What Is a World Model? How It Could Upend the World, Its Operating Principles, the Four Giants’ Battle, and the Controversies

World models (World Model) allow AI to build an internal model of “how the world works,” understanding space, physics, and causality. What they predict is the world’s next state—not, like large language models, just the next word. They have already been deployed in autonomous driving, robotics, and game scenarios, but realistic-looking visuals don’t necessarily mean physical correctness. Whether the right path to AGI is LLMs or world models remains one of the biggest debates in the industry to date.
(Background recap: Fei-Fei Li on LLM’s next step: AI needs “spatial intelligence” to understand the real world—how does the Marble model achieve this?)
(Background supplement: Serenity: world models are the biggest consensus in the AI circle; Chinese funding abandons base models and shifts to physical AI)

Table of contents

Toggle

• What is a world model?
• How does a world model differ from a large language model (LLM)?
• How do world models work? Latent space and “predicting the next moment”
• The 2026 world model battlefield: what are four major players betting on?
• What can world models do? Real-world cases in autonomous driving, robotics, and games
• Controversies and limitations of world models: are they really stronger than LLMs?
• The future of world models: another path to AGI?

A world model (World Model) is one of the most frequently mentioned terms in today’s AI circle, but also one of the most commonly misunderstood. In one sentence: a world model is a system that enables AI to build an internal model of how the world works, helping it understand spatial, physical, and causal relationships, and then predict what will happen next based on that.

It is not the same technical route as ChatGPT, Claude, and other familiar large language models (LLMs). LLMs predict the “next token,” while world models predict the world’s “next state.”

What is a world model?

In simple terms, a world model is a technique that teaches AI to perform “mental simulation.” For example, before humans cross the street, they automatically rehearse in their minds: “If I step out now, how many seconds until that car reaches my front?” This rehearsal doesn’t require actually walking in front of the car once to find out the outcome.

World models aim to give AI this same capability: simulate many possibilities internally first, then decide the next action—rather than trying in the real world every time.

Stanford University computer science professor and a leading figure in AI, Fei-Fei Li, breaks the functions of world models into three parts:

• generation
• understanding
• simulation/interaction

The product her company World Labs builds—Marble—focuses exactly on the third part: simulation and interaction. This is widely viewed as the most critical, and also the most difficult, link in the whole technical chain. Generating a “pretty picture” is not hard; the hard part is generating a space that you can enter, interact with, and whose physical logic also makes sense.

This wave has concentrated into a burst in 2026 for partly a reason: the industry’s anxiety about the ceiling of LLMs has become obvious. As the ceiling of “text continuation” becomes increasingly clear, a group of researchers has started betting on another route—so that AI learns to understand “the world itself,” not merely to imitate how humans describe the world.

It’s worth clarifying a common misunderstanding first: world models are not trying to replace LLMs for conversation or writing. They solve entirely different problems: spatial reasoning, physical simulation, and action planning. That’s also why they first found real-world traction in domains that “need to understand the physical world,” such as autonomous driving, robotics, and games—rather than chatbots.

How does a world model differ from a large language model (LLM)?

Large language models (LLMs) are built on the Transformer architecture. Their operation is to predict the “next token” in a sequence. In simple terms: after seeing an enormous amount of text, the LLM learns “which next character or token is most likely to appear in this sentence.” The entire computation happens in “the space of text.”

They have never truly “seen” or “simulated” the physical world. Internally, there is no gravity, no collision, no spatial coordinates—only statistical relationships between pieces of text.

World models take a different route. Represented by JEPA proposed by Yann LeCun, they don’t predict every pixel of the raw image, nor do they predict text tokens; instead, they predict an abstract representation of future states in “latent space.”

Put plainly, it actively discards unimportant details (for example, a reflection line on a wall in the background) and keeps only the core information needed to understand the world—like which direction that ball is rolling toward, and how many seconds until the car reaches the intersection. Because the output is built on a clear, testable internal state, once the model’s “internal thoughts” don’t match the “output,” this kind of inconsistency is relatively easy to detect—something LLMs struggle to do.

