What happened
- Wired published on November 8, 2026 a guide to world models, a kind of AI that works from visual inputs and generates digital environments that simulate physics.
- Niantic, maker of Pokémon Go, sold the game last year and pivoted to this field through Niantic Spatial. Its chief technology officer, Brian McClendon, says the aim is to reconstruct the real world.
- Wired anticipates that 2027 could be a strong year for home robots and driverless cars, both supported by these advances.
- The piece warns that the public will use these tools without interacting with them.
Why it matters
A language model predicts the next word. A world model receives an image of the state of the world and predicts the next visual state, according to Oliver Cameron, CEO of Odyssey. That difference helps train machines that act in physical space, because it lets them practice without moving a single motor.
For robotics and autonomous-driving companies, simulating costly or dangerous scenarios reduces dependence on road tests. For workers who record the world to feed these systems, it opens a new kind of gig work.
The number
More than $8 billion. That is what, according to Wired, AMD paid this year for World Labs, the company Fei-Fei Li founded in 2024. In addition, Yann LeCun, Meta’s former chief AI scientist, raised more than a billion dollars for his Advanced Machines Intelligence startup.
Context
McClendon describes one use: someone takes photos of the place where a robot will work, generates a 3D model and lets the software run “millions of simulations” so the machine learns the space’s parameters. Jamie Shotton, chief scientist at Wayve, explains that a world model must react to the actions of whoever acts: understand physics, cause and effect and how others change their behavior. Wayve feeds its model in part with dashcam footage to recreate hazardous roads that could not be tested in real life.
Video games are also being explored, since physics and 3D rendering are central to them. Tencent is investing in its own models. The author recounts that, to add data, he strapped an iPhone to his forehead, recorded his household chores and sold the videos for a low payout.
Mamífero has covered this line of simulators. Nvidia opened four Cosmos 3 models to simulate driving and robots, Physicl raised $10 million for more realistic simulations, Mundo turns photos of a room into a simulable world and Ai2 presented ToyBlocks to study robot failures.
What’s next
- Wired expects large labs to keep announcing improvements to their world models.
- As more companies look for data, the work of recording everyday life for pay will grow.
- It remains to be seen whether training in simulation translates into reliable robots in real homes.
Bottom line
World models are not a product users open, but a layer under robots, cars and games. The question the report leaves is who records the physical world’s data and how they are paid.
Sources
- Wired, “World Models Will Soon Be Inescapable. Here’s What You Need to Know”, November 8, 2026, link.
- Cover image: Kartverket, CC BY 2.0, via Wikimedia Commons
Edited by Rodrigo Cornejo. How we select and verify: who writes these notes.
