World Model is the new inflection point for AI.
Rajat's Substack
SubscribeSign in
World Model is the new inflection point for AI.<br>The field is chasing a single world model that can see, build, and act — one system to bind the others together.
Rajat Ghosh<br>Jul 11, 2026
Share
Watch Lionel Messi in the half-second before a pass. He doesn’t look like someone calculating — there isn’t time to calculate. He has already seen it: where each teammate is drifting, where the defenders will close, the gap that is about to open two moves from now. He is running an internal model of the field so faithful that he can predict it, plan inside it, and act on it faster than conscious thought.<br>That inner model is a world model, and some version of it is the most important idea in AI right now. It is also nothing exotic. You have one for your kitchen, your commute, the people you love. We walk around convinced we decide things deliberately, but most of the real work happens underneath — in a compressed internal picture of how our world behaves that we never consciously run. Strip away the mystique and a world model is just that: a representation of a system faithful enough to plan with, control with, and predict from.<br>Thanks for reading Rajat's Substack! Subscribe for free to receive new posts and support my work.
Subscribe
The dream now is to give a machine one of these — and not a modest one. A single model that can see the world as it is, build faithful copies of it, and plan how to act inside it, with every other tool folding into it. One model to rule them all. It is a genuinely thrilling goal, and I think it’s roughly the right one. But anyone who has read the story that phrase comes from remembers the rest of it. The whole drama is not about the power of the ring; it’s about whether anyone can be trusted to carry it — because a thing that powerful fails not when it’s weak, but when the one holding it doesn’t know its own limits.<br>That is the part of the world-model story almost nobody is telling. So let me tell it.<br>What a world model is, abstractly
Let me be precise about the thing itself, because the word is slippery. In the abstract, a world model is three things bound together: a state, a rule for how that state changes, and a way to act on it.<br>Begin with the world. A world, in this sense, is any system that has a state — a complete-enough description of how things stand at a given moment — and that changes over time, sometimes in response to what an agent does. What counts as the world is not handed to you; it is an axiom you choose, fixed case by case. It can be the stress profile of a building under load, the state space of a C++ program, the airflow through a turbine, a position on a chessboard, a human conversation. Name the world, and you have named the thing the model must represent. Many of the worlds in this essay are physical, because that is where the stakes are easiest to see — but the argument is about world models as such, whatever world you point them at.<br>A world model, then, is a compressed internal representation of that state, paired with a dynamics — a rule for how the state moves from one moment to the next — and a way to condition that rule on actions, so it can answer not only “what happens next” but “what happens next if I do this.” From those pieces, three capabilities follow, and they are the whole reason to build one:<br>Predict — given the state now, say what the world will do.
Control — given a goal, work back to the actions that reach it.
Simulate — run the world forward in imagination, including down paths that never actually occurred.
And one property separates a real world model from a mere pattern-matcher: faithfulness. The representation has to track the true system closely enough that a plan made inside the model still works when it meets the world itself. The rest of this essay turns on the single demand that bare definition leaves out: whether the model also knows when its own faithfulness runs out.<br>Every world converges
There is a reason small, faithful world models are possible at all, and it deserves its own moment. Dynamical-systems theory calls it an attractor: the small region of all possible states that a system’s behavior converges onto over time, no matter where it starts. A turbulent flow could in principle occupy an astronomical space of configurations; in practice it circles a handful of recurring patterns. A building’s thermal field could in principle be anything; in practice it settles into a family of profiles shaped by its geometry and its loads. A codebase could in principle reach any state its types allow; in practice its execution traces cluster around the paths it was built to take. Left to run, a world does not wander everywhere it could go. It converges.<br>That is, in a sense, what every world model really is: a model of its world’s attractor. Compression is not a lossy compromise — it is the whole point, and it works only because the world itself...