Large Language Model → Large World Model → Large Universe Model
The third one was always going to happen.
Not because it is better. Because the axis the first two were moving along has an end, and this is it.
Read the three generations as three answers to a single question: what is this system permitted to observe?
| Generation | Permitted to observe | Fails when |
|---|---|---|
| Large Language Model | A corpus, collected once, frozen at a cutoff | The world moves past the cutoff — silently, without notice |
| Large World Model | Sensed experience, while a scene is present | The scene ends, because nothing carries across the boundary |
| Large Universe Model | Every stream still running, without a stopping point | Expensively, but visibly — a wrong belief is one you can audit |
That is movement along one dimension, and the third position is terminal on it. There is no category of evidence beyond everything, continuously. A model that reasons better is an improvement within the third position, not a fourth position — the way a faster car is not a fourth entry in the sequence walk, ride, drive.
Why argue it from outside AI
Because the argument is not really about machine learning. It is about what happens to any system that must stay correct about a world that will not hold still — and that problem has been solved, independently, many times, by people who had never heard of a transformer.
Control theory settled it in the 1940s: open-loop systems fail wherever disturbances exist. Thermodynamics says correspondence is maintained, never achieved. Biology has never once produced a persistent system that configures itself correctly a single time. Navigation could not compute longitude from any snapshot, however rich, and needed a clock that never stopped. Each of these is the same result, reached from a different direction.
This site works through 226 such concepts, drawn from 16 disciplines, across 677 pages. Each one is a separate route to the same conclusion, and several of them narrow it.
Where to start
Definitions
A Large Universe Model (LUM) ingests observation continuously and maintains revisable beliefs about what it observes, each carrying provenance and a stated decay, rather than encoding knowledge fixed at a training cutoff. When evidence contradicts a standing belief, the contradiction is an event — nobody has to have suspected anything.
This matters because the failures that cost the most are not wrong answers. They are absent questions. A retrieval system answers correctly if asked; the loss occurs because nobody knew there was anything to ask about.
A note on what is not here
No roadmap, no benchmark, no company. Several of these 226 concepts are used to argue against the strong form of the thesis, and those pages were not softened. An argument that only ever wins is not being tested.