In short
WIRED says world models are moving from AI theory into practical systems for robotics, driverless cars and games, with 2027 emerging as a likely breakout year. The shift could reshape data collection, simulation and even gig work tied to real-world recording.
- World models are shifting from research buzzword to practical AI infrastructure.
- Niantic, Wayve, Odyssey, Tencent and Google are among the companies pushing the field.
- WIRED says 2027 could be a major year for home robotics and driverless cars.
- The technology depends on real-world video and spatial data, not just text.
- The rise of world models could expand demand for human data collection and recording work.
World models are shifting from a niche AI concept into practical infrastructure for robots, autonomous vehicles and games, with WIRED reporting that 2027 could be the breakout year for home robotics and driverless cars built on the technology. The change matters because the systems are moving beyond text and into software that learns from visual reality, reconstructs environments and predicts what happens next.
What are world models?
World models are AI systems that learn from visual and spatial data rather than only text. In WIRED’s framing, they are built to generate simulated environments with physics, giving machines a way to understand how the world changes when something moves, falls, turns, or collides.
That makes them different from the large language models that dominate public conversation. Instead of predicting the next word, they are designed to predict the next visual state from an image or a scene. In practice, that could help a robot plan a movement, help a self-driving system anticipate a hazard, or help a game engine produce a more responsive digital world.
How are they different from chatbots?
They are different because their input is not primarily text. Oliver Cameron, CEO of Odyssey, described the core idea this way:
“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.”
That visual prediction task is why the technology is attracting attention from companies working on machines that must act in physical space rather than just answer questions on a screen.
Why it matters
The significance goes beyond a technical upgrade. If world models become a standard layer in robotics and transportation, they could shape how machines learn, how much data companies collect, and which parts of everyday life are affected first. WIRED says the technology may spread into daily life without direct user interaction, which means many people could benefit from it before they ever hear the phrase.
The most immediate implications appear to be in three areas:
- Robotics: machines could train in millions of simulations based on photos of the environments they are meant to operate in.
- Autonomous driving: systems could be tested on dangerous roads and unusual conditions without putting vehicles and people at risk.
- Games and entertainment: studios and platform builders could use the models to create more realistic simulated worlds.
There is also a labor angle. WIRED notes that world-model data collection could expand gig work for physical-environment recording. If that happens, the demand for human-recorded real-world video and location data may become part of the model-building economy, not just an engineering detail.
Why are companies racing toward them now?
The commercial case for world models is becoming clearer as AI labs, vehicle companies and game developers look for systems that can reason about the physical world. Robotics and autonomous driving have always depended on simulation because real-world testing is costly, slow or unsafe. World models promise a better way to build those simulations from real scenes instead of from hand-built approximations.
WIRED points to a widening field of players. Niantic, best known for Pokémon Go, has pivoted after selling the game last year and is now pursuing world models through Niantic Spatial. Brian McClendon, the company’s chief technology officer, said the aim is to rebuild the world from the data Niantic already has.
“We’re very much about reconstructing the real world.”
That line captures the strategic logic. Niantic’s old business was built on mapping how people move through physical places. Its new direction treats those maps and location traces as the raw material for AI systems that need to understand the world in three dimensions.
How did the market get here?
The road to world models runs through several years of progress in AI, simulation and sensor-rich consumer products. WIRED traces an early public landmark back to 2016, when Pokémon Go turned mobile phones into a mass mixed-reality interface and, indirectly, showed how valuable environment data could be at scale.
Since then, AI systems have become much better at learning from images, video and multi-modal input. That has made it more plausible to build models that do not just label a scene, but simulate what would happen next. In parallel, self-driving companies and robotics groups have continued to rely on synthetic environments because real-world trial and error can be expensive and dangerous.
Fei-Fei Li’s World Labs is part of the same broad push. WIRED says she launched the company in 2024, and AMD acquired it this year for more than $8 billion. Separately, Yann LeCun recently raised over $1 billion for Advanced Machines Intelligence, another sign that major funding is flowing into AI efforts that focus on understanding the physical world.
Who is building world models?
The emerging field now includes a mix of startups, established AI companies and major platform players. WIRED says Tencent is investing in world models for video game uses, while Google is expected to keep promoting advances in the area. That suggests the technology will be shaped by both consumer entertainment and industrial use cases.
Odyssey, where Oliver Cameron is chief executive, is one of the companies framing the technology in product terms. Cameron’s description of the model as a system that predicts the next visual state from an image gives a sense of how developers see the field: less as a research curiosity and more as an engine for interactive software.
Wayve is taking a slightly different route, using dashcam data to recreate dangerous roads and other hard-to-test driving conditions. For a driverless-car company, the point is not simply to show a machine pictures of roads; it is to teach the system how the road changes in response to its own actions and to the behavior of other drivers.
Jamie Shotton, Wayve’s chief scientist, said driving models need more than pattern recognition. He told WIRED that they must incorporate the deeper rules of the road and the environment itself.
“it must understand physics, cause and effect”
That requirement gets to the heart of why world models are attracting so much attention. A car on a busy street or a robot in a home cannot just identify objects. It needs to infer what happens next if a person steps out, a chair blocks a doorway, or a ball rolls toward a curb.
