Large language models learn from text. They predict which word comes next. AI world models do something fundamentally different: they learn how environments work, predicting how states change when actions are taken. A world model does not just know that “the ball fell” is a plausible sentence — it understands why the ball fell and where it will land.

The concept has roots in cognitive science. Kenneth Craik proposed in 1943 that the human mind creates internal models of reality — a mental simulator that plays out consequences before we act. In AI, Jürgen Schmidhuber formalized this idea in 1990. A 2018 paper by David Ha and Schmidhuber revived it, and Yann LeCun’s influential 2022 position paper made it a central pillar of the field’s ambitions.

How world models work

A world model is built around an encoder-predictor architecture. The encoder takes raw inputs — video frames, sensor readings, images — and compresses them into abstract internal representations capturing what matters about the current state of the environment. The predictor then answers: if I take this action, what comes next?

The key innovation, developed by LeCun’s team at Meta into an approach called JEPA (Joint Embedding Predictive Architecture), is to make predictions in abstract representation space rather than raw pixel space. Reconstructing every pixel of a video frame is wasteful and noisy; predicting “the object moved left and the light changed” is efficient and generalizable.

Training typically happens in two stages. First, the model learns from enormous quantities of video — millions of hours — developing a sense of motion, physics, and causality without any labels. Then a much smaller dataset of action-conditioned examples teaches it what specific actions produce in specific contexts.

What LLMs cannot do — and world models can

Large language models are trained on text. They learn that certain word sequences are probable. They have no grounded understanding of physical cause and effect, spatial relationships, or what happens when an object is pushed.

The practical gap: LLMs struggle with spatial reasoning, cannot plan robustly for physical tasks, and hallucinate facts they have no physical basis for verifying. A world model, by contrast, can simulate — it imagines many possible futures, evaluates them, and selects the action most likely to achieve a goal.

This is why researchers from Yann LeCun to Demis Hassabis argue that world models are essential for AI systems that do not just talk about the world, but act in it.

Notable examples

Meta V-JEPA 2 (June 2025) is a 1.2-billion-parameter world model trained on over one million hours of video. Given a goal and an unfamiliar scene, it can plan robot manipulation tasks without any robot-specific training data — it simply applies its learned model of physics.

Google DeepMind Genie 3 (August 2025) generates interactive 3D environments from text descriptions, modelling water, lighting, and physics in real time. Waymo built on it to synthesize rare driving scenarios for autonomous vehicle training.

NVIDIA Cosmos 3 (June 2026) is an open-weight “omnimodal” world model trained on 20 trillion tokens including video, images, audio, and action data. It generates not just video but physical actions — robot joint angles and trajectories — making it a training environment for physical robots. Companies including Agility Robotics and Figure AI use it.

Alibaba Qwen WebWorld (May 2026) takes the world-model approach into a different domain: the browser. Rather than predicting physical states, it predicts what a webpage looks like after a user action — used to train web-browsing AI agents in simulation before deploying them live.

What they are used for today

The most active application is robotics. Training physical robots in the real world is slow and expensive. World models provide unlimited synthetic environments. Meta’s V-JEPA 2 enables manipulation planning with a fraction of the real-world data that earlier approaches required.

Autonomous vehicles are a close second. Dangerous driving scenarios — the rare events that cause accidents — cannot be collected in sufficient quantity from real roads. World models synthesize them on demand.

Games and interactive media are a growing frontier. World Labs (founded by Fei-Fei Li) generates persistent 3D environments from text or photos, exportable to game engines like Unreal Engine, with a free tier available.

In the news

Alibaba recently open-sourced Qwen WebWorld, a world model for web-agent tasks that outperforms GPT-5.4 on agent benchmarks — the latest sign that world models are moving beyond robotics labs into everyday software. For more on the broader research landscape, see the AI research hub.

FAQ

Is a world model the same as a simulation?
No. A simulation is handcrafted — physics rules written by engineers. A world model is learned from data and generalizes to situations its creators never explicitly programmed.

Do world models replace LLMs?
No — they solve different problems. LLMs excel at language tasks. World models handle spatial, physical, and action-conditioned reasoning. Future AI systems are likely to combine both.

Can I use a world model today?
Several are publicly available. Meta V-JEPA 2 is open-source on GitHub and Hugging Face (commercial use permitted). NVIDIA Cosmos 3 is open-weight. Alibaba Qwen WebWorld is on Hugging Face. World Labs offers a free tier for interactive environment generation.

How is this different from a digital twin?
A digital twin is a live virtual replica of one specific real object — a factory or a vehicle — fed by real-time sensor data. A world model is a general-purpose learned predictor that generalizes across environments it has never seen, without needing to correspond to any single real-world object.