What are world models, and why AI's biggest minds say they're the future beyond ChatGPT

The future of artificial intelligence isn't more ChatGPT. It's teaching machines to actually understand reality.

While tech giants have spent billions scaling up large language models, a quieter revolution is underway. World models—AI systems that grasp physics, space, and cause-and-effect—are emerging as the technology that could finally bridge the gap between clever chatbots and truly intelligent machines. Think self-driving cars that don't panic in snowstorms, surgical robots that understand anatomy, and AI assistants that know you can't walk through walls.

Your cat is smarter than ChatGPT

Let's start with an uncomfortable truth. ChatGPT can write sonnets and explain quantum mechanics, but it would fail spectacularly at being a house cat.

Your tabby navigates window ledges, calculates pouncing trajectories, and understands that knocked-over water glasses create puddles. These are trivial observations for any mammal. For AI? Nearly impossible.

"We can't even reproduce cat intelligence or rat intelligence," Yann LeCun told a room full of AI researchers in Paris recently. LeCun won the Turing Award—basically the Nobel Prize for computer science—for pioneering the neural networks that power today's AI. He knows what he's talking about. "Any house cat can plan very highly complex actions."

This isn't just about party tricks. Watch an AI-generated video long enough and weird things happen. A dog runs behind a couch and its collar vanishes. The couch becomes a sofa. A person's hand sprouts an extra finger.

These glitches reveal something fundamental: current AI systems don't understand three-dimensional space. They predict pixels based on patterns, not reality. They've never knocked a glass off a table or watched it shatter. They don't know what "behind" means in any meaningful way.

What world models actually are

World models are AI systems that build internal representations of how reality works.

Think about learning to drive. You don't memorise every possible road scenario. You develop an intuitive understanding of physics, momentum, and spatial relationships. You know that cars ahead of you might brake suddenly. That ice patches are slippery. That pedestrians can step into the street.

World models aim to give AI this same intuitive grasp. They learn by observing video, processing sensor data, and building abstract representations of objects, scenes, and physical dynamics. Instead of predicting the next word in a sentence, they predict what happens next in the world—how a ball bounces, how water flows, how a robot arm moves through space.

A baseball batter has about 150 milliseconds to decide whether to swing. That's less time than it takes for visual signals to travel from their eyes to their brain and back. They succeed because their brain predicts where the ball will be based on an internal physics model built from thousands of previous pitches. This is exactly what world models try to replicate digitally.

Two ways to build a digital world

Developers are taking two different approaches to world models. Each has its advantages and brutal limitations.

The first method generates environments on the fly. Google's Genie 3, currently in research preview, works this way. You type a text prompt—say, "a volcanic landscape with lava pools"—and it creates a navigable 3D world in real time at 24 frames per second.

You can move through this world, look around, interact with objects. The AI continuously generates new frames based on your actions and its understanding of how volcanic landscapes behave. It's like a video game, except the game designer is an AI that's never seen the level before.

The catch? Computational intensity. Even Google's cutting-edge system maintains consistency for only a few minutes. After that, the simulation starts breaking down. Objects drift, physics gets wonky, details contradict themselves.

The second approach, used by companies like World Labs (founded by AI pioneer Fei-Fei Li), takes a different tack. Their Marble platform converts text, images, or video into persistent 3D models—digital assets with geometry and physics properties that you can download and edit in other software.

This method creates stable environments that don't degrade over time. But it lacks the dynamic responsiveness of real-time generation. It's the difference between exploring a pre-built video game level and having one generated around you as you walk.

Why the smartest people in tech are betting on this

Jensen Huang, Nvidia's CEO, spent a chunk of his CES 2026 keynote talking about world models. Demis Hassabis, who runs Google DeepMind and just won a Nobel Prize, calls them "a key stepping stone on the path to AGI." Meta built an entire platform called Habitat 3 for training robots in simulated worlds. Even Elon Musk's xAI is developing one.

The reason? Language skills alone won't get AI where it needs to go.

Consider autonomous vehicles. Waymo has spent years and billions of dollars teaching cars to drive in San Francisco and Phoenix. But edge cases still trip them up. A child on a bicycle. Construction zones. Weather conditions the system hasn't seen before.

Toronto-based Waabi took a different approach. They built Waabi World, a complete simulation where AI drivers can log millions of virtual miles. When their trucks encounter a situation in real life, they've often experienced something similar ten thousand times in simulation. The company expects its software to pilot actual trucks autonomously by late 2026—a timeline that would have seemed absurdly aggressive just two years ago.