By the numbers
| Number | What it refers to | Why it matters |
|---|---|---|
| 2016 | The year Pokémon Go became a mass mixed-reality example | Shows how long location and environment data have been building value |
| 2024 | The year Fei-Fei Li launched World Labs | Marks the recent acceleration of the field |
| $1 billion+ | Funds raised for Yann LeCun’s Advanced Machines Intelligence startup | Signals serious investor backing for physical-world AI |
| $8 billion+ | Reported AMD acquisition price for World Labs | Shows how highly the market may value world-model technology |
| 2027 | The year WIRED says may be major for home robotics and driverless cars | Points to the likely next stage of consumer impact |
| millions | Simulations McClendon says a robot could train on | Highlights the scale of synthetic training now being imagined |
What makes the training data so important?
World models depend on data that captures how the real world looks and changes. WIRED says AI labs need human-recorded real-world video to build better systems, which is why the quality and breadth of that video matters so much. The goal is not simply to gather more footage, but to capture environments rich enough to teach physics, movement and cause-and-effect.
McClendon says robots could be trained in millions of simulations based on photos of target environments. That is an important clue to how the technology may scale: a company might photograph a warehouse, a room or a street, then use those images as the starting point for synthetic training runs that would be impossible or impractical to do in the real world.
The same logic applies to self-driving systems. Wayve’s use of dashcam data to recreate difficult driving scenarios suggests that the real value lies in edge cases: the rain-slick curve, the blind corner, the illegal U-turn, the sudden pedestrian movement. Those are the situations that matter most when a machine must react safely.
Background: how the field evolved
World models are not a brand-new idea, but they are newly visible because the rest of AI has moved closer to making them useful. In broad public terms, the field has grown out of a few overlapping trends: better computer vision, stronger generative models, cheaper storage for video, and a rising need for simulations in robotics and transport.
Pokémon Go is an unusual but useful marker in that history. The game used mobile augmented reality and map-like environment data at global scale, making it one of the clearest consumer examples of a product that depended on the physical world being digitally represented. Niantic’s later pivot, after selling the game last year, shows how those assets can be repurposed for a different AI era.
Now, the same broad category of data is being redirected toward machine learning systems that need to understand spaces, not just images. That shift is helping world models move from research discussion to product strategy.
How are the big players positioning themselves?
Different companies are approaching the problem from different angles, but the common thread is that each wants a model that can reason about the real world more faithfully than a text-only system.
- Niantic Spatial is focused on reconstructing the real world from location data.
- Odyssey is framing world models as systems that predict the next visual state from an image.
- Wayve is using driving data to create more realistic test conditions.
- Tencent is investing for game-related use cases.
- Google is expected to keep pushing the category forward.
Those approaches reflect different markets, but they all point to the same underlying need: AI that can model motion, context and physical change.
What happens next?
The next phase will likely be determined by three questions: which companies commercialize world models first at scale, how much data they need, and whether the first visible consumer effects show up in robots, cars or games. WIRED says 2027 is likely to be the big expansion year for home robotics and driverless cars, which gives the market a rough timetable to watch.
Another key issue is data sourcing. If AI labs continue to need human-recorded video of the physical world, that could create new demand for contractors, gig workers and specialized recording services. It could also raise practical and regulatory questions about consent, privacy and the collection of spatial data in homes, streets and workplaces.
For now, the clearest sign of momentum is that the field is no longer confined to academic discussion. Major funding, major acquisitions and direct product pivots are all pointing in the same direction. World models are becoming part of the infrastructure story behind robotics, autonomous vehicles and interactive digital worlds.
What to watch over the next two years
- Commercial launches: whether Niantic Spatial, Odyssey or similar companies ship products that use world models in visible ways.
- Driverless-car testing: whether companies like Wayve show measurable gains from physics-aware simulation.
- Robotics training: whether the “millions of simulations” approach becomes common in home and warehouse robots.
- Gaming use cases: whether Tencent and others use the technology to build richer simulated environments.
- Data rules: whether regulators begin paying closer attention to the collection of real-world video and spatial data.
If WIRED’s timeline proves accurate, the next major wave of AI headlines may not come from chat interfaces at all. It may come from systems quietly learning how rooms, roads and physical objects behave, then using that knowledge to power machines that operate in the real world.
Frequently asked questions
What are world models in AI?
World models are AI systems that learn from visual and spatial data to simulate how the physical world behaves. Instead of predicting text, they aim to predict what happens next in a scene, which makes them useful for robots, self-driving cars and games.
Why are world models important now?
World models are important now because they are moving from a research idea to practical infrastructure for machines that operate in the real world. WIRED says the technology could spread into daily life without users noticing, especially as robotics and autonomous driving mature.
Which companies are working on world models?
Niantic Spatial, Odyssey, Wayve, Tencent and Google are all part of the story WIRED describes. The field also includes Fei-Fei Li’s World Labs and Yann LeCun’s Advanced Machines Intelligence startup.
When could world models become mainstream?
WIRED points to 2027 as a likely breakout year, especially for home robotics and driverless cars. That does not mean the technology will be everywhere by then, but it suggests major commercial expansion could arrive around that time.
What data do world models need?
World models need human-recorded real-world video and other spatial data to learn physics, cause and effect, and how environments change. WIRED also says that demand could expand gig work for people recording physical spaces.
