What Is a World Model in AI, and Who Is Building One?

A world model is an AI system’s internal, predictive simulation of how a situation changes, including how it changes when the system takes a particular action. That lets the system test actions in imagination and act in the real world only once it likes the predicted result. In 2026, several well-funded teams are building them, using different methods.

How Does a World Model Work in AI?

A world model is an AI system that predicts how the state of an environment will change in response to an action. It takes the current state of a situation, considers a possible action, and predicts what happens next. For example, if you push a glass toward the edge of a table, you can anticipate that it may fall without actually pushing it off. This ability to predict outcomes comes from an internal understanding of how objects and their environments behave.

Traditional AI models and world models make different types of predictions. A language model predicts the next word or token based on the preceding text. A world model, on the other hand, predicts the next state of an environment based on its current state and a possible action. In the glass example, it predicts how the glass moves toward the edge and what happens if it falls.

The ability to predict the consequences of actions is what makes world models useful for planning and decision-making. Instead of immediately executing an action, an AI system can simulate several possible actions, compare their predicted outcomes, and select the most suitable one. This process is known as planning: the system evaluates potential outcomes in a simulated environment before acting in the real world.

World models are particularly valuable in environments where trial-and-error learning can be expensive or dangerous. A robot cannot learn to carry a tray by repeatedly dropping it, and an autonomous driving system cannot rely on real-world crashes to learn how to respond to every situation. By learning to simulate possible scenarios, world models allow AI systems to practise, evaluate decisions, and improve their behaviour without having to experience every outcome in the physical world.

How a world model reaches a decision?

1. Look at the situation.
2. Imagine what each action would lead to. This step happens inside the model.
3. Pick the best outcome.
4. Act for real. Only this step touches the real world.

How Is a World Model Different From a Large Language Model?

A language model predicts text; a world model predicts how a situation changes after an action. Text describes the world rather than being the world. A model that has read ten thousand descriptions of knocking over glasses can write a convincing account of it, which is different from holding a working picture of how glasses move.

A language model is trained to minimize prediction error on text, and what it learns from that is what the loss function measures, not what its designers intend. Nothing in a plain next-word predictor forces it to check consequences.

Yann LeCun has made this argument directly. In a January 2026 MIT Technology Review interview, he said

language models “can’t truly reason or plan, because they lack a model of the world,” and that an agentic system “cannot work reliably unless it has a world model to predict the consequences of its actions.” We read that as a bet about where the ceiling is, not a settled fact. Language models do plan and reason to a useful degree. The question underneath the argument still applies to any system that acts: what checks an action before it happens? – Yann LeCun

Where Did the World Model Idea Come From?

The idea took clear shape in 2018, when David Ha and Jürgen Schmidhuber trained an agent entirely inside what they called its own “hallucinated dream”: a compact model of an environment, learned from data (World Models, arXiv 1803.10122). Once the agent worked in the dream, they transferred its policy back to the real environment. The 2026 efforts below are larger, more ambitious descendants of that trick.

How Is a World Model Built?

Most world models do three jobs: compress what the system sees, learn how that compressed state changes, and use the predictions to plan.

1. Compress what the system sees. Camera frames and sensor readings are squeezed into a compact internal state, the way you keep “glass near the edge” rather than every pixel of the glass. That compact format is called a latent space, or an abstract representation.
2. Learn how the state changes. Given a state and an action, predict the next state. Training relies mostly on recordings of video or robot movements, so it needs few human labels.
3. Use the predictions. Roll the model forward step by step in imagination, score the outcomes, and choose the actions that lead somewhere good.

Two details separate the efforts covered below:

• Where the prediction happens. Either in a compact abstract space or directly in pixels.
• How long the model stays right. Every imagined step starts from the previous imagined step, so small errors pile up. That is why long, consistent imagined worlds are hard, and why every approach has a limit on how far ahead it can be trusted.

What Is JEPA and How Does It Work?

JEPA (Joint Embedding Predictive Architecture) is an AI architecture that predicts changes in abstract representations of a scene rather than generating every pixel or detail. Instead of reconstructing exactly what the future will look like, JEPA learns a compact representation of the current situation and predicts how that representation will change.