The differences between the two can be quickly compared in the table below:

| Comparison item | | --- | Large language model (LLM) | World model (World Model) | | --- | --- | | Prediction target | Next token (text) | World’s next state | | Operating space | Text / language space | Latent space | | Core architecture | Transformer | JEPA, diffusion models, autoregressive Transformers, etc. | | Understanding physics | Does not simulate physics; no internal world model | Learns physical rules such as object movement, collisions, and gravity | | Typical errors | Hallucination (still over 15% for most models) | Visually realistic but physical correctness not guaranteed | | Representative applications | Dialogue, writing, code generation | Autonomous driving simulation, robot training, explorable 3D worlds |

One point to remind readers: “looks realistic” does not equal “is physically correct.” For world models built on video generation, the training objective more directly targets “whether it looks real,” and may not optimize equally for “whether the physics logic is reasonable.” In high-failure-tolerance scenarios like autonomous driving and robotics, this is a crucial difference—and one of the core sources of ongoing controversy.

How do world models work? Latent space and “predicting the next moment”

To understand how world models operate, you need to first understand two terms: latent space and predicting the next moment.

Latent space is, in simple terms, where the AI “compresses” the scene it sees into a set of abstract numbers—keeping only the information useful for understanding the world and discarding unnecessary visual noise.

For example, in a dashcam video, latent space might only preserve abstract descriptions like “there is a car ahead; the distance is shrinking; the speed is moderate,” without recording the color of every leaf in the scene. The benefit is that computation drops dramatically, and the model learns “patterns” instead of memorizing the scene.

With latent space in place, the next step is “predicting the next moment”: given the current state plus an action (for example, turning the steering wheel left), the model predicts what the next state will look like.

By repeating this process, it becomes an internal simulator you can “roll forward.” Give the AI a moment in time, and it can run a few possible futures in its head before deciding which action to take—without having to crash into the real world once just to find out the result.

JEPA (Joint Embedding Predictive Architecture) is currently the most representative architecture for this logic, proposed by Yang LeCun in his 2022 paper “A Path Towards Autonomous Machine Intelligence.”

JEPA’s core claim is: rather than forcing the model to generate every pixel detail (which wastes compute and can easily get misled by irrelevant noise), it should directly compare, at the abstract level, whether the “predicted future” matches the “actual future.”

That’s why LeCun says world models make errors “detectable.” Because their outputs must be aligned with a clear internal representation; if they don’t match, it’s a signal—not like LLMs, where errors can silently blend into a seemingly fluent stream of text.

However, latent space isn’t a cure-all. Which details to discard and which to keep is an engineering judgment with no standard answer. Discarding too much makes simulations inaccurate; discarding too little drags down computational efficiency—this is one of the biggest points of divergence among current world model architectures.

The 2026 world model battlefield: what are four big players betting on?

In 2026, the world model track has already gathered the most heavyweight players in the industry, and the scale of funding is large enough to show this is not a one-off technical gimmick.

World Labs, founded by Fei-Fei Li, was one of the earliest companies to go mainstream in this wave. In September 2024, World Labs came out after raising $230 million at a $1 billion valuation. In January 2026, it was reported to be negotiating a new round at roughly a $5 billion valuation. By February 18, 2026, World Labs officially announced it had raised $1 billion, with investors including AMD, Autodesk, Emerson Collective, Fidelity, NVIDIA, and Sea. Autodesk led with $200 million, which is the largest investment Autodesk has ever made in a startup.

In total, World Labs has raised about $1.23 billion to date. Its flagship product, Marble, opened a limited beta test in November 2025, and became commercially available in February 2026. Inputs can be text, photos, videos, panoramas, or rough 3D models. Outputs are navigable, persistently stored, editable 3D environments—and each generation produces two types of meshes at the same time: a triangle mesh for high-precision geometry, and a collider mesh designed for collision detection in physics engines.

Google DeepMind’s Genie 3 is currently the only world model that has achieved “real-time interaction.” It appeared first as a research preview in August 2025, became officially available on January 29, 2026, and was opened for use to Google AI Ultra subscription users in the United States. At the same time, it launched Project Genie online demos, letting users directly generate worlds, go explore them, and even remix modifications.

On technical specifications, Genie 3 is an 11B-parameter autoregressive Transformer that can generate navigable worlds in real time. Its resolution is 720p, with 24 frames per second (24fps). Consistency can remain stable for several minutes without collapsing.

