World Models: The AI Technology Teaching Machines to Understand Physics
A new generation of AI systems can simulate gravity, predict collisions, and generate interactive 3D worlds. Here is what world models are, who is building them, and why they matter.
Key points
- Nvidia released Cosmos 3, a family of open-weight world model AI systems, in June 2026, with versions ranging from 16 billion to 64 billion parameters.
- Google DeepMind's Genie 3, introduced in August 2025, generates photorealistic interactive 3D worlds from a text description at 24 frames per second.
- Waymo adopted Genie 3 in February 2026 to simulate rare driving scenarios, such as unusual pedestrian behaviour, without putting a vehicle on public roads.
- World Labs raised $1 billion and Yann LeCun's AMI raised $1.03 billion to fund world model research in 2025 and 2026.
- Alibaba unveiled its world model, Happy Oyster, in April 2026, capable of generating three-minute in-world video clips.
Most AI you've heard about, the chatbots, the image generators, the writing assistants, works by predicting patterns in text or pixels. Ask one what happens when a ball rolls off a table and it'll give you a plausible-sounding answer. It doesn't actually know what gravity does. It has never modelled the physics.
World models are different. A world model is an AI system that builds an internal picture of an environment, cause and effect included, and uses that picture to predict what will happen next. Think of it as the difference between a parrot that can describe a chess game and a player who actually understands the consequences of each move.
Where did world models come from?
The idea is older than most people realise. Researcher Jürgen Schmidhuber introduced the concept in 1990, proposing recurrent neural networks (layers of software loosely inspired by the brain) that could predict future states of an environment. David Ha and Schmidhuber revived it in a 2018 paper, in which AI agents learned to drive virtual cars by practising inside simulations they generated themselves.
Meta's chief AI scientist Yann LeCun pushed the idea further in a 2022 paper arguing that true machine intelligence needs predictive world models, not just pattern matching. His team developed V-JEPA 2, which reached top performance on video understanding and physical reasoning and can guide a robot in a room it has never seen before. We've followed LeCun's work since our first report on 16 July 2026.
What is actually being built right now?
The field has moved fast. Here is a snapshot of the major releases.
| Model / Product | Creator | Date | What it does |
|---|---|---|---|
| Genie 3 | Google DeepMind | August 2025 | Photorealistic interactive 3D worlds from text, at 24 fps |
| Waymo World Model | Waymo / DeepMind | February 2026 | Simulates rare driving edge cases for self-driving cars |
| Happy Oyster | Alibaba | April 2026 | Generates 3-minute in-world video; includes a directing mode |
| Spark 2.0 | World Labs | April 2026 | Open-source 3D rendering engine for smartphone-class hardware |
| Cosmos 3 | Nvidia | June 2026 | Combines physics reasoning, simulation, and action generation |
Nvidia's Cosmos 3 is the most technically ambitious entry. It processes text, images, video, audio together with physical action sequences, rather than treating each as a separate task. Cosmos3-Nano, at 16 billion parameters (a parameter being a single adjustable value inside the model, with more parameters generally meaning more capability), is designed to run on a workstation. Cosmos3-Super, at 64 billion parameters, targets research labs.
Cosmos 3 is released as open-weight, meaning anyone can download and inspect the model's internal values. That matters for researchers who want to build on it without paying API fees.
What does this mean for ordinary people?
Most near-term impact lands in a few places.
Self-driving cars get safer training. A robotaxi can't wait to encounter a tornado or a toddler running into traffic before learning to handle it. World models let companies like Waymo simulate those situations thousands of times in software first. Waymo is currently expanding into three new cities, and the ability to stress-test edge cases cheaply is part of what makes that pace possible.
Robotics gets cheaper to develop. Robots trained inside world model simulations can transfer those skills to the physical world, cutting the number of expensive real-world trials. Our 8 September story on the AI researcher who taught robots to imagine shows how seriously independent researchers are taking exactly this bet.
Games and creative tools change shape. Genie 3 lets a user type a description and walk around inside the resulting environment seconds later. Game studios are already using this to prototype levels faster.
Funding is flowing heavily into the space. General Intuition closed a $133.7 million seed round, World Labs raised $1 billion, and LeCun's AMI raised $1.03 billion. When that much capital concentrates on a single approach, products tend to reach consumers faster than the underlying research timelines suggest.
Common questions
Is a world model the same as ChatGPT or Claude?
No. Large language models, the technology behind ChatGPT and Claude, predict the next word in a sequence and work almost entirely in text. World models predict what happens next in a physical environment, working with video, sensor data or spatial information. They overlap in some architectures but serve different goals.
Can I use any of these tools today?
Nvidia's Cosmos 3 is available as open-weight software for researchers and developers. World Labs released Spark 2.0 as open-source. Consumer-facing products built on these systems are beginning to appear, but most current tools target developers and enterprise customers rather than everyday users.
Are world models safe?
Video prediction accuracy and robot success rates inside a simulation don't always predict real-world performance, which is one reason safety testing remains a live concern. The field is still defining standards, and no major regulator has issued specific guidance on world models as of mid-2026.
The honest read on this moment: the physics really does work inside these systems, which is new, but the gap between a convincing simulation and reliable real-world behaviour is where the next five years of hard work will happen.
