World Models: AI's Next Big Infrastructure Layer
Original: World Models Will Soon Be Inescapable. Here’s What You Need to Know
Why This Matters
World models are becoming the core training layer for robotics and autonomous vehicles.
World models simulate real-world physics using visual inputs, training robots and autonomous vehicles in digital environments. Companies like Niantic, Wayve, and Odyssey are building them from dashcam footage and street-scan data.
While large language models predict the next word in a sentence, world models predict the next visual state of the world—making them far better suited for training robots and self-driving cars. 'You give it a visual state of the world, an image, and it will then predict the next visual state,' says Odyssey CEO Oliver Cameron. These generated 3D environments act as digital training grounds that 'look and feel real,' letting robots run millions of simulations before ever moving in physical space.
Niantic—the Pokémon Go creator that sold the game last year—has pivoted entirely to world models, planning to leverage years of environmental scan data collected passively by players. CTO Brian McClendon says the goal is reconstructing the real world at scale so robots can train in photo-based replicas of their intended spaces.
Autonomous vehicle startup Wayve uses generative world models fed by dashcam footage to recreate hazardous road conditions too dangerous to test physically. Chief scientist Jamie Shotton notes these models must understand physics, cause and effect, and how other agents respond to actions in real time.
Video game companies, including Tencent, are also investing heavily given how central physics and 3D rendering are to games. Major AI labs like Google are expected to accelerate their world model announcements. The tech is positioned to arrive in everyday life through products—home robots, driverless cars—rather than as a consumer-facing interface.