A world model is a generative system that keeps an internal state of an environment, so a space stays consistent as you move through it and act on it. Going from prompt to world changes the hard problem from making one convincing frame to keeping every frame in agreement, in real time, while you steer.
- A generated world is judged by whether it stays consistent over time, not by its best frame.
- Persistence is the hard constraint. Google DeepMind says its Genie 3 model keeps environments largely consistent for several minutes, not hours.
- Frame-by-frame generation lets small errors pile up, so controlling drift matters more than raw resolution.
- Once a viewer can steer, faithful response to input beats visual polish.
- There are two routes to an explorable space: learned models that predict what you see, and engines that keep state explicitly in code and data. The second is what you can build and share in a browser today.
- Shared spaces add a stricter rule: every visitor must see the same facts.
What is a world model?
A world model is a system that learns how an environment behaves and uses that knowledge to predict what happens next, including what you would see after you move or act. The idea is older than today's video generators. In 2018, David Ha and Jürgen Schmidhuber described a model that learns "a compressed spatial and temporal representation of the environment" and showed that an agent could be trained "entirely inside of its own dream environment generated by its world model" (World Models).
What changed since then is scale and interactivity. In August 2025, Google DeepMind described Genie 3, a model that generates worlds from a text prompt that a person can navigate in real time at 24 frames per second and 720p resolution. The company says the environments stay largely consistent for several minutes, with visual memory reaching back about one minute, and that text prompts can change a world while you are in it, for example by altering the weather or adding objects (Genie 3 announcement).
The output of such a system is not a picture or a clip. It is a place: something you enter, move through and act on, which has to hold together while you do.
Why generated output is becoming a space
For most of its short history, generative output has been terminal. A prompt resolved into one image or one clip, and the result was complete the moment it appeared. Nothing later depended on what the system had decided three seconds earlier.
An interactive environment removes that luxury. The moment a viewer can change viewpoint, the system owes consistency to every choice it has already made. A wall placed on the left has to stay on the left. A shadow cast at one angle cannot quietly change when the camera pans away and returns.
This is a different kind of problem, not a harder version of the old one. A convincing frame needs local plausibility. A convincing space needs agreement across many frames that no viewer will ever see at once. It is the same demand that makes character and style consistency hard across a series of still images, multiplied by time and interaction.
It also changes what "finished" means. A still is done when it looks right. A space is proven only through use: someone has to walk the corridor, double back and find the door where they left it.
What persistence demands from a generated space
Persistence is the property that a generated environment keeps its established facts when attention moves away and comes back. It sounds modest. In practice it is what separates a navigable world from a sequence of attractive but unrelated views.
Consider the cost of forgetting. If a room rearranges itself each time it leaves the frame, the viewer cannot form a mental map, and the sense of place collapses. Memory is not a feature added on top of generation here. It is what makes the output feel like a world at all, in the same way that agent memory across sessions is what makes an assistant feel like it knows your project.
There is a trade-off between how much state a system carries and how freely it can improvise. A model that remembers everything resists surprise; a model that remembers nothing contradicts itself. The useful region sits in between: established facts are honoured, and unobserved space stays open to invention until someone looks.
Persistence should also be selective. Not every detail deserves the same weight. The exact texture of a distant wall can be regenerated freely; the position of a doorway cannot. A good environment commits hard to the facts a viewer can check and stays loose about the rest, spending its memory where a contradiction would be noticed.
Temporal coherence vs frame quality
Temporal coherence is the agreement of a generated sequence with itself over time: objects keep their identity, motion stays continuous, and physical relationships hold from one moment to the next. A single stunning frame can still belong to an incoherent sequence. The failure modes are specific:
- Identity drift. A face, a sign or an object slowly changes while no one is looking at it directly.
- Physical contradiction. Something falls upward, or passes through the table it rested on a second ago.
- Continuity breaks. The camera turns cleanly, but the world behind it has quietly reset.
These errors are small in any one frame and serious in aggregate. The Genie 3 announcement explains why: when a world is generated frame by frame, each new frame building on the ones before, inaccuracies tend to accumulate over time. Holding coherence is therefore less about any single output and more about stopping accumulated error before a viewer notices it.
People also judge the two kinds of error differently. A slightly blurry texture reads as a limitation. A chair that vanishes reads as a broken world. That is why coherence often matters more to perceived quality than resolution does. The same accumulation problem appears whenever a generator runs repeatedly; our piece on same prompt, different output covers it for text and code.
Why controllability matters more than fidelity
Once a person can steer, the question changes from how good the picture looks to how faithfully the space responds. A beautiful environment that ignores where the viewer goes is less useful than a plainer one that obeys.
