World Models Will Soon Be Inescapable. Here’s What You Need to Know

World models are the attempt to give machines a sense of physical space — software that takes a visual state of the world and tries to predict what happens next, rather than the next word. Niantic, after selling Pokémon Go last year, has pivoted into that work. Brian McClendon, Niantic Spatial's chief technology officer, says the company wants to reconstruct the real world and use its trove of data to build replicas. Wayve is feeding dashcam footage into a generative model so cars can rehearse hazardous roads impractical to stage in real traffic.

The timing is the point. WIRED's Reece Rogers writes that 2027 could be a major year for the expansion of home robotics and driverless cars, both bolstered by these advances. The software is poised to seep into everyday life without people necessarily interacting with a world model directly. If the pitch holds, a robot practices inside a digital copy of a room before it ever moves through it, and a car practices the crash it should never cause.

Investors are already treating this as a founder-name race. Yann LeCun, Meta's former chief AI scientist, raised over a billion dollars recently for his Advanced Machines Intelligence startup. Fei-Fei Li launched World Labs in 2024; AMD acquired it this year for over $8 billion. Large companies like Google will keep boasting about improvements. Tencent is investing in world models for games, where physics and 3D rendering are already the job.

We're for the useful version of this. A robot that can run "millions of simulations" from photographs of the space it will work in is a real gain — practice without breaking objects or putting a person in the way. Recreating unsafe conditions from dashcam footage, instead of waiting to meet them on a public road, is safety work autonomous vehicles have owed people for years. Jamie Shotton, chief scientist at Wayve, sets the right bar: a world model has to react to the actions we take, understand physics, cause and effect, and how other objects and people change in response.

I'm watching the data underside just as closely. To simulate the physical world, labs need people to record it. Earlier this year Rogers strapped an iPhone to his forehead, filmed hours of life in his apartment — folding laundry, washing dishes, pouring Diet Coke — and sold the video through an app for a fairly measly payout. As more companies try to build better world models, more of that gig work is likely to show up. Homes and chores become training material. The payout does not have to be fair for the footage to be valuable.

The privacy problem is quieter than a launch event and harder to refuse. Hundreds of millions of people played Pokémon Go in 2016 without realizing those environmental scans might one day help robots navigate obstacles and interact with objects. Most of those players were not signing up to train a household machine. World models do not need a conversation. They can sit underneath a robot, a car, or a game and never introduce themselves.

Oliver Cameron, CEO of the world-model lab Odyssey, calls those 3D generations digital training grounds — simulations meant to pass as the real thing. Useful, if the practice stays a rehearsal for safer machines.

Where this leads if it scales is more of physical life recorded so bots can practice being in it. The labor market for filming ordinary tasks expands. The upside is robots and cars that fail in simulation instead of in a kitchen or an intersection. The downside if we sleep: replicas built from underpaid footage of private rooms, and a public that only notices when the machine finally arrives to use the map.

Unlike text-based language models, world models take visual inputs and generate digital environments that simulate physics. McClendon's training loop is concrete: photograph the place a robot is meant to function, generate a 3D model, then let the software run those simulations so the machine learns the space's parameters without physically moving through it.