Wired explainer: world models are heading into robots, cars and games
Wired's explainer opens in 2016, with players hunting Snorlax in Pokémon Go, alongside what the author calls hundreds of millions of other players. The author's point is that scans of those players' surroundings will one day help robots navigate obstacles and interact with objects. Niantic, the creator of Pokémon Go, sold the game last year and pivoted to world models. Brian McClendon, Niantic Spatial's chief technology officer, says: "We’re very much about reconstructing the real world." He hopes to use the company's trove of world data to build replicas.
The article defines the term by contrast. Unlike text-based large language models, world models use visual inputs and can generate digital environments that simulate physics. Oliver Cameron, CEO of Odyssey, an AI lab focused on world models, puts it this way: a language model takes a sentence and predicts the next word, while with a world model you give it a visual state of the world, an image, and it predicts the next visual state. These 3D generations can serve as digital training grounds. "They’re simulations, but they look and feel real," he says.
The author argues the software is poised to seep into everyday life, without people necessarily interacting with world models directly, and says 2027 could be a major year for the expansion of home robotics and driverless cars, both bolstered by advances in world models.
Robots are the first use case. McClendon sees world models letting robots practice without physical movement or being in their intended environments. Someone could photograph the place where a robot is meant to work, generate a 3D model, and let the software run "millions of simulations" to train the robot on that space's parameters. Driverless cars are the second: without a human at the wheel, split-second software reactions are paramount. Jamie Shotton, chief scientist at Wayve, which builds world models designed for cars, says a world model must be able to react to the actions we take, so it must understand physics, cause and effect, and how other objects and people change their actions in response. Wayve's generative world model is powered in part by dashcam footage, and the startup uses it to recreate hazardous roads and other unsafe conditions that would be impractical to test in real life. Video games are a third area: since physics and 3D rendering are core to games, major players like Tencent are investing in their own world models, and developers are analyzing the impact.
Money is following. The article says large AI companies such as Google will keep boasting about world-model improvements, and investors are backing well-known computer scientists who have turned to the field. Yann LeCun, Meta's former chief AI scientist, raised over a billion dollars recently to fund his Advanced Machines Intelligence startup. Famed Stanford computer scientist Fei-Fei Li launched World Labs in 2024, and AMD acquired it this year for over $8 billion.
The article closes on labor. To simulate the physical world, AI labs need real people to record it. The author says that earlier this year they strapped an iPhone to their forehead and recorded hours of video in their apartment: hands folding laundry, washing dishes and pouring Diet Coke. They sold the footage through an app for a fairly measly payout. As more companies try to build better world models, even more gig work may demand this kind of data collection.
Key facts
- World models take visual input and generate digital environments that simulate physics; Odyssey's CEO Oliver Cameron says they predict the next visual state of the world the way a language model predicts the next word.
- Brian McClendon of Niantic Spatial says a robot's workplace could be photographed, turned into a 3D model and run through "millions of simulations" to train the robot.
- Wayve's generative world model is powered in part by dashcam footage and recreates hazardous roads that would be impractical to test in real life; Tencent is among those investing in game-focused world models.
- Yann LeCun's Advanced Machines Intelligence startup raised over a billion dollars recently, and AMD acquired Fei-Fei Li's World Labs this year for over $8 billion.
- The author sold hours of household-chore video through an app for a fairly measly payout, and expects more gig work to demand such data as companies build better world models.
Why it matters
The article frames world models as one of Silicon Valley's buzziest AI topics and a different kind of system from text-based language models: they work from visual input and generate environments that simulate physics. The author argues the technology will reach ordinary people indirectly, through home robots, driverless cars and video games, and says 2027 could be a major year for the first two. The funding named in the piece, over a billion dollars for LeCun's startup and over $8 billion paid by AMD for World Labs, shows how much money is moving into the area.
Who it affects
Robotics and self-driving companies are the main users in the article: Niantic Spatial on robot training, Wayve on cars. Game developers are weighing the technology, with Tencent investing in its own world models. People who record video for AI labs are affected too. The author describes selling footage of household chores and expects more such gig work as companies build better world models. Ordinary consumers may meet world models only through the robots and cars built on them.
How to use it
The piece describes no product to try or buy. It does sketch a workflow from McClendon: photograph the place where a robot will operate, generate a 3D model, then let the software run millions of simulations to train the robot on that space. For individuals, the practical angle is the data side: apps pay gig workers to record everyday tasks on video, and the author's experience was a fairly measly payout.
How solid is it
This is a general explainer written in the first person, not a report on a new release or a study. The technical explanations come from people at companies in the field: Niantic Spatial's chief technology officer, the CEO of Odyssey and the chief scientist at Wayve. The financial figures are given only as lower bounds ("over a billion dollars", "over $8 billion"). The 2027 outlook is the author's own statement of what "could" happen.
Risks and caveats
The article gives no sales or deployment numbers for home robots or driverless cars in 2027, and it describes no technical architecture of any world model. It does not say how much Tencent is investing, and it names no investors in LeCun's raise. The author raises one concern: world models are already changing the labor market behind the scenes, and these systems are ingesting more and more of the author's physical life, with pay for the recorded data described as measly.
“They’re simulations, but they look and feel real”
— Oliver Cameron, CEO of Odyssey, an AI lab focused on world models