What Is Artificial World Intelligence™?

Large language models changed software because they made prediction over language extraordinarily cheap. That was a genuine breakthrough. In a few years, LLMs went from research curiosity to infrastructure: search, writing, coding, customer support, legal review, medical triage, education. The impact has been real and enormous.

But language is not the world.

An LLM predicts the next token in a sequence. It is very good at this, and that capability turns out to be useful for a remarkable range of tasks. But it does not understand what happens when you push a glass off a table. It does not know that a door swings on a hinge. It cannot predict what a car will do at an intersection based on the physics of the situation, only based on what text about intersections it has read. Language models are, as Fei-Fei Li has put it, "eloquent but ungrounded."[1]

Yann LeCun has made the same argument more bluntly. He left Meta in late 2025 to found AMI Labs around the conviction that LLMs will never produce systems with human-level intelligence, because they learn from the wrong substrate. "Real intelligence does not start in language," AMI’s mission statement reads. "It starts in the world."[2]

We are not here to declare LLMs dead. They are clearly not. But we do think the field is arriving at a recognition that language alone is not enough, and that the next generation of AI will need to learn from environments, not just text.

A world model is an AI system that learns what tends to happen next in an environment, given a current state and a possible action. That is a deceptively simple description for a genuinely hard problem.

Where an LLM predicts the next word, a world model predicts the next state of the world. Not just what looks plausible in a generated image, but what would actually happen if an agent took a specific action in a specific context. If I push this object, where does it go? If I turn left here, what do I see? If this character jumps, what does gravity do?

David Ha and Jürgen Schmidhuber helped formalize this idea in 2018, showing that agents could learn compressed internal simulations of their environments and even train inside their own imagined worlds.[3] Since then, the concept has expanded from reinforcement learning research into a broad family of systems: predictive latent models, controllable video simulators, physics-aware 3D generators, and embodied reasoning engines.

Most world model architectures share a common structure. An encoder compresses observations, whether video, images, audio, sensor data, or text, into a latent representation. A dynamics model predicts how that representation evolves over time, often conditioned on an action. And a planner or controller uses those predicted futures to decide what to do next. Some systems can also decode those latent states back into video or 3D scenes, which is why today’s world models often look like video generators, even though their core purpose is prediction and planning, not media production.

The technical details vary. DeepMind’s Genie line emphasizes interactive, action-controllable environments generated in real time.[4] Meta’s V-JEPA architecture emphasizes prediction in representation space rather than pixel-perfect reconstruction.[5] NVIDIA’s Cosmos emphasizes photoreal, physics-aware synthetic data for robotics and autonomous vehicles.[6] Different architectures, but they are all converging on the same destination: models that internalize consequences.

We use the term Artificial World Intelligence™ because we think it captures something that “world models” alone does not.

World models are a technical category. AWI is a thesis about where AI capability comes from and what it will require. It is the idea that the next meaningful leap in artificial intelligence will not come from predicting text more accurately, or generating images more beautifully, or even producing more convincing video. It will come from AI systems that can understand, predict, and reason about how environments work. Systems that grasp cause and effect. That can plan across time. That can act in the world, not just describe it.

This is what robotics needs. It is what autonomous vehicles need. It is what any AI system operating in the physical world, or in a complex interactive environment, will eventually need. And it demands a completely different kind of training data than the one the industry has been built on.

LLMs were trained on text: books, web pages, code, conversations. The entire open web became the default training corpus. Video models were trained on flat media: YouTube clips, stock footage, social media content. World models need something else entirely. They need structured, interactive, multimodal data from environments where actions produce consequences, where physics is real or at least consistent, where states change over time, and where the relationship between observation and action is preserved.

That is a fundamentally different data problem. And it is the data problem Origin Lab was built to solve.

The shift toward world intelligence is not a forecast. It is already underway, and it is moving fast.

In February 2026, World Labs raised over a billion dollars to build what Fei-Fei Li calls “spatial intelligence,” with backing from AMD, Autodesk, NVIDIA, Fidelity, and others. Their Marble product generates persistent, editable 3D environments from text, images, or video, with export support for game engines and VR headsets.[7]

Weeks earlier, Google DeepMind opened Project Genie to public users, powered by Genie 3, the first general-purpose world model capable of generating real-time interactive environments at 24 frames per second with visual consistency lasting several minutes. DeepMind has framed this explicitly as a stepping stone toward artificial general intelligence.[4]

Yann LeCun’s AMI Labs raised $1.03 billion in March 2026 at a $3.5 billion valuation, the largest seed round in European history. The company is built on JEPA, an architecture that learns from video and sensor data rather than text. AMI’s CEO Alexandre LeBrun put it bluntly: “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.”[2]

NVIDIA’s Cosmos platform, which has been downloaded over two million times, provides open world foundation models trained on 9,000 trillion tokens from 20 million hours of data. Agility Robotics, Figure AI, Uber, and others are already using Cosmos to generate training data for physical AI systems. Jensen Huang has said that world foundation models are “a breakthrough for physical AI” comparable to what LLMs were for generative and agentic AI.[6]

And Runway released GWM-1, its first dedicated world model, with variants for environments, robotics, and avatars, positioning video generation as a stepping stone toward general-purpose simulation.[8]

Over two billion dollars raised for world models in the first three months of 2026 alone. Multiple independent bets from the field’s most respected researchers. Real products shipping. This is not a research tangent. It is the next platform shift.

Here is the part that matters most to us, and to anyone building in this space.

Every one of these world model efforts faces the same constraint: data. Not just any data. Structured, multimodal, consequence-rich data from environments where physics, causality, and state changes are observable and learnable.

NVIDIA’s own Cosmos documentation says it directly: building physical AI requires petabytes of video data and massive simulation effort.[6] Scientific American noted that training world models demands data capturing how the real world works, including motion, depth, spatial relationships, and temporal dynamics.[9]

And this is where the current data supply chain falls apart. The open web was a reasonable, if ethically fraught, substrate for language models. It is a terrible substrate for world intelligence. Scraped video does not carry action labels. It does not include engine state or camera telemetry. It has no chain of custody, no structured metadata, no synchronized signals across modalities. It is flat, passive, and disconnected from the environment that produced it.

The data that world models actually need looks much more like what comes out of video games, 3D simulations, and interactive digital environments. These are systems where actions have consequences, where states are measurable, where observations are synchronized with inputs, and where the environment itself is structured, replayable, and rich with causal information.

That is the data Origin Lab builds. Licensed, enriched, structured, and captured at the source, not scraped after the fact. We believe this is not a niche product for a small market. It is the training infrastructure for the next generation of AI.

In our next post, we will make the case for why the scraping era is ending and why what replaces it will define the quality of the AI systems we all live with.

For now, the point is this: AI learned language, and that changed everything. Now it needs to learn worlds. And the worlds it learns from will shape the intelligence it develops.

We think that matters. And we think the term for it is Artificial World Intelligence™.