World models

Richard Sutton, winner of the Turing Award and author of the 2019 essay “The Bitter Lesson,” argues that more computing power alone is not enough to advance AI: new learning methods are needed. He advocates AI agents that build world models and continuously learn from their own experience. Other researchers agree that the success of scaling large language models does not prove that this approach alone will lead to more capable AI. But leaders at OpenAI and Anthropic are confident that scaling still has room to grow.

In 2016, Pokémon Go players scanned the world around them while hunting virtual Pokémon. Data from those recordings is now helping develop world models: AI systems that build digital environments and learn to predict how objects change within them. Niantic, the company behind Pokémon Go, sold the game and shifted its focus to this technology. World models could become an important part of the development of home robots and self-driving cars. A robot, for example, could practice tasks in a virtual copy of a room, while a car could learn from simulated hazardous situations.

Google DeepMind has released Genie 4, a world model that adds audio, more realistic scenes and real-time control over what happens inside them. The update arrives as rivals build interactive worlds of their own and companies look to simulation for applications from games to robotics. Genie 4 can keep a scene consistent for up to two minutes—twice as long as Genie 3—but Google says the system is still early, with no settled product plan and unresolved problems with audio accuracy and readable text.

Odyssey has opened public access to Odyssey-3, a model that generates interactive worlds in real time and predicts how they respond to user actions. Its free demo lets people explore environments built from text prompts, while the company is also positioning the technology for robotics, drones and games. The release adds technical details and benchmark results to a model founders Oliver Cameron and Jeff Hawke first introduced on September 15.

JEPA-Anything asks whether one world model can handle tasks from physics and robotics to biology. PhAI Labs researchers and colleagues at Stanford, Oxford and Princeton built the model around Yann LeCun’s JEPA approach, then tested it against a standard JEPA across ten tasks. Its strongest results came in dynamic systems, while its most consequential claim is biological: the model identified a potential combination for attacking liver cancer, then researchers tested it in cells, organoids, tissue samples and mice. That is evidence of a promising research lead, not a treatment.

Microsoft Research has introduced Quine, an AI system designed to help researchers study complex biological processes. Rather than treating each kind of biological data as a separate problem, Quine aims to connect evidence across scales, from genes and proteins to cells, and use it to help plan experiments. The project is an early step toward a biological “world model”: one that predicts how a system might respond to an intervention before researchers commit scarce laboratory time to testing it.

A British startup wants to turn video-game play into training data for AI world models, systems designed to learn how actions change the physical world. Worldmodeldata says it has licensed nearly one million hours of gameplay from studios behind popular games. The pitch is that games already produce vast streams of visual scenes paired with player inputs, offering researchers data that is difficult to collect by hand. But the approach has a central limitation: a game can look real without behaving like the world a robot must handle.

Sakana AI has appointed Jürgen Schmidhuber as a scientific advisor, bringing a researcher it calls the “father of modern AI” into its effort to build self-improving systems. The company points to his work on world models in 1990 and deep-learning methods published in 1991. The appointment also connects his earlier research on meta-learning and recursive self-improvement with Sakana AI’s plans for autonomous scientific and physical systems.

At this week’s All In conference, I moderated a panel on world models—an area with major backers, prominent founders and almost no clear product roadmaps. AMI Labs, founded by Yann LeCun, and World Labs, founded by Fei-Fei Li, have drawn significant attention and funding, yet neither has made a convincing case for where its technology will first earn money. That secrecy may reflect early-stage research. It may also be a deliberate strategy: in a field broad enough to cover robots, video, graphics and autonomous systems, revealing the destination could help competitors catch up.

A year ago the daily Scholar Inbox digest of robotics papers was almost entirely VLA — vision-language-action models built on a pretrained vision-language backbone. Now a second acronym turns up nearly every day: WAM, world-action models, which start from a pretrained video or world model instead and predict future states and robot actions together. In October 2025, in a piece called "The…

Runway has introduced Solaris, a research system that generates a program's interface in real time instead of running one. A language model decides how the screen should change; a world model draws each new frame. Underneath there is no application — no layout, no components, no code. Solaris is built on Gen-4.5, Runway's video model, and continues the direction the company opened with GWM-1.…

Nvidia has published a recipe for turning Cosmos 3 Edge, a 4-billion-parameter world model, into a robot manipulation policy that runs entirely on a Jetson Thor with no server GPU in the loop. In closed-loop testing on the RoboLab benchmark, the post-trained policy completes 22.9% of tasks. Nvidia's own larger Cosmos 3 Nano reaches 36.8% on the same benchmark. The company is shipping the…

Hafner, 31, has left Google DeepMind and started a robotics company whose name is not yet on the door. The office in San Francisco's SoMa district is close to empty: one employee on the day of a visit, almost no furniture. The exception runs down the middle of the room, where humanoid robots of different shapes and sizes hang from racks like marionettes. The robots are shipped in from China.…

Generative AI comes with a convenient illusion of competence: the model writes, draws, sometimes even "plans," and it looks as though there is something like a picture of the world inside it. But the moment a system is asked not to continue text but to act toward a goal over a long horizon — drive a robot, navigate websites, negotiate with…

Generative models have learned to produce striking clips, but that kind of video has a weak spot: nothing forces it to obey physics. An object can drift in mid-air, particles can ignore gravity, rigid bodies can pass through each other. For spatial intelligence that is not enough. What you need is a world model that doesn't just look plausible but behaves plausibly, because a simulation is…

Over the past couple of years it has become fashionable to talk about world models: systems that do not merely continue text or fill in the next frames, but understand at least a little of how reality is put together and how it changes over time. The authors of Research on World Models Is Not Merely Injecting World Knowledge into Specific Tasks…

Not long ago, models learned to generate video from text with a few seconds of coherent motion. But ask one of those systems to walk forward, look back and return to a familiar object, and the magic stops. Objects trade places, details drift, and causal logic gives way to statistical coincidence. That gap between a nice-looking picture and an actual simulation of a world is what the authors…

When we ask a machine to help us cook dinner or swap a SIM card, it has to do more than recognize the objects in frame — it has to picture how the world will change from one step to the next. Most systems today see pixels and answer in short phrases, and long-horizon planning still does not work. The VLWM (Vision Language World Model) team proposes a different route: describe the future in…

Interactive world models are a way to teach AI to sense the world rather than only describe it in words. Until recently, three obstacles stood in the way: there was not enough quality data with precise action labels; classic video diffusion models were too slow to compute and "forgot" the start of the clip; and errors compounded from frame to frame. Matrix-Game 2.0 offers a clear, practical…