LUM Large Universe Model

Large Universe Model/Essays/Why the third generation looks different

Essay

Why the third generation looks different

The first two generations were a single artefact. The third is not, and the reason is unavoidable rather than aesthetic.

What the first two shared

A Large Language Model is a file of weights. A Large World Model is a file of weights. You can copy either one, run it anywhere, and it behaves identically. Everything it knows is inside it.

This property is enormously convenient and it is precisely what continuous ingestion breaks.

The forcing argument

Suppose you want a model that never stops learning, and you try to do it the obvious way: keep training the weights as evidence arrives.

Three things go wrong. Training is expensive per update, so per-observation revision is uneconomic. Training is lossy and interfering, so new evidence degrades unrelated knowledge. And most fundamentally, continually trained networks lose plasticity — they progressively stop being able to learn, as documented in Loss of plasticity in deep continual learning.

Continuous ingestion into weights does not work. Something else is required.

The separation

The Large Universe Model splits the artefact in two.

Capability — language understanding, reasoning, extraction — stays in weights, changes rarely, and is inherited more or less directly from the Large Language Model and the Large World Model.

Belief — what is currently true about this domain — lives in explicit, external state. Updating it costs nothing, cannot cause forgetting, can be inspected, weighted, decayed and reversed.

The Large Language Model put knowledge in weights. The Large World Model put dynamics in weights. The Large Universe Model took the part that changes and moved it out of the weights entirely — which is the single structural idea of the third generation.

What follows from the split

Nearly everything distinctive follows mechanically.

Provenance is possible because beliefs are built from separable observations. Reversal is possible because evidence can be reweighted. Decay is possible because beliefs carry timestamps. Calibration is measurable because confidence is explicit. Revision is reportable because there is a previous position to compare against.

None of these could be bolted onto a pure-weights model, which is why they were absent from the first two generations rather than merely unimplemented.

The cost of the split

A Large Universe Model is no longer a portable file. It is a running system with state that grows, requires maintenance, and cannot be copied and handed over the way a model checkpoint can.

That is a real loss, and it is the price of continuity.