What Is a World Model in AI? How It Differs From an LLM

A world model is an AI system trained to predict how an environment will change next, rather than which word comes next in a sentence. Instead of learning from text, it learns from video, images, sensor data, or simulated environments, building an internal sense of physics, space, and cause and effect. That distinction is why Yann LeCun left Meta to build one, why Google DeepMind and Fei-Fei Li's World Labs have both shipped their own versions in the past year, and why Waymo now uses one to train self-driving cars on scenarios it has never actually encountered on the road.

What is a world model in AI?

In AI research, a "world model" is a system that learns an internal, predictive model of how an environment behaves — physically, spatially, or causally — so it can simulate what would happen next given a current state and an action. The term was popularized by researcher David Ha and Jürgen Schmidhuber's 2018 "World Models" paper, but it has since become the umbrella term for a wave of 2025–2026 products: Google DeepMind's Genie line, Fei-Fei Li's World Labs and its Marble model, Meta's JEPA family, and Yann LeCun's new venture AMI Labs.

What unites them is the training objective. A world model is trained to predict the next frame of a video, the next state of a 3D scene, or the sensory consequence of an action — essentially, to answer "what happens next?" for the physical or visual world, rather than "what word comes next?" for a sentence.

How is that different from a normal LLM?

A large language model such as GPT, Claude, or Gemini is trained almost entirely on text, and its core task during training is next-token prediction: given a sequence of words, predict the most likely next word, over and over, across trillions of tokens scraped from the web, books, and code. That process turns out to be extraordinarily good at producing fluent language and even reasoning-like behavior, but the model's internal "understanding" of the physical world is second-hand — inferred entirely from descriptions of the world in text, not from directly observing it.

A world model flips the training data and the objective. Instead of text tokens, it is fed video, images, 3D scans, or robot sensor logs. Instead of predicting the next word, it predicts the next frame, the next camera view, or the next sensor reading given an action. Google DeepMind's Genie 3, for example, generates a continuously playable 3D environment frame-by-frame in response to a text prompt and the user's movements, rather than generating a paragraph describing a 3D environment. Meta's V-JEPA 2 does something related but more abstract: it watches video and predicts what an unseen portion of that video will "look like" in an internal mathematical representation, not in pixels.

Why the training data difference actually matters

Text is a compressed, human-curated summary of the world — someone already did the work of observing a physical event and describing it in words. That makes text efficient to learn from, but it also means an LLM never directly learns the raw physical rules (gravity, occlusion, momentum, object permanence) that produced the description. It only learns statistical patterns in how humans talk about those rules.

Video and sensor data, by contrast, contain that physical information directly, but in a much noisier, higher-bandwidth, less labeled form — which is exactly why world models are a newer and harder engineering problem than LLMs. Meta says its V-JEPA 2 model was pretrained on more than one million hours of internet video and one million images, then separately fine-tuned on under 62 hours of real robot interaction video from the open-source DROID dataset to learn action-conditioned planning, according to Meta AI's official release. That lopsided ratio — enormous amounts of passively watched video, then a comparatively tiny amount of "hands-on" robot data — is a recurring pattern in world-model training, and it's part of why AI researchers worry about training-data quality more broadly as these data-hungry systems scale up.

Meta's JEPA: predicting in "concept space," not pixels

Meta's world-model research is built around an architecture Yann LeCun proposed in a 2022 position paper called JEPA — Joint Embedding Predictive Architecture. The idea: rather than generating realistic video frame-by-frame (which LeCun has argued wastes enormous compute modeling irrelevant pixel-level detail, like the exact texture of leaves on a tree), JEPA encodes both the observed and the predicted portions of a video into a shared, abstract "embedding space," and the model only has to predict the future embedding — a compressed representation of what matters — not the future pixels.

