The news lately has been full of releases called “world models,” along with articles describing how they differ. These are notes I’ve been keeping as I follow the space.

What I’m mainly interested in is how a new kind of AI model lets other companies build products around it. Language models first gave us chat-based assistants tailored to different verticals. Once they started to loop, call tools and write code, that led to an explosion of products beyond chat. What kinds of products could “world models” enable? That question guides these notes.

What is a world model?

I’m using “world model” broadly here, because the releases I’m following do quite different things. Some generate the next view as a person moves through a scene; others predict what is likely to happen without rendering a picture at all. What they share is the ambition to learn enough about an environment to predict or generate what follows from an action.

The comparison with language models helps me see that ambition. A language model predicts what text comes next; these models try to retain something about a place or physical situation so that an action has a plausible consequence. But I don’t think the label alone tells me much about what another company could build. Generating a convincing scene and maintaining a world that someone can keep using are different requirements.

Project Genie

What interests me about Google’s Project Genie is that the user can move through the generated scene. Instead of requesting a video and watching it play, a person describes a world and explores it while the model generates what comes next in response. Google released the experiment to U.S. Google AI Ultra subscribers on January 29, 2026. Each exploration lasts up to 60 seconds. Project Genie announcement

It helps me to picture a still image of a room. Generating that picture is one task. Letting me walk around the corner means deciding what exists outside the original frame. And if I then turn back, I expect the room to still be there, with its objects in the same places. My actions now set requirements for what the model has to remember.

That memory is what I want to follow beyond the demo. Google warns that controls can lag and that a generated world may ignore the prompt or ordinary physics. A short exploration could still be useful for sketching an idea. But for a product someone keeps coming back to, I need to understand how much of the world survives each return.

Games and reusable 3D scenes

Could a developer build a game around this? I can imagine using a generated scene as a starting point, but I’m less sure what happens once the player changes something. If I open a door, it should still be open when I come back. A game designer also has to decide what opening that door does, then keep that rule consistent across repeated play.

That is why I’m following World Labs’ Marble alongside Genie. Marble, launched in November 2025, generates reusable 3D scenes that can be explored and used as 3D content. Its World API, announced in January 2026, lets developers request those worlds from their applications. Marble; World API

With Marble, my product question is what a developer can do with the scene afterward. Can they edit it and make it part of something larger? Genie’s interactive generation raises a different question: how the world responds while someone is exploring it. I don’t yet see a reason to judge the two by which demo looks more impressive. The useful comparison is what each leaves another company able to build.

Prediction without pictures

Most of the examples so far are easy to discuss because I can imagine looking at the output. My interest in applied AI also points me toward a less visible use for a world model: predicting what will happen well enough to inform a decision. Does the model need to render a picture for that?

Yann LeCun’s JEPA line of research explores prediction in a learned representation space. In V-JEPA, the model predicts information about missing parts of a video without reconstructing each pixel. The representation is a compressed description learned by the model, not an image meant for someone to watch. Meta’s V-JEPA explanation

Suppose I only need to predict where an object will move. Position and motion matter for that task; the exact texture of the wall behind it may not. I can see the appeal of learning a representation that keeps the relevant information without spending computation on every pixel. What I still need to know is whether that representation improves the prediction or decision of the system that uses it.

I’m following LeCun’s new company, AMI Labs, with that question in mind. Judging JEPA by video quality would miss its intended use. A convincing result would show up in what a downstream system can predict or decide.

Synthetic data for robotics

In robotics, I want to know whether world models can reduce the work of collecting training data. Recording many variations of a task with a physical robot takes equipment and operator time. If a model can produce useful variations, a developer could train or test a policy in situations that would otherwise be expensive to record.

NVIDIA’s Cosmos platform addresses that part of development. Its January 2026 announcements included Predict 2.5 and Transfer 2.5 for generating physical-world scenarios and supporting robot-policy evaluation. NVIDIA’s January announcement

The question for me is which variations help once the robot returns to the real task. Changing the lighting could help it recognize an object under conditions it will encounter in a warehouse. A convincing video of a grasp is less useful if the model gets the contact wrong and the real object slips. I want to see the robot’s behavior improve, not just a larger collection of generated footage.

What I’m following next

I came to these releases asking what another company could build around them, so the most useful next evidence would come from companies that try. A developer reusing a generated scene would tell me something different from a robotics team improving a policy with synthetic data. I’m keeping both in view, because “world models” may turn out to support several kinds of products with quite different requirements.