Interactive · World models
What is a world model in AI?
A world model is an AI system that simulates an interactive environment instead of storing one. It generates what you see and how it reacts to you in real time, predicting the next moment from the moments just before it. The interesting consequence is below, and you can walk through it.
Walk through a world that forgets
Below is a strip of world. You can only see part of it at once, exactly like a screen. Scroll along it and come back, and watch what happens to the tiles you stopped looking at.
Move right, then come back left. With durable state off, the world re-invents itself.
What is actually happening here
With durable state off, each tile is regenerated when it re-enters view, so it comes back plausible but different: right sort of place, wrong specifics. That is the behaviour people report in generative game models, where a building becomes a cliff when you turn around.
With durable state on, the tile is written down the first time and read back afterwards. Nothing about the generator changed. The only difference is that something was stored.
What is the difference between a world model and an LLM?
An LLM predicts the next token of text. A world model predicts the next state of an environment, usually the next frame, given your input. Both are next-step predictors. One produces language, the other produces somewhere you can move through.
| LLM | World model | |
|---|---|---|
| Predicts | The next token | The next frame or state |
| Output | Text | An interactive environment |
| Memory | Context window | Recent frames |
| Forgets when | Context is exceeded | Something leaves the screen |
| Fails as | A confident wrong answer | A confident wrong world |
The failure modes rhyme, which is the useful part. Both produce something fluent and plausible when they have lost track, rather than stopping and telling you they have lost track.
Why do world models forget?
Because they condition on a short window of recent context rather than a stored map. Anything that leaves that window was never written down, so when you look back the model regenerates something plausible instead of recalling what was there.
This is not a bug awaiting a patch. It follows from the architecture. A next-frame predictor works from what it can still see, and a world only persists if something outside the model persists it.
You can watch this happen in a real one. Oasis generates a playable Minecraft-like world with no game engine underneath, and it loses anything you stop looking at. The same limitation shows up from the other side when an AI plays a real game and loses track of its own inventory.
Why this matters outside games
World models are now being pointed at robotics, autonomous vehicles and manufacturing simulation, where an agent has to predict what happens next in a physical space. The persistence problem travels with them.
In an operation the equivalent of a forgotten tile is a stock figure nobody wrote down, an order that exists in one person's memory, or a production change that never reached the schedule. The business keeps running and looks fine. It has simply stopped agreeing with itself about what is true.
The fix is unglamorous and it is the same fix as in the demo above. Write it down once, in one place, and read it back instead of re-deriving it. That is what a bill of materials does for a product and what a safety stock level does for a buffer.
Frequently Asked Questions
What is a world model in AI?
A world model is an AI system that simulates an interactive environment instead of storing one. Rather than loading a saved world, it generates what you see and how it reacts to you in real time, predicting the next moment from the moments before it.
What is the difference between a world model and an LLM?
An LLM predicts the next token of text. A world model predicts the next state of an environment, usually the next frame of video, given your input. Both are next-step predictors, but one produces language and the other produces a place you can move through.
Why do world models forget?
Because they condition on a short window of recent context rather than a stored map. Anything that leaves that window was never saved, so when you look back the model regenerates something plausible instead of recalling what was actually there.
Are world models useful outside games?
They are being aimed at robotics, autonomous vehicles and manufacturing simulation, where an agent needs to predict what happens next in a physical space. The same persistence limitation applies, which is why these systems are paired with durable state rather than trusted alone.
If your operation forgets things, that is the same problem
Most businesses we meet have a version of the demo above running every day: the number is right while someone is looking at it, and quietly re-invented afterwards. We build the system that writes it down once and reads it back.