Meta's FAIR lab shipped this in stages: I-JEPA for still images in 2023, V-JEPA for video in 2024, and V-JEPA 2 in June 2025 — a 1.2-billion-parameter model Meta describes as state-of-the-art for visual understanding, prediction, and zero-shot robot planning. In Meta's own lab tests, robots using V-JEPA 2 for subgoal-based pick-and-place of unfamiliar objects in new environments succeeded 65–80% of the time, per Meta's release. Meta frames JEPA as a step toward what it calls "Advanced Machine Intelligence" — the same phrase, not coincidentally, that LeCun used to name his new startup after leaving the company.

Google DeepMind's Genie 3: a playable world, generated live

Google DeepMind took the opposite approach from JEPA: generate the actual pixels, in real time, and make the result interactive. Genie 3, announced August 5, 2025, is described by DeepMind as a general-purpose world model that generates an "unprecedented diversity of interactive environments" from a text prompt. A person (or an AI agent) can then move through that generated world as if it were a video game, at 24 frames per second and 720p resolution, with the model generating each new frame based on the scene so far and the latest action — rather than relying on a pre-built 3D map, the way a game engine or techniques like NeRFs and Gaussian splatting do.

DeepMind says Genie 3 can hold a consistent, navigable world in memory for a few minutes, and supports "promptable world events" — typing an instruction mid-session to add a character, change the weather, or alter the scene. DeepMind's own framing is telling: it positions world models like Genie 3 as an effectively unlimited training curriculum for AI agents, and used it to generate novel environments for testing its SIMA agent. The company has called generalizable world models a key stepping stone toward artificial general intelligence. At launch, Genie 3 was limited to a small research preview of academics and creators, and DeepMind has acknowledged real limitations: it can't model real locations with geographic accuracy, text rendered in scenes is often illegible, and sessions are capped at a few minutes.

From video games to the road: Waymo's world model

The clearest real-world payoff of a generative world model so far is autonomous driving. In a February 6, 2026 blog post, Waymo introduced the "Waymo World Model," a driving simulator built directly on top of DeepMind's Genie 3 and adapted through additional training to output the camera and lidar sensor data Waymo's own self-driving hardware actually uses (rather than just a video feed for a human to watch).

Because Genie 3 was pretrained on huge amounts of general video, Waymo says the resulting simulator can generate driving scenarios its fleet has never recorded firsthand — the company's own published demo scenes include tornadoes, flooded streets, and encounters with loose animals. Engineers can steer a simulation three ways: directly controlling the simulated vehicle's driving actions (to test counterfactuals, like "what if the car had braked sooner"), adjusting the scene layout, or simply describing changes in plain language, such as switching the time of day or weather. Waymo also says the system can convert an ordinary dashcam clip into a simulation of how its own Waymo Driver software would have perceived and handled that same scene. Waymo frames this as a way to prepare its driver for rare, dangerous situations before encountering them for real — on top of the nearly 200 million fully autonomous miles the company says its fleet has already driven.

Fei-Fei Li's World Labs and Marble

World Labs, the startup co-founded by Stanford computer scientist Fei-Fei Li, took a third angle: generating static, editable 3D worlds rather than live, agent-driven ones. On November 12, 2025, World Labs announced that its model, Marble, was generally available, calling it "a first-in-class generative multimodal world model." Marble can build a persistent 3D environment from a single image or a short text prompt, or stitch together several photos or a short video of a real space into one coherent 3D world. Users can then edit it — removing an object, restyling a room, or reshaping the layout — export it as a Gaussian splat, a textured mesh with physics colliders, or a flythrough video with precise camera control, and combine multiple generated worlds into a larger one.

Alongside Marble, World Labs launched Marble Labs, a hub of workflows and case studies aimed at game developers, VFX artists, and — notably — roboticists, who can use a generated 3D environment as a training ground for a simulated robot before testing it physically. World Labs frames all of this as a step toward "spatial intelligence": AI that understands and can act within 3D space, not just describe it in words. The company, founded in 2024, has raised well over $1 billion across multiple funding rounds since launching, reflecting how much investor appetite has shifted toward this category.