Where language models hit walls

Here's a weird fact that illustrates the problem: an Atari 2600 from 1979 can beat GPT-4 at chess.

The chatbot has absorbed millions of chess games and rulebooks. It can discuss Kasparov's playing style and explain the Sicilian Defense. But it attempts illegal moves and loses track of where pieces are on the board. The Atari wins because it maintains a simple database—a crude world model—tracking every piece's position.

"These bots tend to attempt illegal moves and quickly lose track of the positions of their pieces," notes a recent analysis. "The Atari wins because it keeps the locations of pieces straight using an ancient and humble version of an internal world model: a database."

This limitation extends everywhere. Language models struggle with spatial reasoning, can't reliably plan multi-step actions, and have no intuitive sense of cause and effect. They're brilliant at manipulating symbols but blind to the three-dimensional space those symbols describe.

LeCun puts it bluntly: "We're never going to get to human-level intelligence by just training on text." He points out that a four-year-old processes as much data through vision alone as the largest language models consume in text. Blind children achieve similar cognitive development through touch. The common thread isn't language—it's interaction with physical reality.

Real applications that actually matter

World models aren't just about making better video games, though that'll happen too.

In manufacturing, they could simulate complex industrial processes with thousands of sensors—jet engines, steel mills, chemical factories. Right now there's no technique to build complete, holistic models of these systems. A world model could learn from sensor data and predict how the entire system will behave when one variable changes.

Healthcare offers compelling use cases. Doctors could simulate molecular reactions to test drug efficacy before clinical trials. Surgeons could practice procedures in environments that respond realistically to every cut and movement. The models work at molecular scales as well as macro ones.

Architecture firms could test buildings before laying a single brick. How does sunlight move through the space across seasons? How do people naturally flow through the floor plan? What happens during an earthquake? Right now these questions require expensive physical prototypes or limited computer simulations. World models promise comprehensive answers.

Smart city infrastructure could leverage world models for video analytics that actually understand context. Not just "person detected" but "pedestrian appears to be injured and lying in street during rush hour." That level of comprehension requires understanding physics, human behaviour, and causal relationships.

The data problem nobody's solved

Building world models requires solving a problem that was easier for language models: where do you get the data?

ChatGPT trained on essentially the entire internet. Text is everywhere, neatly organised in databases and web pages. World models need something different—high-quality video, sensor readings, spatial information, 3D representations. That data isn't sitting around waiting to be scraped.

Encord, which runs one of the largest open-source world model datasets, has assembled 1 billion data pairs across images, videos, text, audio, and 3D point clouds. Company president Ulrik Stig Hansen calls this "just a baseline." Production systems will need orders of magnitude more.

This creates obvious problems. A world model trained mostly on sunny European cities might fail spectacularly in snowy Seoul. Or worse, it might confidently generate incorrect representations of Korean urban environments. The bias issues that plague language models could be even more dangerous when they control physical robots or autonomous vehicles.

"Training data for a world model must be broad enough to cover a diverse set of scenarios," explains Alex Mashrabov, formerly Snap's AI chief and now CEO of Higgsfield. "But also highly specific so the AI can deeply understand the nuances of those scenarios."

The contrarian leaving Meta

LeCun spent over a decade at Meta as chief AI scientist, building FAIR, the company's influential research lab. Last November he left to start his own company in Paris.

His departure wasn't quiet. "The entire industry has been LLM-pilled," he said recently. "The AI industry is completely LLM-pilled. In Silicon Valley, everybody is working on the same thing. They're all digging the same trench."

LeCun's new venture, Advanced Machine Intelligence (AMI Labs), focuses exclusively on world models. Specifically, on a framework called JEPA—joint embedding predictive architecture—that he developed at Meta. Instead of predicting every pixel or every word, JEPA learns abstract representations and makes predictions in that simplified space.

It's the difference between memorizing that basketballs bounce and understanding the physics of elastic collisions. The first requires endless examples. The second generalizes to tennis balls, superballs, and objects you've never seen.

"LLMs are not a path to superintelligence or even human-level intelligence," LeCun argues. "I have said that from the beginning."

His timing is pointed. Meta has pivoted hard toward language models and chatbots, spending billions on data centers to train ever-larger systems. LeCun apparently decided the company was digging that same trench he's now criticizing.

