Most AI tools you have used predict the next word. Type a sentence into a chatbot and it guesses what should come after, one token at a time. A newer kind of model is being built to predict something else entirely: the next moment. Not the next word in a sentence, but what a room looks like if you turn left inside it, or what happens half a second after a ball leaves your hand. That is the idea behind a world model, and a handful of labs now think it matters more than the next chatbot upgrade.
What a “World Model” Really Means
The term can sound like marketing, so it helps to start with how the people building these systems define it. Google DeepMind describes Genie 3 as “a general purpose world model that can generate an unprecedented diversity of interactive environments,” built on the idea that a model can use its understanding of the world to simulate parts of it, so an AI agent can predict how a space will change and how its own actions affect that space.
World Labs, a separate company founded around the same idea, frames it almost identically in its own announcement of a product called Marble: “World models should reconstruct, generate, and simulate 3D worlds; and allow both humans and agents to interact with them.” Two different teams landed on the same basic claim: a useful world model does not just describe a scene in words, it lets something, human or AI, move through that scene and see consequences.
This is a different job than the one large language models are trained for. A language model’s whole task is predicting the next token in a stream of text. A world model’s task is predicting the next state of an environment, which is why the two are often described as complementary rather than competing approaches.
Genie 3: Real-Time Worlds From a Text Prompt
DeepMind’s Genie 3, announced on August 5, 2025, generates an interactive 3D-feeling environment that a user or an AI agent can navigate in real time, at 24 frames per second and 720p resolution, with the generated world staying consistent for a few minutes before drifting. That consistency window is the hard part: earlier video-generating models could produce footage that looked like a place, but they could not reliably let you walk back to a spot you had already visited and find it unchanged.
DeepMind ties the project directly to training other AI systems, not just entertainment. The company has long used simulated environments to train game-playing and robotics agents, and it describes world models as a way to give an agent “an unlimited curriculum of rich simulation environments” instead of a fixed set of hand-built training levels. Genie 3 is presented as a step along that path, following earlier, more limited versions the company calls Genie 1 and Genie 2.
Marble: Worlds You Can Export and Edit
World Labs, started by AI researcher Fei-Fei Li, took a related but distinct approach with Marble, which went generally available on November 12, 2025. Rather than generating a world purely from a text prompt, Marble accepts text, a single image, multiple images, video of a real location, or a rough 3D layout sketched with blocks and planes, then produces a 3D world you can export as a point-cloud-style Gaussian splat, a textured mesh, or video. World Labs lists use cases including gaming, visual effects, design, and robotics simulation.
The practical difference from Genie 3 is persistence and portability. Genie 3’s worlds exist as a real-time generated stream you explore in the moment. Marble’s worlds are meant to be saved, edited, and dropped into other software, closer to a 3D asset than a live simulation. Neither approach has fully replaced the other, and both companies have kept shipping newer models since these first releases, so treat any specific version as a snapshot of a fast-moving field rather than a final answer.
Why This Matters Beyond Demos
The most concrete reason researchers care about world models is training data for robots and other physical agents. Teaching a robot arm to handle a new object by trial and error in the real world is slow and can damage equipment. A world model that accurately simulates how objects fall, collide, and respond to force lets an agent practice far more attempts than any physical lab could run, before any of it touches real hardware. This is the same logic behind existing game-playing AI agents, just extended to open-ended, generated environments instead of a fixed game.
Game and film production is the other obvious use. Blocking out a 3D level or location normally takes artists and engineers real time. A model that can rough out an explorable space from a sketch or a reference photo does not replace that work, but it can shorten the earliest, most exploratory stage of it.
What’s Still Missing
Both companies are upfront that this is early. Genie 3’s environments drift and lose consistency after a few minutes rather than staying stable indefinitely, so it is not yet a substitute for a hand-built game engine where a room looks the same no matter how long you stay in it. Judging how good any of these systems actually are is its own open problem, the same challenge covered in our explainer on how AI models are compared and ranked, since there is no single agreed test for “does this world behave correctly.”
It is also worth keeping expectations calibrated against the hype. Headlines sometimes describe world models as a near-term path to general intelligence. DeepMind and World Labs make a narrower claim: that simulating physical environments is a missing piece next to language and image models, not a finished replacement for either.
FAQs
Is a world model the same thing as a video generator?
No. A video generator produces a clip you watch from start to finish. A world model produces a space you or an AI agent can move through and interact with as it is being generated, and it has to keep track of what you have already seen so the space stays consistent as you move around.
Can I try Genie 3 or Marble myself?
Marble is available directly from World Labs as a product. Genie 3’s broader access has been more limited and research-focused, so check each company’s current site for what access looks like now, since availability for tools this new tends to change quickly.
Do world models replace large language models?
No, they solve a different problem. Language models predict text. World models simulate environments and physical consequences. Current AI products increasingly combine several kinds of models rather than relying on just one.
Why do companies want AI to simulate physical environments?
The biggest reason is training other AI systems, especially robots, cheaply and safely. Practicing a task thousands of times in a realistic simulation is far faster and less risky than practicing it the same number of times with real hardware.
How long can a generated world stay consistent?
For Genie 3, DeepMind states consistency holds for a few minutes before the environment can start to drift. That is a meaningful improvement over earlier systems, but it is still well short of a traditional game engine’s indefinite consistency.












Discussion about this post