Why Yann LeCun bet his career on this instead of LLMs

No single event has done more to legitimize "world models" as a distinct, investable category than Yann LeCun's exit from Meta. LeCun, Meta's chief AI scientist and head of its FAIR research lab since 2013, confirmed on November 19, 2025 that he would leave the company to start his own venture, after reportedly arguing in press interviews that large language models are "a dead end" on the path to more general machine intelligence. He co-founded the new company, AMI Labs — short for Advanced Machine Intelligence Labs — in December 2025, taking the role of executive chairman while Alexandre LeBrun, previously CEO of the healthcare AI startup Nabla, became chief executive.

In March 2026, AMI Labs announced a seed round widely reported at roughly $1.03 billion, at a pre-money valuation of about $3.5 billion — described by multiple outlets as the largest seed round in European AI history. The company is headquartered in Paris, with additional hubs in New York, Montreal, and Singapore. Its stated mission is to build world models that learn directly from multimodal sensory data about the physical world, continuing the JEPA-style research direction LeCun pursued at Meta, rather than scaling up text-based LLMs further.

Why this matters for robots, agents, and games

An LLM can describe how to pick up a glass of water in perfect, fluent detail without having any real internal model of gravity, friction, or how glass behaves when it tips. That gap is exactly what's limiting today's AI agents and robots from operating reliably in physical or dynamic environments. A world model's entire purpose is to close that gap by learning directly from the modality — video, 3D scenes, sensor streams — where that physical information actually lives.

That's why the earliest practical applications of world models cluster around three areas: robotics (Meta's V-JEPA 2 doing zero-shot pick-and-place planning), autonomous vehicles (Waymo generating rare-event driving scenarios it could never safely capture on real roads), and game and simulation generation (DeepMind's Genie 3 and World Labs' Marble building explorable 3D environments on demand). The common thread in every case is using a world model to generate an effectively unlimited supply of realistic training or testing scenarios — something a text-only model structurally cannot do, because text can describe a scenario but can't simulate the second-by-second physical consequences of acting inside it. This is also conceptually distinct from the reasoning models that have dominated recent AI headlines, which improve an LLM's answers by spending more computation at inference time on a problem — they're still fundamentally text-in, text-out systems, not systems that learn from or predict physical environments.

The limitations nobody has solved yet

Every major world model shipped so far comes with the same family of caveats, acknowledged by their own makers. DeepMind says Genie 3 sessions are capped at a few minutes, can't reproduce real-world geography accurately, and struggle with multiple interacting agents in one scene. Meta says there's still a "noticeable gap" between V-JEPA 2's robot-planning performance and human performance on the same tasks. Waymo, for its part, hasn't published independent benchmark results for its Genie 3-based simulator, and World Labs' Marble is, for now, generating static or lightly editable worlds rather than ones that evolve with the complex physics of a live simulation. None of these systems yet model the open-ended, long-horizon physical world as reliably as humans navigate it — which is precisely the gap AMI Labs, DeepMind, Meta, and World Labs are all racing, with very different architectures, to close.

What's next for world models

Expect the next year to be defined by three things: longer context (today's world models hold a consistent scene for minutes, not hours), tighter integration with physical robots and vehicles beyond research demos, and consolidation of the underlying architecture debate between "generate the pixels" approaches like Genie 3 and Marble versus "predict in embedding space" approaches like JEPA. With AMI Labs now fully funded and staffed, DeepMind iterating publicly on Genie, World Labs scaling Marble commercially, and Waymo already shipping a production use case, world models look less like a research curiosity and more like the architecture the industry is quietly betting will matter as much as the transformer did for language — even though, for now, no world model can yet sustain the kind of open-ended, hours-long physical reasoning that would be needed to fully replace task-specific robotics or simulation software. For readers tracking the broader AI agent landscape, it's worth comparing this shift against other architecture changes reshaping how AI models are built and scaled, since world models are arguably the biggest structural departure from the transformer-and-tokens recipe that has defined AI since 2017.

Written byTheo Park

Theo Park runs the AI desk at Pandromeda. He follows model launches from the frontier labs and the open-weight community, tracks the assistants and developer tools built on them, and explains what each release changes on pricing, capability and safety. His reporting leans on primary sources: model cards, technical reports, API documentation and the companies' own announcements.

More from Theo ParkAll AI coverage