What Are World Models in AI?

To understand why this shift matters, it helps to look at how LLMs actually work.

An LLM learns by reading vast quantities of text—trillions of sentences drawn from books, websites, and articles. Over time, it becomes extremely good at predicting which word is likely to appear next in a sentence.

This is why LLMs can produce such convincing responses. But it also reveals their core limitation: they understand patterns in language, not the physical world.

An LLM knows that the word glass is often followed by shatters. But it does not know what a glass actually is, how much it weighs, or what gravity does to it. In many ways, it is like a librarian who has read every book ever written but has never once stepped outside.

For many digital tasks this limitation doesn’t matter. But when AI is expected to operate in real environments, three challenges quickly emerge:

• Physical hallucinations. Because LLMs generate responses by predicting the most likely next word, they can confidently suggest actions that are physically impossible or unsafe in real-world environments. 
• Lack of embodied intelligence. Humans develop an intuitive understanding of space, motion, and objects from a very early age. A toddler knows that objects continue to exist when hidden behind something else, or that pushing something harder makes it move faster. These forms of “spatial common sense” are largely absent in text-trained AI systems. 
• Data inefficiency. Humans learn how the world works through observation and interaction. LLMs, by contrast, require enormous amounts of written information to approximate concepts that people learn simply by watching and touching their environment.

These limitations have prompted researchers to ask a fundamental question: what if AI learned about the world the way humans do?

The idea behind World Models is relatively simple in principle: instead of learning primarily from text, AI systems should learn by observing and interacting with reality.

A World Model is an AI system that builds a dynamic internal simulation of the environment around it. Rather than relying mainly on written knowledge, these systems are trained on large datasets of video, sensor data, and physical interactions.

The goal is for the AI to develop an internal representation of how the world behaves.

Imagine an AI that doesn’t just read about a warehouse, but develops a kind of “mental map” of how objects move, how light reflects, and how friction affects motion. When given a task, such a system could simulate the consequences of different actions internally before choosing what to do. Instead of predicting the next word, the AI would be predicting the outcome of actions. 

It is important to distinguish between the experimental systems seen today and the longer-term goal of the technology.

Many systems currently described as “world models”—including video-generation research emerging from major AI labs—are best understood as partial world models. These systems are often highly sophisticated video prediction engines. They can generate plausible sequences of events, predicting what the next frame in a video might look like.

However, they typically lack a complete architecture for taking actions, learning from outcomes, and improving through experience.

The longer-term ambition is a Full World Model Architecture—a system capable not only of predicting how environments change, but also of making decisions within those environments.

Several research groups and startups are now exploring this approach. One example is AMI (Advanced Machine Intelligence), a company pursuing a more comprehensive world model architecture.

AMI recently raised $1.03 billion in seed funding to pursue this goal, with backing from investors including Nvidia, Toyota Ventures, Samsung, Bezos Expeditions, Temasek, Mark Cuban, and Dassault Group - an extraordinary level of support for a still-emerging research direction.

One proposed architecture divides this system into five interacting components:

1. The Perception Module – interprets incoming sensory data. 
2. The World Model (Simulator) – predicts the physical consequences of actions. 
3. The Memory Module – stores past experiences and learned behaviours. 
4. The Actor Module – decides which physical action to take next. 
5. The Critic Module – evaluates whether those actions align with the overall objective.

While current AI might be able to generate a video of a glass breaking, a system built on this type of architecture could potentially understand why the glass broke, remember the outcome, and plan a different approach next time.

If LLMs are transforming knowledge work, World Models could reshape the physical economy. Industries that depend on physical processes—manufacturing, logistics, robotics, and transport—have historically been difficult environments for AI systems that rely primarily on text-based training.

World Models could change that:

• Adaptive robotics. Today’s industrial robots often rely on rigid programming and struggle when unexpected changes occur. A robot equipped with a world model could adapt to unfamiliar situations by reasoning about physical cause and effect rather than following a fixed script. 
• Autonomous logistics. From warehouse automation to self-driving delivery systems, spatial awareness and environmental reasoning are essential. World Models could provide a more reliable form of navigation grounded in an understanding of physical environments. 
• Simulation and design. By modelling how components interact and degrade over time, world models could allow engineers to simulate equipment failures or design flaws before physical prototypes are built. 

The first major wave of AI innovation was generative - systems capable of creating text, images, code, and media.

The next phase may be interactive.

Rather than operating solely within a chat interface, future AI systems could interact with the physical world: navigating spaces, manipulating objects, and anticipating the consequences of their actions.

In this paradigm, AI is no longer just a conversational tool. It becomes a system capable of understanding and operating within the environments it inhabits.

The growing interest in World Models marks the early stages of what some researchers describe as the embodied phase of artificial intelligence.

Large Language Models will likely remain our primary interface for language, analysis, and digital workflows. But as AI expands beyond the screen and into physical environments, new architectures will be required.

For organisations working in environments where physics matters more than grammar—in factories, warehouses, and transport networks—this shift could prove significant.