The Brief
How AI Works 4 min read

World Models: The AI Layer Being Built Beneath Everything Else

NAVION

Share

Most coverage of artificial intelligence focuses on chatbots and text generators. World models are something fundamentally different, and they are quietly becoming the infrastructure layer that will power robotics, autonomous vehicles, and possibly much more. Understanding what they are, and how they are already being built, matters for anyone trying to make sense of where AI is actually heading.

From Predicting Words to Predicting Reality

To understand world models, it helps to start with what most people already know: large language models. A language model takes a sentence and predicts the next word. The logic is sequential, text-based, and trained on written human output.

World models operate on a different axis entirely. As Oliver Cameron, CEO of AI lab Odyssey, explains it: you give the system a visual state of the world, an image, and it predicts the next visual state of the world. The output is not text. It is a simulated environment, one that must account for physics, cause and effect, and how objects and people respond to actions taken within it.

This is a harder problem than language prediction. A word follows grammatical and semantic rules that are, at least partially, consistent. The physical world is messier. Objects fall, surfaces have friction, people react unpredictably. A world model that cannot handle these dynamics is not useful for training anything that needs to operate in reality.

The Training Ground That Does Not Exist Yet

Here is what makes world models practically significant: they allow machines to practice in environments that have not been built yet, and to rehearse scenarios that would be dangerous or impractical to stage in the real world.

Brian McClendon, chief technology officer of Niantic Spatial, describes a workflow where someone photographs the space where a robot is intended to operate, generates a 3D model from those images, and then runs millions of simulations to train the robot on that environment’s specific parameters. The robot learns without physically moving. It fails, adjusts, and improves inside a digital replica before it ever touches a real surface.

Wayve, a startup that builds world models designed for autonomous vehicles, uses a similar logic. Its generative world model is powered in part by dashcam footage, which it uses to recreate hazardous roads and unsafe conditions that would not be practical to test in real life. Jamie Shotton, Wayve’s chief scientist, frames the core requirement clearly: a world model must be able to react to actions taken within it, which means it must understand physics, cause and effect, and how other agents change their behavior in response.

Niantic itself is a telling case. The company built Pokémon Go, a game that had hundreds of millions of players scanning their physical environments through their smartphones. Those environmental scans, collected over years of gameplay, represent a substantial trove of real-world spatial data. Niantic has since sold the game and pivoted toward world models, with McClendon stating directly that the company’s focus is on reconstructing the real world. The data collected through a consumer game is now being repurposed as training material for AI systems designed to navigate physical space.

The Labor Market Nobody Is Talking About

The investment side of world models is already substantial. Yann LeCun, Meta’s former chief AI scientist, raised over a billion dollars to fund his Advanced Machines Intelligence startup. Fei-Fei Li, a Stanford computer scientist, launched a startup in 2024 called World Labs, which AMD acquired for over $8 billion. Companies like Google and Tencent are developing their own world model capabilities. The capital flowing into this space signals that major players view world models not as a niche research area but as a foundational technology.

What receives less attention is the labor structure being built to feed these systems. World models need data about the physical world, and that data does not generate itself. It requires people to record ordinary environments: kitchens, living rooms, hands performing routine tasks. This is already producing a new category of gig work, where individuals record hours of video in their own homes and sell that footage through apps, often for modest compensation. As more companies compete to build better world models, demand for this kind of data collection is likely to grow.

This is worth sitting with. The people doing this work are not AI researchers or engineers. They are contributing the raw material that makes physical AI systems possible, often without a clear understanding of how that material will be used or what it will eventually train.

In Short

World models are AI systems that simulate physical reality rather than generate text. They allow robots and autonomous vehicles to train in digital environments before operating in real ones. The technology is attracting significant investment, with major acquisitions and funding rounds already on record. Behind the scenes, it is also creating new forms of data labor, where ordinary people record their physical environments to feed systems they will likely never interact with directly. The gap between who builds these systems and who supplies them is one of the defining tensions of this technology’s early development.

Based on reporting from Wired.

Written by

NAVION