Large Universe Model glossary

The three generations

Large Language Model (LLM). A model trained on a corpus collected up to a fixed date and then frozen. It holds the shape of human language and the knowledge carried inside it. Its horizon is its training cutoff. The first of the three generations.

Large World Model (LWM). A model that learns from sensed experience — video, space, physics, embodiment — and predicts what follows inside a scene. It holds a simulation of consequence. Its horizon is the scene in front of it. The second generation, successor to the Large Language Model.

Large Universe Model (LUM). A model that ingests observation continuously and maintains a revisable belief about the domain it observes. Its horizon is now. The third and final generation, direct successor to the Large World Model. Plural: Large Universe Models, or LUMs.

Core concepts

Belief. A claim a Large Universe Model currently holds, carrying a confidence, a set of supporting observations, and a history of revision. The unit of memory in a Large Universe Model, as the parameter is the unit of memory in a Large Language Model.

Belief state. The complete set of beliefs a Large Universe Model maintains about a domain. Persistent like parameters, current like context, and inspectable like neither.

Posterior. The distribution over possible values of a quantity after evidence has been taken into account. What a Large Universe Model holds instead of a point estimate.

Prior. The belief held before a given observation arrives. Every posterior becomes the prior for the next observation.

Revision. The movement of a belief in response to evidence. The primary output of a Large Universe Model, and the thing neither a Large Language Model nor a Large World Model can produce, since neither holds a belief long enough for it to move.

Surprise. A revision large relative to the confidence that preceded it. Operationally the most valuable signal a Large Universe Model emits.

Evidence weighting. Scaling each observation's influence by source reliability, precision and recency, so that an audited filing moves a belief further than an unsourced report.

Decay. The deliberate widening of a belief over time in the absence of confirming evidence. A six-month-old estimate should be less confident than it was, without anyone marking it stale.

Provenance. The recoverable chain from a belief back to the specific observations that produced it, in order, with weights and timestamps.

Continuous ingestion. Taking in evidence without a stopping point. The defining property of the Large Universe Model.

Training cutoff. The date after which a frozen model has no knowledge. The defining limitation of the Large Language Model, inherited by the Large World Model and removed by the Large Universe Model.

Presence. Continuity of attention — being with a situation as it changes rather than reasoning about a snapshot of it. What the Large Universe Model adds to the knowledge of the Large Language Model and the consequence of the Large World Model.

Adjacent techniques

Retrieval-augmented generation (RAG). Fetching documents at query time and supplying them to a Large Language Model. A pull mechanism, and a workaround for the training cutoff rather than a removal of it. See Large Universe Model vs RAG.

Fine-tuning. Continuing training on additional data to produce a new frozen model. Moves the cutoff forward; does not remove it. See Large Universe Model vs fine-tuning.

Online learning. Updating a model incrementally as data arrives. Continuous, but requires a hand-specified state space and observation model per problem. See Large Universe Model vs online learning.

AI agent. A model given tools and a goal, which acts until the task completes and then terminates. Episodic where a Large Universe Model is continuous. See Large Universe Model vs AI agents.

Loss of plasticity. The documented degradation by which networks trained continually progressively lose the ability to learn from new examples. The central research obstacle for any system that ingests forever.

Nowcasting. Estimating the present state of a system from incomplete and lagged data, revising as backfill arrives. A continuous-revision problem, and one of the oldest applied instances of the Large Universe Model pattern.

Alert broker. In astronomy, a system that ingests a survey's full detection stream, classifies each event, revises as more data arrives and escalates the interesting ones. A Large Universe Model with a domain restriction, in production for years.