The term world model has become one of the most frequently used phrases in AI, yet it often means very different things depending on who is using it. To understand why it matters, it helps to look at how the idea has evolved.
Robotics began with the concept of state
For decades, robotics relied on an internal notion of state. Every robot maintained an estimate of itself: joint positions, velocities, localization, sensor readings, and often a representation of its surroundings. Together, the robot state and environment state provided the information needed for planning and control.
This representation was never intended to be intelligent by itself. It simply described the robot’s best estimate of reality so that planners and controllers could decide what to do next. The reasoning largely happened outside this representation, inside planners, controllers, and optimization algorithms.
LLMs showed that impressive reasoning was possible without persistent state
Large language models took a very different approach. They demonstrated remarkable reasoning despite having almost no explicit notion of a continuously evolving world. Every interaction was largely confined to a context window, while longer-term memory lived outside the model through retrieval systems or application logic.
That abstraction works surprisingly well for language because most conversations do not require maintaining a precise, evolving representation of the physical world. Robots face a fundamentally different problem. They cannot treat every camera frame as an independent prompt because the world continues to evolve whether or not the robot is looking.
Physical AI is bringing these two traditions back together
Objects move behind occlusions. People change direction. Forklifts disappear behind shelves before reappearing a few seconds later. Every action changes the set of actions that remain possible in the future.
Physical AI therefore requires continuity. A robot needs an internal representation that evolves with the world rather than restarting from scratch every few hundred milliseconds.
The idea of a world model has evolved beyond representing the world
The progression is subtle but important. Early robotics focused on robot state: Where am I? It then expanded to world state: What objects exist around me, and where are they?
The emerging generation requires something richer: a true world model that asks What is happening? Why is it happening? What is likely to happen next?
This is a fundamentally different role. The world model is no longer just memory or a snapshot of the environment. It becomes the workspace in which reasoning takes place, allowing the robot to connect observations over time, understand physical constraints, anticipate future events, and evaluate the consequences of its actions.
Not everyone means the same thing when they talk about world models
Today, the phrase has become overloaded. For some researchers, a world model is primarily a generative model that predicts future observations. For others, it is an internal simulator that evaluates multiple candidate futures before selecting an action.
Both are valuable interpretations, but I think there is another perspective that will become increasingly important as Physical AI matures.
A world model should become the substrate for reasoning
Rather than existing only to generate future images or simulate hypothetical trajectories, a world model should continuously represent both the robot and its environment through geometry, physics, causality, and temporal relationships.
Its purpose goes beyond remembering the world: it organizes it into a representation that supports reasoning
That internal model becomes the place where the robot asks questions, evaluates alternatives, predicts outcomes, understands constraints, and selects actions that remain safe as the environment changes. In that sense, the world model becomes more than another module within the autonomy stack. It is the medium through which the robot thinks about the physical world.
The next generation of AI may be defined by the world models it builds
For years, progress in AI has largely been measured through larger datasets, larger context windows, and larger models. Physical AI introduces a different question.
Instead of asking how much information a model can process, we should ask what kind of world it constructs internally, how faithfully it maintains that representation over time, and how effectively it reasons within it.
That is the shift ahead: from models that respond to information to models that reason over an evolving understanding of the physical world. It may become one of the defining ideas of the next generation of AI.
