World models, the AI systems designed to let machines simulate how the world works before acting in it, have attracted roughly $4 billion in funding in a single year, yet independent research shows the most visually impressive ones may understand physics the least. Two major studies found that popular video generators like Sora, Runway, and Pika fail fundamental physics tests, scoring near random chance when asked whether impossible scenarios (objects vanishing, sliding through walls) are actually possible. The disconnect between investment hype and actual understanding is reshaping how the field thinks about what these systems can really do.

The money arrived before anyone could agree on what a world model actually is. In March 2026, Yann LeCun, a Turing Award winner who spent over a decade as Meta's Chief AI Scientist, left to found AMI Labs, which raised $1.03 billion at a $3.5 billion pre-money valuation, backed by Bezos Expeditions, Nvidia, Eric Schmidt, and Mark Cuban. Fei-Fei Li, the Stanford professor known as the "godmother of AI" for creating ImageNet, founded World Labs on the thesis that spatial intelligence (understanding 3D space) is the next frontier after language. The robotics side of the movement proved even larger, with Skild AI raising $1.4 billion, Physical Intelligence $600 million, Odyssey $310 million, and Decart $300 million.

Yet ask five of these companies to define a world model and you get five different answers. Some describe it as an interactive video game, others as a 3D asset generator, a robot brain, a driving simulator, or a new path to human-level AI. LeCun himself captured the absurdity: "My prediction is that 'world models' will be the next buzzword. In six months, every company will call itself a world model to raise funding." He said this while raising a billion dollars for a world-model lab.

What Exactly Is a World Model, and How Does It Differ From Language AI?

The core difference between a world model and a language model (LLM) is fundamental. A language model predicts the next word in a sequence of text. A world model makes a different prediction: given the current state of the world and an action, what happens next? Not the next word, but the next state of reality.

Think of two chess players. One has memorized every recorded game in history and plays brilliantly as long as the board stays within known patterns. The second simulates moves in her head: "If I take the knight, the pawn structure collapses, then..." The first player recalls. The second plans in imagination. That distinction captures the difference between a pattern model and a world model. The idea is older than the current hype. In 2018, researchers David Ha and Jürgen Schmidhuber published a paper titled "World Models," where a small agent learned a compressed model of its environment and then trained inside its own dream of it. The concept matured through DeepMind's Dreamer series, culminating in April 2025 when DreamerV3 was published in Nature as the first agent to find diamonds in Minecraft with zero human data, purely by imagining futures inside its learned model.

Why Do the Most Photorealistic World Models Fail Basic Physics Tests?

Here is the uncomfortable finding that should calibrate every investment decision in this space: two rigorous studies tested whether the most photorealistic video generators had actually learned physics. Google DeepMind's Physics-IQ benchmark tested fluid dynamics, optics, solid mechanics, magnetism, and thermodynamics, concluding in five words: "visual realism does not imply physical understanding." ByteDance's phyworld study went further, finding that video generation "fails to learn physical laws from video data, even with scaling." The models perform case-based mimicry of their nearest training examples; on out-of-distribution scenarios (situations unlike anything in the training data), errors jump by an order of magnitude.

Meta's IntPhys 2 benchmark makes the gap human-legible. The test is disarmingly simple: show short videos in which one object quietly does something impossible (slides through a wall, vanishes behind a box, floats away) and ask, "possible or impossible?" Humans get it right 85 to 95% of the time. Current models score at or near chance, roughly a coin flip. They cannot reliably tell physics from magic.

The implications are stark. Models like DeepMind's Genie 3 (released August 2025) and Nvidia's Cosmos predict the future as literal video frames. Type a prompt, get a navigable 720p world at 24 frames per second, consistent for a few minutes, with about one minute of visual memory. It is spectacular and easy to demo, but computationally very heavy: the system renders every raindrop, including the ones that do not matter for the decision you need to make.

How Are Researchers Approaching the Physics Problem Differently?

• Latent-space prediction: Meta's V-JEPA 2 (June 2025) predicts the future in a compressed, abstract representation instead of pixels. No visual frames; the model predicts a summary of the future. A 1.2 billion parameter model trained on over 1 million hours of video plus only 62 hours of robot data achieved 65 to 80% zero-shot success picking and placing objects it had never seen in environments it had never visited.
• Efficiency through abstraction: LeCun argues that only the latent-space lane scales to real understanding, because predicting every pixel wastes computational capacity on unpredictable detail. The model learns causal structure rather than memorizing visual patterns.
• Practical value despite imperfect understanding: Waymo, a subsidiary of Alphabet working on self-driving technologies, is shipping a production simulation tool on top of Genie 3, proving that pixel-space models create real value even without deep physical understanding. Usefulness and understanding are different axes.

The real fault line in this field is not "LLMs versus world models." It is generating convincing pixels versus learning causal structure. A cook who has memorized ten thousand recipes reproduces them flawlessly, until the kitchen changes. New altitude, different butter, and the sauce splits. A chef who understands why emulsions hold adapts on the spot. Photorealistic video models are, today, mostly the first cook. The bet of the latent camp is that you can train the second one.

Two rival architectures are now competing to become the standard. They share a goal (predict the future, then plan against the prediction) and disagree on everything else. The pixel-space camp emphasizes visual fidelity and ease of demonstration. The latent-space camp emphasizes efficiency, scalability, and the possibility of learning genuine causal understanding. As the field matures, the distinction between hype and substance will increasingly depend on which approach can deliver both visual credibility and genuine physical reasoning.

" }