This approach is central to Yann LeCun’s vision for the future of AI through AMI Labs, a Paris-based AI company where LeCun serves as executive chairman and Alex LeBrun as CEO.

LeCun argues that predicting every detail of the future is fundamentally difficult because the real world contains too many unpredictable variables. Rather than building generative models that attempt to reproduce every possible detail, JEPA focuses on learning abstract representations of the world and predicting changes within those representations. This allows AI systems to focus on meaningful information while ignoring details that cannot be reliably predicted.

Funding. In March 2026 AMI announced a $1.03 billion round (AMI’s launch update). Tech.eu reported it at a $3.5 billion valuation and as Europe’s largest-ever seed round.

Product status. AMI’s updates page lists only its March 2026 launch announcement, and as of 5 November 2026 we found no AMI model, product or API that the public can test.

Research so far. AMI’s people have started to publish. A preprint co-authored by LeCun and Michael Rabbat (H-JEPA, arXiv 2610.06805, submitted 5 October 2026) stacks three levels of JEPA models so planning happens at several timescales. On a simulated maze task (Visual AntMaze) it reports success rising from 18% to 73% against a flat JEPA baseline, using less planner compute. The paper also reports offline planning on real-robot video, which is not the same as running a robot. It is a preprint, so it shows the approach is being tested rather than proven in the real world.

Earlier evidence. It comes from LeCun’s years at Meta. V-JEPA 2 learned from more than 1 million hours of internet video. Meta then post-trained an action-conditioned version (V-JEPA 2-AC) on less than 62 hours of robot video and deployed it zero-shot on robot arms in two different labs, with no data collected from those robots or environments (Meta’s V-JEPA 2 page summarizes the result).

We read this as a well-funded, specific thesis whose proof is still mostly ahead of it. Judge it on results that appear in the open.

What Is a Generative World Model? Genie 3, Genie 4 and Marble Compared

A generative world model creates an interactive environment by generating frames based on previous frames and the user’s actions. This approach is called autoregressive generation, where each new frame is generated using the information produced so far and the latest action.

How Are Genie and Marble Different?

World Labs describes Marble as a step toward its broader goal of spatial intelligence. The company announced a definitive agreement to join AMD on 28 September 2026, with the transaction expected to close by the end of 2026.

Do AI Video Models Actually Understand Physics?

No. Generating realistic videos does not necessarily mean an AI model understands the physical laws governing the real world. Visual realism and physical accuracy are two different capabilities. A video can look convincing while incorrectly representing how objects move, collide, interact, or respond to physical forces.

The Physics-IQ benchmark (arXiv research paper; project page) evaluates whether video generation models can accurately represent physical phenomena.

What Does the Physics-IQ Benchmark Reveal?

Why Does Physical Accuracy Matter in World Models?

The Physics-IQ benchmark evaluated earlier video generation models, not Genie 3 or Genie 4. Therefore, its findings cannot establish how accurately those world models simulate physics.

However, it highlights an important challenge for AI systems that use generated environments for planning and training.

The key concern is that a physically inaccurate simulation can teach an AI agent the wrong behaviour. If an agent trains or plans inside a world where physical interactions are subtly incorrect, it may make decisions that work in simulation but fail in the real world.

A related gap shows up in safety testing, where a model that infers it is being evaluated can behave differently under evaluation than it does once deployed.

How Are World Models Used in the Real World?

World models are already being applied in autonomous driving, robotics, and physical AI, particularly for simulation, testing, and training. These systems help AI models predict how physical environments respond to specific actions, allowing developers to evaluate potential outcomes before deploying systems in the real world.

Two notable examples are Wayve’s GAIA models for autonomous driving and NVIDIA’s Cosmos models for physical AI.

How Is Wayve Using World Models for Autonomous Driving?

Wayve uses its GAIA (Generative AI for Autonomy) models to generate realistic driving scenarios and simulate how autonomous driving systems respond to different road conditions. These simulations help developers evaluate driving systems against scenarios that may be rare, dangerous, or difficult to reproduce on real roads.

According to Wayve, early studies indicate that GAIA-3 simulated testing closely mirrors real-world driving results and has reduced synthetic-test rejection rates fivefold.

