Post

Knowledge as a World Model

Reframes knowledge for this workspace as the revisable world model constructed by a learner.

Knowledge as a World Model

Chinese version: /zh-hant/knowledge/knowledge-as-world-model/

Status: Workspace working model revised on 2026-09-20 at the learner’s request. This is a cognitive and learning framing, not a final solution to every epistemological question and not evidence of independent learner performance.

Core correction

For this workspace, a person’s knowledge is the revisable world model that their brain constructs: an organised network of representations about objects, categories, relations, causes, conditions, procedures, and likely consequences.

The model is useful when it helps the person explain what is happening, predict what may happen under stated conditions, choose an action, notice a mismatch, and revise its assumptions. It is not a literal copy of the world. It is selective, incomplete, and sometimes wrong.

What this changes

An external book, a note in this repository, a video, a number, or an LLM response is not automatically the learner’s knowledge. It is a potential model resource. It becomes part of the learner’s knowledge only to the extent that the learner has constructed, retained, retrieved, and can use the relevant representation or relation in their own world model.

Likewise, an isolated fact is not useless, but it is not the whole target. A fact becomes more valuable when the learner can place it in relations: what it refers to, when it applies, what changes it, what it predicts, and which action it informs.

A compact map

TermRole in this workspace
WorldEvents, objects, people, and processes independent of the learner’s current representation of them.
RepresentationA partial encoding of some aspect of the world: a word, image, measurement, example, rule, or memory.
World modelThe learner’s organised, internal set of representations and relations used to make sense of and act in the world.
KnowledgeThe relatively dependable, usable parts and organisation of that world model.
Information resourceAn external representation that may update the model but is not yet the learner’s knowledge merely by being available.
UnderstandingDemonstrated ability to use and revise the relevant part of the model for explanation, prediction, application, and transfer.
Skill / masteryReliable performance using the model under the required conditions; mastery additionally needs stability, independence, flexibility, and error correction.

Example: a route delay

The world contains a real service disruption. An app alert saying “12-minute delay” is an external representation. The learner’s world model may include the relation: “When a substantial disruption occurs on Route A during this commute, Route A’s usual reliability is no longer a good guide; compare alternatives.”

That relation is knowledge only as far as it is sufficiently accurate for its conditions and the learner can retrieve and use it. Predicting a late arrival, switching routes, and correcting the relation when new evidence disagrees demonstrates understanding of that part of the model. One lucky switch does not demonstrate mastery.

Accuracy and revision

Calling knowledge a world model does not make every belief correct. The learner may hold a model that is incomplete, poorly calibrated, or false. Therefore the workspace continues to ask:

  1. What does the model claim about the world?
  2. Under what conditions should it apply?
  3. What observation, source, or performance supports it?
  4. What prediction or action follows from it?
  5. What result would require the learner to revise it?

This preserves the earlier concern with support, truth, uncertainty, and conditions while locating learning in the learner’s internal model-building process.

Learning implication

Do not aim merely to collect notes or repeat definitions. Learning changes a world model through targeted revision and reconstruction: identify the narrow or failed relation, take in relevant information, integrate the update with what still holds, and test the revised model.

Aim to build and test a relation in the world model:

  1. Identify the part of the world or task being modelled.
  2. State the objects, variables, relation, mechanism, or procedure.
  3. Name conditions and limits.
  4. Predict a changed case before checking the answer.
  5. Compare the result with evidence and revise the model if needed.

Do not treat revision as deleting everything previously believed. Preserve relations that remain accurate, narrow claims whose scope was too broad, and replace only the assumption, variable, mechanism, or boundary condition that failed. See Learning as Model Revision and Reconstruction for the full cycle.

The smallest next evidence is a no-notes explanation of this distinction using one original everyday example, including what is in the world, what is an external representation, what relation is in the learner’s model, and what would revise it.

This post is licensed under CC BY 4.0 by the author.