World models are moving artificial intelligence from text to 3D space, and they are about to become very hard to ignore. Unlike large language models, which predict the next word in a sentence, a world model takes a visual state of the world — an image — and predicts the next visual state, simulating real-world physics so robots and autonomous cars can learn how to navigate reality before they ever hit the street.
The concept is not new, but the money and momentum behind it are. Wired reports that Niantic, the company behind Pokémon Go, sold the game and pivoted hard into world models through Niantic Spatial, betting that billions of player scans of streets and parks can be turned into digital replicas. Meanwhile, investors are funding startups such as World Labs, founded by Stanford computer scientist Fei-Fei Li and acquired by AMD for more than $8 billion, and Advanced Machines Intelligence, launched by former Meta chief AI scientist Yann LeCun with over a billion dollars raised.
How world models actually work
The idea is simple to describe and hard to build. "So with a language model, you have a sentence, and it will predict the next word in the sentence," says Oliver Cameron, CEO of the AI lab Odyssey. "With a world model, you give it a visual state of the world, an image, and it will then predict the next visual state of the world." Those 3D generations become digital training grounds: simulations that look and feel real.
For robots, that means practice without physical movement. Brian McClendon, chief technology officer at Niantic Spatial, says someone could photograph the space where a robot is meant to work, generate a 3D model of it, then run millions of simulations so the machine learns the layout and its limits. Wayve's chief scientist, Jamie Shotton, adds that a world model must understand physics and cause and effect — how objects and people react to the car's own actions — because split-second decisions matter when no human is behind the wheel.
- Niantic Spatial is rebuilding the real world in 3D from Pokémon Go scan data.
- Wayve uses dashcam footage to recreate hazardous roads for driverless car training, including conditions too dangerous to test in real life.
- Odyssey builds world models that turn images into predicted 3D scenes that look and feel real.