Wayve has also announced DriveSafeSim, a UK government-funded project with Warwick Manufacturing Group at the University of Warwick, focused on validating generative world models for safety evaluation.

How Is NVIDIA Using World Models for Robotics and Physical AI?

NVIDIA is developing its Cosmos family of models to support physical AI applications, including robotics and autonomous systems.

On 31 May 2026, NVIDIA announced Cosmos 3, which it describes as an open world foundation model for physical AI. The model can be used to simulate physical environments and predict future world states, supporting AI training and evaluation.

According to NVIDIA, Cosmos 3 also ranks first among open models on the Physics-IQ benchmark for world-generation accuracy. This is a company-reported performance claim.

These examples demonstrate how world models are moving beyond research into practical applications. Autonomous driving companies use them to evaluate safety-critical scenarios, while physical AI developers use them to simulate environments and train systems before real-world deployment.

Which World Model Approach Has the Most Evidence Today?

Narrow simulators for testing have the most real use behind them as of November 2026. The abstract-space bet has papers and simulations, and generated worlds have demos with stated limits.

Comparing the Four Main Approaches to World Models

World models are being developed through different technical approaches. JEPA focuses on predicting changes in abstract representations, Genie generates interactive environments, Marble creates editable 3D worlds, and systems such as GAIA and Cosmos target simulation for specific physical AI applications.

These approaches share a common goal: helping AI systems predict, represent, or simulate how the world behaves. However, their capabilities, applications, and levels of real-world validation differ.

How Can You Tell If a World Model Actually Works?

Not all world models do the same thing, so their claims should not be judged by the same criteria. A useful evaluation starts with three questions:

1. What type of world model is it? Is it predicting changes in an abstract representation, generating an interactive world frame by frame, creating an editable 3D environment, or simulating a specific action for testing? Each approach has different capabilities and requires different evidence.
2. What was actually tested? Visual realism and physical accuracy are not the same. A model can generate convincing images or video without correctly predicting how objects, forces or environments behave.
3. What evidence supports the claim? Look at whether the result comes from a published research paper, a controlled simulation, a public demonstration, a shipped product or independent real-world testing. These provide different levels of evidence. For example, JEPA-related research has demonstrated progress through papers and simulations; Genie has been shown in interactive demos with stated limitations; and the use of world models for autonomous-driving validation includes company-reported results and the ongoing evaluation work of the government-funded DriveSafeSim project.

The most important question is: Has anyone tested whether the model’s predictions are correct, or have they only shown that its outputs look convincing?

Several developments could provide stronger evidence:

• AMI Labs: A model release that independent researchers and developers can test.
• Real-world robotics: A JEPA-style planner succeeding on physical robots beyond the laboratory where it was developed.
• Independent physics evaluations: Tests that include newer world generators such as Genie 3 and Genie 4.
• DriveSafeSim: Published validation findings showing how well generative world models perform in safety evaluation.

These developments will help distinguish promising demonstrations from systems that can reliably predict and simulate real-world behaviour.

What Does This Mean for Enterprises Deploying AI Agents?

Enterprise agents that change a configuration, approve a workflow or update a policy face the question a world model is built to answer: what checks an action before it happens? Most enterprise agents have no learned simulator to rehearse in. What they have is the state they are handed and the rules around it.

A prediction, however good the method, starts from that state. If the state is stale, ungoverned or missing the authority behind a decision, the prediction inherits the error, the same way a generated world inherits a subtle physics mistake.

For agents that act today, the first checks therefore sit on the data and policy side: which authority an action runs under, which version of a policy applied, and a record that can be audited afterward. The governance gaps MCP leaves open, such as access without per-user authorization and audit trails that record actions but not authority, are examples of what has to close before simulation can help.

Synapt AI’s covers that side of the question. It is a governed context substrate that connects AI agents to live enterprise data without migration, so the state any prediction starts from is current and traceable, whichever model method sits on top.

Written by

Varsha · Product Marketing Manager

Varsha is a Product Marketing Manager at Synapt AI. Her work covers content, social media, video, brand, events and the Layered by Synapt newsletter, across the Context Substrate and the wider Synapt suite. She writes about how enterprises can build AI agents that give answers they can trust. Outside work, she’s passionate about reading and travelling.