On 8 November 2026 Wired published a survey of the state of world models — software that predicts the next image of its surroundings rather than the next word. The piece gathers statements from Niantic Spatial, Odyssey and Wayve, and shows how the technology is already reshaping the labour market around data collection. What is at stake is training home robots and autonomous cars without driving onto a real street.
Key takeaways
- A world model takes a visual state of the world as input and predicts its next visual state — unlike a language model, which predicts the next word
- Niantic, the maker of Pokémon Go, sold the game and pivoted to world models, drawing on its accumulated spatial scans
- Wayve trains a world model on dashcam footage in order to recreate hazardous road conditions that are impractical to test live
- AMD is acquiring World Labs, Fei-Fei Li's startup, for $8.2B — as TechCrunch reported on 28 September 2026
- Wired names 2027 as a pivotal year for home robotics and driverless cars, both driven by progress in world models
Where does this start?
The year is 2016. Hundreds of millions of people are walking their neighbourhoods with a smartphone, hunting a Snorlax in Pokémon Go. Along the way their phones are scanning the surroundings — pavements, kerbs, interiors, objects.
A decade later those same scans become material for building digital replicas of the world, in which robots learn to avoid obstacles and pick up objects. Niantic, the game's creator, sold it in 2025 and pivoted to precisely this technology — world models.
But a problem appears: a model that describes the world beautifully in words does not know what happens when a mug slides off a table. Next-token prediction does not supply physics.
And that is the gap world models are meant to close.
How does a world model differ from a language model?
The difference lies in what the model is given and what it returns.
“So with a language model, you have a sentence, and it will predict the next word in the sentence,” Oliver Cameron, CEO of the Odyssey lab, told Wired. “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.”
Everything else follows from that single change. If the output is an image rather than text, the output can also be an entire generated 3D space. And if it is a space, a robot can be placed in it and told to try.
Cameron describes these generations as “simulations, but they look and feel real”. Brian McClendon, CTO of Niantic Spatial, sets out a concrete use: photograph the place where a robot is meant to operate, generate a 3D model from those images, and run “millions of simulations” to teach the software that space's parameters.
This is a simplified flow reconstructed from the description in Wired, not documentation for a specific product.
| Company | Where its data comes from | What it uses the world model for |
|---|---|---|
| Niantic Spatial | environment scans from Pokémon Go | reconstructing real interiors for robot training |
| Wayve | dashcam footage | recreating hazardous road situations |
| Odyssey | a visual state of the world given as input | generating subsequent states as a training ground |
| World Labs | not publicly disclosed | the commercial product Marble |
What do these names mean?
A few terms recur in discussions of this technology that cannot be guessed from the name.
- World model — a model predicting the next state of the environment from an image and an action taken. The equivalent of a game physics engine, except learned from footage rather than programmed as rules.
- Sim-to-real gap — the divergence between what works in simulation and what works on real hardware. The main reason good physics simulation is a prerequisite rather than an extra.
- Dashcam data — recordings from vehicles in motion, used as training material. Wayve builds its generative world model on them.
- Spatial reconstruction — turning photographs or scans of a real place into a 3D model. This is the stage where datasets gathered by consumer apps become useful.
Taken together, these four elements explain why mobile-game companies and autonomous-vehicle companies suddenly compete over the same technology: all of them need a learned model of physics, and all of them hold data from which one can be built.
Why not train the robot in the real room?
That is the most obvious question: if a robot is to work in a kitchen, let it learn in a kitchen.
The constraint is the number of attempts. McClendon speaks of millions of simulations per room. A physical robot managing one attempt per minute would need years, wearing out gearboxes and batteries along the way. Simulation shifts the cost from mechanics to compute.
The second problem concerns cases that must not be triggered deliberately. Jamie Shotton, chief scientist at Wayve, states the requirement: “a world model must be able to react to the actions that we take in the world, and so it must understand physics, cause and effect, and how other objects and people change their actions in response to our own”. The company recreates hazardous roads and conditions whose live testing would be impractical.
The third item is the cost of data — and here the bill lands on people. Reece Rogers, writing for Wired, describes his own experiment: he strapped an iPhone to his forehead and recorded hours of video in his apartment, capturing laundry folding, dishwashing and pouring a drink, then sold the footage through an app for what he calls a measly payout. The more companies build world models, the more gig work of this kind will be required.
Is it working already?
Partly — the technology is in production at autonomous-vehicle companies, but consumer applications remain a forecast. Wayve has a working generative world model trained on road footage. Home robotics and driverless cars are, per the Wired piece, set to accelerate only in 2027.
Hard confirmation comes from the deal side. TechCrunch reported on 28 September 2026 that AMD would acquire World Labs, Fei-Fei Li's startup, for $8.2B — the same company raised $1B in February 2026. Yann LeCun, Meta's former chief AI scientist, runs his own world-model startup (AMI Labs), and Wired states he raised over a billion dollars for it. Tencent is investing in its own world models on the gaming side.
What is not known: the piece gives no quality figures and no benchmarks. It is a survey of the market and its applications, not a technical comparison of specific models.
Why does this matter?
World models are the first branch of AI that will work for most people with no interface at all. There will be no chat window and no prompt — there will be a car that brakes differently and a robot that grips differently. At the same time the cost of entry shifts from compute to physical data, and someone has to physically record it. That creates a new, poorly paid category of work.
What next?
- Wired names 2027 as the point at which home robotics and driverless cars expand — both segments dependent on progress in world models
- AMD's acquisition of World Labs for $8.2B, announced on 28 September 2026 per TechCrunch, moves world-model technology into a chipmaker's portfolio
- Identified risk: demand for recordings of everyday chores is growing faster than the rates paid for supplying them, which the Wired piece documents first-hand





