marble from World Labs, which was founded by Fei-Fei Li. She is a professor of computer science at Stanford and a former chief scientist at Google Cloud. She led the project that created ImageNet, which helped make modern computer vision practical. AI video-generation company Runway has built a family of world models (GWM-1) designed to simulate interactive environments, generate digital characters and support robotics applications. Another company working on this is xAI, which in 2025 hired two researchers, Zeeshan Patel and Ethan He, from NVIDIA to work on this. Meta released a world model, V-JEPA 2, in June 2025. British company Wayve, an autonomous driving company with a supervised robotaxi service running with Uber in London since September 2026, has a world model called GAIA that helps develop and test driving systems; GAIA is not necessarily the model controlling a vehicle in real time.
Many of the newer world models are still in their early days. A May 2026 benchmark called What-If World tested nine models on their ability to generate physically consistent videos when environmental conditions changed. None scored above 52%, while open-source models clustered around 28%. AMI Labs CEO Alexandre LeBrun predicted that “world models” would become the next buzzword, with every company calling itself one within six months to raise funding. Some of the biggest names in AI are now leading world model efforts, so this is an area to watch closely.
