Yann LeCun just raised €890 million for a new lab called Advanced Machine Intelligence. The bet is simple: current AI systems don't understand the physical world, and scaling language models won't fix that.
This isn't a side project. AMI Labs launched with one of the largest seed rounds in AI history, backed by serious institutional money. The plan is to build AI systems that model how the world actually works - what LeCun calls world models.
The World Models Thesis
Language models like GPT-4 or Claude learn from text. They're brilliant at language patterns but fundamentally limited - they've never seen, touched, or moved through physical space. They learn associations between words, not how gravity works or why objects fall.
World models take a different approach. Instead of predicting the next word, they predict what happens next in the physical world. Show them a video of someone throwing a ball, and they learn physics. Show them enough examples, and they build an internal model of how objects move, collide, and interact.
LeCun has been advocating for this approach for years, particularly through his work on JEPA (Joint Embedding Predictive Architecture). The core idea: learn by predicting what you'll see next, not by reconstructing what you've already seen.
Why This Matters Now
The timing is pointed. While most AI labs are scaling language models larger and larger, LeCun is explicitly positioning AMI as an alternative path. It's a public disagreement with the "scale is all you need" philosophy that dominates current AI development.
The €890M seed round suggests investors see merit in the argument. For context, that's more than most AI companies raise across their entire lifetime. It's a bet that understanding the physical world is valuable enough to justify building completely new architectures.
AMI has already attracted elite vision researchers - people who've spent careers working on how machines perceive and understand visual information. That's significant. World models need different expertise than language models, and AMI seems to be assembling the right team.
The Practical Implications
If world models work as promised, they'd be transformative for robotics. Current robots struggle with basic physical reasoning - they can't predict what happens if they push an object or how to manipulate something they've never seen before.
A robot with a working world model could plan actions by simulating outcomes internally. Instead of learning specific tasks through trial and error, it could reason about physics and predict consequences. That's the difference between a robot that can stack boxes in a warehouse versus one that can adapt to any physical task.
For autonomous systems generally - drones, vehicles, industrial machinery - better physical understanding means safer operation. These systems currently rely on enormous amounts of specific training data. World models could generalise better from fewer examples.
The Sceptical View
Here's the counterargument: language models already show surprising physical reasoning abilities. They can predict physics outcomes, understand spatial relationships, and even plan physical tasks - all without explicit world modelling. Maybe scaling language models gets you there anyway.
The practical challenge is also substantial. Building world models requires massive amounts of video and physical interaction data, then training systems that can actually learn useful representations from it. It's computationally expensive and architecturally complex.
LeCun has been making this argument for years. The question is whether €890M and a dedicated lab can finally prove it works at scale.
What to Watch
AMI represents a genuine fork in AI development philosophy. If they succeed, it validates an alternative to pure language model scaling. If they struggle, it suggests the current paradigm was right all along.
For builders and businesses, the near-term impact is limited - world models are research-stage work. But the long-term implications are significant. The AI systems we're building today might be fundamentally limited by their lack of physical understanding.
LeCun is betting nearly a billion euros that fixing that limitation is the next frontier. That's a bet worth paying attention to.