NVIDIA is taking a platform route. It launched the Cosmos world foundation model platform on January 1, 2025, and on June 1, 2026 it introduced Cosmos 3, claiming it is the first fully open omnimodel. It uses a mixture-of-experts-style Transformer architecture, integrating visual reasoning, world generation, and action prediction into a single system, supporting multiple modalities including text, images, video, environmental audio, and actions.

The scale of training data is staggering: multimodal data totaling about 20 trillion tokens, including nearly 1 billion images and 400 million real and synthetic videos. NVIDIA says this system shortens the training and evaluation cycles for physical AI (embodied AI) from several months to just a few days. Under Cosmos 3, it also has three variants: Super (a high-precision version for robots and autonomous driving), Nano (pursuing extreme speed), and Edge. NVIDIA also assembled companies like Runway, Black Forest Labs, and Skild AI into the Cosmos Coalition alliance.

Yann LeCun’s move is the most dramatic. In November 2025, after leaving Meta where he had worked for 12 years and originally served as the director of FAIR, he co-founded AMI Labs with Alexandre LeBrun. The headquarters are in Paris, with additional sites in New York, Montreal, and Singapore.

On March 10, 2026, AMI Labs announced it raised $1.03 billion in a seed round (about €890 million), with pre-investment valuation of $3.5 billion. This is the largest seed round in European startup history. Investors include Bezos Expeditions, NVIDIA, Samsung, Temasek, Toyota Ventures, and well-known individual investors such as Mark Cuban and Eric Schmidt. LeCun has long publicly argued that “LLMs can’t reach general artificial intelligence,” and now, after leaving Meta, he is putting his entire post-Meta career—along with $1.03 billion—on the world model route.

The positioning of the four can be summarized as follows:

| Company / team | | --- | Key figure | Representative product | Latest funding / scale | Technical positioning | | --- | --- | --- | --- | | World Labs | Fei-Fei Li | Marble | Total about $1.23 billion (one round $1 billion in 2026/2) | Navigable, persistent 3D space generation | | Google DeepMind | DeepMind team | Genie 3 | Google internal resource support | 11B parameters, real-time interactive world generation | | NVIDIA | NVIDIA team | Cosmos 3 | NVIDIA in-house resources + Cosmos Coalition alliance | An open omnimodel platform, focusing on physical AI | | AMI Labs | Yann LeCun | JEPA-series research | Seed round $1.03 billion, valuation $3.5 billion | JEPA latent space prediction architecture; most aggressive theoretical route |

The four companies’ technical routes are not completely the same. World Labs emphasizes usable 3D space products, Genie 3 focuses on real-time interaction, Cosmos 3 takes an open platform approach, and AMI Labs concentrates on foundational architecture research. But the shared belief is consistent:

The next wave of AI breakthroughs won’t rely only on making LLMs bigger—it will come from enabling AI to truly “understand” the world it operates in.

What can world models do? Real-world cases in autonomous driving, robotics, and games

The scenarios where world models land first are almost all highly related to “needing to understand the physical world.” Autonomous driving is the most direct example.

Startup Decart launched Oasis 3 in June 2026. It can generate highly realistic driving environments in real time for up to several hours. It aims to help autonomous vehicles simulate the “rare scenarios” that are hard to encounter in real life but crucial for safety—for example, pedestrians suddenly appearing in a sudden downpour, or oncoming vehicles losing control and drifting out of their lane.

Oasis 3’s API is already open for use, priced at about 2 cents per second. Its goal is to capture the “simulation layer” position in the physical AI supply chain. Waymo also unveiled its own generative simulation architecture in February 2026 built on Genie 3. Companies like Tesla, NVIDIA, and Wayve are each working on similar things. The core value of this kind of technology is straightforward: it allows autonomous vehicles to repeatedly train across an “infinite variety of simulated scenarios,” without needing to take the car on the road and crash to learn lessons once.

Robotics is another key battlefield. Systems like Cosmos 3 and Oasis 3 can generate controllable, multi-view simulation environments. Robots can learn to grasp, walk, and manipulate objects in this virtual space, repeatedly improving their action policies. Replacing real-world data collection with synthetic data can greatly reduce the cost and time of training. Actions that previously required physical robots to try thousands of times in real environments can now run a first round in simulation.