Controllability means the space changes in the direction the user intended, and only in that direction. It pulls against fidelity, because every degree of freedom given to the viewer is another constraint the system must satisfy in real time. Genie 3's stated limitations reflect this: the range of actions an agent can perform directly is limited, and interactions between several independent agents remain an open research problem.
The table compares three ways of producing something a person can look at or move through.
| Property | Static image or clip | Learned world model | Engine-built browser world |
|---|---|---|---|
| Where the state lives | Nowhere after rendering | In the model's learned memory | Explicitly, in code and saved data |
| How long facts persist | Not applicable | Minutes, in published systems | As long as the data is kept |
| Content range | Anything you can describe | Open-ended, invented on the fly | What the rules and assets define |
| Response to input | None | Predicted, within a limited set of actions | Exact, defined by the rules |
| Several people at once | Not applicable | An open research problem | Routine, through shared realtime state |
| Main risk | Implausible detail | Accumulated drift | Bugs in the rules |
The rows show a real trade. A learned world model can invent a place nobody designed, but it remembers for minutes and responds within a narrow set of actions. An engine-built world contains only what its rules and assets define, but every fact persists and every input has an exact effect.
Explicit state vs learned state: two ways to build a world
Under the surface, the two routes answer the persistence question differently. A learned world model holds the world implicitly: the layout of a room exists only as patterns in its memory, and returning to the room means predicting it again. An engine holds the world explicitly: the room is a record with coordinates, the door is an object with a position, and returning to the room means reading those records back.
Explicit state is why video games have offered persistent, shared worlds for decades. It is also why they are expensive to make: someone has to define every rule and supply every asset. Learned world models promise to remove much of that authoring work, at the price of the coherence problems described above.
In practice the routes are converging from the other side too. AI now writes much of the engine. You can describe a game's rules in plain language and have an AI build agent write the code that enforces them, with scores, positions and player state saved explicitly. The world is not predicted frame by frame, but the work of building it starts from a prompt. For anyone who wants to ship something people can visit this week, that is the practical route from prompt to world.
How an explorable space holds together
Whichever route you take, a navigable space has to satisfy the same sequence. The order matters because each step protects the next.
- Commit the facts. Decide the layout, the objects, the lighting and the rules that relate them.
- Store them beyond the current view. Facts must outlive the viewpoint so they can be recalled when attention returns.
- Treat input as a stream. Turn each movement or action into a bounded change, not a fresh generation of the whole scene.
- Reconcile every new view. Check it against committed facts and correct drift before it accumulates.
- Leave the unseen open. Resolve unobserved regions only when someone moves to look, so you do not commit detail nobody asked for.
- Audit the transitions. Viewpoint changes, scene cuts and players joining are where contradictions slip in.
- Share one source of truth. When several people are present, every client must read the same facts, not its own version of them.
Most visible failures are failures of bookkeeping, not imagination: a fact was generated well and then lost. Generation is the easy half. Reconciliation is where the difficulty lives, because it runs constantly and gets no credit when it works.
Where explorable generation gets harder
As output becomes space, difficulty moves from making things look right to making them stay right. Three pressure points stand out.
Duration comes first. A world that holds for a minute is a demo; one that holds for an hour of free movement is a different order of problem. Genie 3 was announced with a few minutes of continuous interaction, which shows how far the persistence horizon of learned worlds still has to grow.
Interaction depth is next. Looking is cheap; acting is expensive. When a viewer can alter the space, every change becomes a new fact to remember and reconcile, and the bookkeeping grows with the freedom granted.
Shared presence is the third. A space several people move through at once must stay consistent for all of them, which turns persistence from a private ledger into a public one. This is where explicit state still has a decisive advantage, and why multiplayer is the clearest test of any generated world. Interfaces are going through a similar shift from fixed output to something composed in use, which we cover in generative UI.
Build an explorable world in the browser with WisyLink
WisyLink takes the engine route. You describe the world in a chat, and its build agent writes a browser game or interactive scene and hosts it at its own address, with state kept explicitly in code and saved data rather than predicted frame by frame.
- 2D and 3D in the browser. Games, canvas animation and real-time WebGL scenes such as product viewers or camera fly-throughs.
- Shared presence. Realtime channels for who's online, shared cursors and multiplayer state, with anything that must last saved to the database.
- Persistent facts. Scores, leaderboards and player progress are stored records, so they are still there next week.
- Your own assets. Use your own images, audio and 3D models (.glb, .gltf or .fbx), or ask for art drawn in code as vector graphics. There is no image, video or world-model generation.
The result is a web page anyone can open from a link, not an App Store game. Our guide to making a multiplayer browser game with AI walks through a full example, and the capabilities overview lists what a build can include. To try it, start free, describe the rules and the space in one message, and walk through the result yourself before you share it.
F.A.Q.