The startup making Chess-playing AI look simple

Logical Intelligence, a San Francisco startup, just appointed LeCun to its board. They're building something called energy-based reasoning models, or EBMs.

Here's how founder Eve Bodnia explains the difference: imagine climbing Mount Everest. An LLM climber picks one direction and keeps going. If there's a hole, they fall into it. They can't deviate until they complete the task.

An EBM climber sees the whole map. They evaluate multiple paths simultaneously. When they encounter an obstacle, they backtrack and try another route. The summit is always in mind, but the path can change based on conditions.

Their debut model, Kona 1.0, solves sudoku puzzles many times faster than leading LLMs despite running on just one Nvidia GPU. The model doesn't just predict likely numbers—it understands the constraints and reasons through valid solutions.

"This is not a guessing game," Bodnia says. "It's actual reasoning."

The company sees applications in grid management, drug discovery, and chip manufacturing—anywhere you need error-free optimisation within complex constraints. They've already talked to one of the world's largest chip manufacturers and multiple data centres.

What "AGI" actually means now

The debate over artificial general intelligence—AI that matches human-level reasoning across any domain—has gotten messy.

Anthropic CEO Dario Amodei told an audience at Davos that AI would replace all software developers within a year and achieve "Nobel-level" scientific research within two. He predicts 50% of white-collar jobs gone in five years.

OpenAI's Sam Altman talks about superintelligence—AI smarter than all humans combined—as if it's around the corner.

LeCun and Hassabis are more cautious. Hassabis puts genuine AGI at "five to 10 years" with a 50% probability, and only if researchers make "one or two more breakthroughs" beyond current approaches. He lists missing capabilities: learning from few examples, continuous learning, better long-term memory, improved reasoning and planning.

LeCun has abandoned the term AGI entirely. "The reason being that human intelligence is actually quite specialized," he explains. "So calling it AGI is kind of a misnomer." He prefers "advanced machine intelligence"—AMI, conveniently the name of his startup.

The disagreement isn't just semantic. It reflects fundamentally different views on whether current approaches can reach human-level intelligence or whether something entirely new is required.

The Chinese wild card

While Silicon Valley debates, Chinese companies have fully embraced open-source AI. Tencent, DeepSeek, and others are releasing powerful world models that anyone can download and modify.

LeCun sees this as a strategic advantage for China. "All leading open-source AI platforms are Chinese," he notes. "The result is that academia and startups, outside of the US, have basically embraced Chinese models."

He's not anti-Chinese—he calls their engineers and scientists "great." His concern is different. In a future where AI mediates most of our information diet, do we want a choice between proprietary American models and open Chinese ones that might be fine-tuned to avoid certain topics?

"If there is a future in which all of our information diet is being mediated by AI assistance, and the choice is either English-speaking models produced by proprietary companies always close to the US or Chinese models which may be open-source but need to be fine-tuned so that they answer questions about Tiananmen Square in 1989—you know, it's not a very pleasant and engaging future."

His solution? European companies building open-source world models independent of both superpowers. "There is a huge demand from the industry and governments for a credible frontier AI company that is neither Chinese nor American," he argues.

What happens next

World models are coming whether Silicon Valley pivots or not. The applications are too valuable, the limitations of language-only AI too obvious.

Nvidia is positioning its Cosmos platform as infrastructure for the next AI generation. Google DeepMind just released Genie 3 to public. World Labs shipped Marble commercially in late 2025. Meta's Habitat 3 is training robots right now.

The timeline depends who you ask. Optimists like Hassabis see AGI-level systems within a decade. Skeptics like Ravi Kumar, CEO of Cognizant, think capturing AI's current value matters more than chasing theoretical superintelligence. "That $4.5 trillion will generate real value in enterprises if you start to think about reinvention," he said at Davos, referring to potential productivity gains from today's technology.

But the direction is clear. AI can't remain blind to the physical world indefinitely. Language skills are impressive—they're just not enough for robots, autonomous vehicles, or any system that needs to take actions in three-dimensional space.

"We need machines that understand the world," LeCun said recently. "Machines that can remember things, that have intuition, have common sense—things that can reason and plan to the same level as humans."

That's world models in a sentence. Not AI that sounds smart. AI that actually is.

GN Awards 2025: Vote for your favorite Gadgets