Games and content creation are relatively “more appealing” application scenarios. Genie 3 and Marble can both generate a world you can enter and explore from a single text description, which directly serves as a productivity tool for game prototype development and virtual scene construction. At the same time, startups that originally built video generation models—like Luma and Runway—are beginning to pivot their “physics-aware” video generation models toward the direction of world models. This shows that world models are not an entirely new track created out of thin air, but rather the natural next step in the evolution of video generation toward “interactive” and “controllable.”

A common thread across these cases is that the value of world models is not “how pretty the visuals are,” but whether they can be used to train, simulate, and make decisions. That is precisely the biggest line of separation from simple video generation tools.

Controversies and limitations of world models: are they really stronger than LLMs?

The most attention-grabbing bet within the world model camp comes from Yann LeCun. After leaving Meta, founding AMI Labs, and raising $1.03 billion in a seed round, he effectively bet his entire career on a single judgment: the LLM route will never reach general artificial intelligence, and the way out must be world models. This stance is not a consensus in the industry.

In the opposing camp, the loudest voice is Anthropic CEO Dario Amodei. His claim is that as long as model scale continues to increase, along with incremental improvements to architectures, LLMs themselves are sufficient to overcome today’s limitations. Even optimistically, he expects a capability leap in 2026—something he describes as a “land of genius” inside data centers.

In other words, one side believes the ceiling of the current route has already been seen at the top and requires switching routes; the other side believes the ceiling hasn’t arrived yet, so keep pushing forward. As of now, neither side has produced decisive evidence to persuade the other. To some extent, both sides are using real money to bet on a question that still has no answer.

World models themselves are not without flaws. Earlier, we mentioned that the latent space architecture makes “internally inconsistent” errors easier to detect. But that doesn’t mean world models won’t make mistakes. It only means they solve “whether errors can be caught,” not “whether errors will happen.” The more critical issue is: visuals being realistic doesn’t mean the physics is correct.

For world models based on video generation, the training objective more directly targets “whether it looks real,” and may not optimize just as strongly for “whether the physical logic is unreasonable.” The trajectory of a rolling ball may look natural and smooth, but the actual acceleration and friction calculations may not hold up under scrutiny. This risk cannot be ignored in applications like autonomous driving and robotics, where one failure can be an accident.

Compute and latency costs are also real barriers. For world models based on diffusion, generating each frame requires multiple rounds of denoising computation. For scenarios requiring real-time reactions—like robots and autonomous vehicles—this latency cost is quite high. If decisions are delayed even slightly, in the real world it could result in a collision.

Additionally, according to estimates from the Open-Sora team, reproducing a commercially capable video generation model would require about $200k in compute costs. This is only the threshold to “reproduce” existing results, not including the R&D investment needed to develop new architectures from scratch.

The future of world models: another path to AGI?

Putting these clues together reveals a clear industry signal: in 2026, world models have already evolved from an academic concept into a route battle involving tens of billions of dollars and gathering the industry’s heaviestweights. Fei-Fei Li used Marble to prove that a navigable 3D world can become a product. Google DeepMind used Genie 3 to show that real-time interaction is technically feasible. NVIDIA used Cosmos 3 to turn it into an open platform and aims to capture the entire physical AI supply chain. LeCun bet his entire career—plus the largest seed round in European startup history—on the judgment that “LLMs can’t reach AGI.”

Whether this path can truly lead to general intelligence is something no one can give a definite answer to. But it is certain that as long as applications that need to “understand the physical world,” like autonomous driving and robotics, continue to expand, the commercial value of world models will keep being validated—without needing to wait until the ultimate question of “AGI” is answered.

For readers, rather than obsessing over whether world models will replace LLMs, it may be better to understand it from another angle: these two technical routes are likely to coexist for the long term. LLMs handle language and knowledge, while world models handle space and physics—each solving problems the other can’t solve.

In summary, world models (World Model) represent a serious divergence in the AI industry’s roadmap: when the hallucination problem in LLMs has not been solved for a long time, a group of top researchers has chosen to bet on enabling AI to directly understand the physical world itself. From Fei-Fei Li’s Marble, Google’s Genie 3, and NVIDIA’s Cosmos 3, to AMI Labs where Yann LeCun bet his entire career, more than $3 billion has already been poured into this battle.

Whether it will be another path to AGI is still too early to conclude—but one thing is clear: the debate about how AI should understand the world is only just beginning.