Agents
World Model
A language model has learned which words follow one another. Whether it has also learned how the things it talks about fit together is an open and seriously contested question. A world model would be an internal representation of objects, causes, and consequences: something with which one could play through what happens before acting. The debate about it is not an academic game: it decides whether today's methods, with more data and compute, lead to reliable action in the physical world or whether something else is needed. The question becomes practical wherever an error spoils more than a text: in robots, vehicles, and plant controls.
An example
A language model knows that “the glass falls off the table and” is usually followed by “breaks.” Whether it understands that glass is brittle, height converts into impact energy, and a carpet changes the outcome is another question, and the cases in which the answer is no are exactly those where unforeseen situations fail. A world model would be an internal representation with which consequences can be played through before acting.
A closer look
What is meant is a maintained, machine-readable picture of one's own reality: which people, companies, projects, and events exist and how they are connected. It is the prerequisite for an AI being able to answer questions about the company rather than about the world in general.
Why it matters
The term is one of the central bets of the industry and at the same time one of the least settled. For investment decisions it should therefore be treated as an open research question, not as a product feature. It becomes practically relevant where systems act in the physical world (robotics, vehicles, plant control), because there an error spoils more than a text. Anyone reading a “world model” in an offering today should ask for the measurement that supports the claim.
Commonly misunderstood
A model of one's own company ages faster than one of the world. Without a maintenance process, after half a year it is a source of convincing false information.
Where this term does its work in the magazine.
1 article uses the term or takes up its principle.