You can hand an AI a legal brief and it'll tear through it faster than any paralegal alive. You can ask it to write code, explain molecular biology, translate Mandarin into Portuguese. It'll do all of it well. Sometimes remarkably well. Large language models (LLMs) — are a real form of intelligence.
But here's the thing about them that nobody talks about enough.
They've read everything. And they've experienced nothing.
Everything a large language model knows about the physical world came to it secondhand — a human experienced something real and described it in words. The model read the description. Millions of descriptions. It got so good at the descriptions that it can produce new ones indistinguishable from the originals. But it never smelled the smoke. It never felt the rain.
You know the scene in Good Will Hunting — the one on the bench in the park? Robin Williams tells Matt Damon: you can tell me everything about Michelangelo, but you've never stood in the Sistine Chapel and looked up. You know about things. You don't know things.
That's the LLM problem in one scene. These models have read the entire internet. They can tell you about anything. But they've never stood anywhere. They've never looked up.
How Humans Really Learn
You learned to read, sure. You sat in classrooms and absorbed information from books and lectures. That's real and it matters. But that's not all you did.
Think about being a kid on a road trip, sitting in the back seat. You felt the pull of a sharp turn in your stomach. You noticed nearby trees moving fast while distant mountains barely moved at all. Nobody explained parallax or inertia to you. You just felt them. And your brain quietly built a model.
A surgeon doesn't learn to operate by reading about surgery. A chef doesn't learn to cook by reading recipes. A pilot doesn't learn to fly by studying aerodynamics. They all study, yes — but the knowledge that matters, the knowledge that keeps people alive, comes from doing. From seeing. From feeling the feedback of the real world and letting their brains build models from that raw sensory experience.
Humans learn from text and from reality. LLMs learn from text only.
That's the gap. And it's not a gap you can close by adding more text.
The Technical Divide
Under the hood, the difference is structural.
LLMs take in tokens — words, fragments of words — and learn to predict what comes next. Language is a compression of reality. When a human writes "the car skidded on the wet road," that sentence is a dramatic reduction of what actually happened — the physics of friction, the weight transfer, the curvature of the road, the depth of the water. The LLM gets the sentence. All of that underlying reality is gone.
World models learn from the uncompressed version. They take in raw sensory data — video, depth, motion, spatial measurements — and build internal representations of how things in the physical world relate to each other and change over time. They don't try to predict reality at full resolution. Instead, they learn abstractions that capture the structure of what's happening without getting lost in surface-level detail.
This is the idea behind JEPA (Joint Embedding Predictive Architecture), the framework Yann LeCun developed at Meta. Rather than reconstructing a scene pixel by pixel, JEPA predicts what will happen at the level of concepts. A ball is in the air — where will it be in half a second? You don't need to render every frame to answer that. You need a model of how objects move through space. That's a fundamentally different kind of learning than next-token prediction, and it's closer to what biological brains actually do.
LeCun puts it plainly: "The idea that you're going to extend the capabilities of LLMs to the point that they have human-level intelligence is complete nonsense."
Language is a protocol designed for communication between minds that already understand the world. When you read "the glass shattered on the floor," you understand it because you've seen glass break, heard the sound, felt the sharpness. The sentence activates a world model you already have. An LLM processes the same words with no model underneath. It can use the word "shatter" perfectly. It has never understood shattering.
LeCun left Meta in late 2025 and founded AMI Labs in Paris. Last week, AMI announced a $1.03 billion seed round — one of the largest ever — to build world models for robotics, industrial automation, and healthcare. As AI researcher Luiza Jarovsky put it: "the beginning of the end of language supremacy in AI."
You Can't Scrape the Physical World
This is where it gets real.
LLMs had it easy on the data front. Trillions of words were already sitting on the internet — books, papers, Reddit, Wikipedia, Stack Overflow — digitized, scrapable, free. The training data for language models was, in a sense, a solved problem before anyone started building them.
World models have no such luck.
Training a world model requires physical-world data: continuous video streams, GPS trajectories, IMU readings, depth maps, radar returns, sensor data. This data is expensive to collect — you need cameras on vehicles, sensors on robots, hardware in the field. Every data point requires atoms, not just bits. It's hard to label — physical data needs spatial annotation, object identification, temporal alignment, all orders of magnitude more complex than text. It's messy — a dashcam in Phoenix and a dashcam in Oslo produce fundamentally different data. And it's siloed — most of it sits inside companies that have no reason to share it.
This flips the competitive dynamics of AI completely. In the LLM era, the moat was compute and architecture. The data was available to everyone. In the world model era, the moat is the data itself.
The companies that have been collecting real-world sensor data at scale — fleet operators with thousands of vehicles running dashcams and GPS, autonomous vehicle programs with petabytes of driving footage, robotics companies with millions of hours of manipulation data — are sitting on something that just became dramatically more valuable. They may not know it yet. But the researchers building world models know it.
The winners of this next era won't be determined by who writes the best architecture paper. They'll be determined by who has the best data about how the physical world actually works. The question is whether you're collecting that data now, or whether you'll be buying it from someone else later.
Follow us on X or Try Bee Maps for Free.