LeCun vs. Li: Inside the $1B War to Build AI World Models

TL;DR

• AMI Labs (Yann LeCun) and World Labs pulled in over $1 billion combined to build world models — AI systems that simulate environments instead of just predicting words.
• New analysis breaks world models into five distinct categories, including JEPA and spatial intelligence approaches.
• V-JEPA 2 from AMI Labs achieved zero-shot robot planning after training on just 62 hours of data, signaling a leap for physical AI.
• The funding war between LeCun and Fei-Fei Li’s teams blurs the line between pure research and robotics infrastructure bets.

The Billion-Dollar Bet on World Models

AMI Labs and World Labs just raised over $1 billion between them, and both companies are chasing the same prize: world models that can simulate reality well enough to plan actions in it. Yann LeCun’s AMI Labs and Fei-Fei Li’s World Labs represent two of the biggest names in AI research jumping from academia into the startup arena. The money signals a shift in where investors think the next breakthrough lives — not in bigger language models, but in systems that understand physics, space, and causality.

AMI Labs announced V-JEPA 2, a model that can plan robot actions in zero-shot scenarios after training on just 62 hours of video data. That’s a remarkably small dataset by modern AI standards. Most robotics models demand thousands of hours of labeled demonstrations or millions of simulated interactions. V-JEPA 2 apparently skips that grind by learning a compressed representation of how the world works, then using that internal model to figure out what to do next.

World Labs hasn’t detailed a specific product yet, but Li’s team reportedly focuses on spatial intelligence — teaching AI to reason about 3D environments the way humans do. The company’s pitch centers on models that don’t just recognize objects in images but understand how those objects relate in physical space. If you want a robot to navigate a cluttered warehouse or assemble furniture, spatial reasoning matters more than language fluency.

Five Flavors of World Models

The analysis from Radical Data Science breaks world models into five categories, a taxonomy that clarifies what’s actually being built under the hype. JEPA — Joint Embedding Predictive Architecture — sits at the core of LeCun’s approach. It trains models to predict missing parts of a scene in an abstract representation space, not in raw pixels. The idea is that predicting high-level features forces the model to learn the underlying structure of reality.

Spatial intelligence, World Labs’ focus, emphasizes geometry and 3D reasoning. Then there are diffusion-based world models, which generate plausible future states of a scene by iteratively refining noisy predictions. Reinforcement learning world models simulate environments to train agents through trial and error. And finally, hybrid approaches mix several techniques.

These aren’t just academic distinctions. Each category optimizes for different trade-offs — sample efficiency versus generalization, interpretability versus raw performance, speed versus accuracy. V-JEPA 2’s ability to train on 62 hours of data suggests JEPA-style models win on sample efficiency. But whether they scale to complex, long-horizon tasks remains an open question.

Why V-JEPA 2 Matters for Robotics

Zero-shot robot planning sounds like marketing speak, but it’s actually a meaningful milestone. Most robot learning systems need task-specific training. You want the robot to pick up a mug? Fine, show it 500 examples of mug-picking. V-JEPA 2 reportedly generalizes across tasks without that per-task overhead. It builds an internal simulation of the world, then uses that simulation to reason about new problems.

Think of it like this: V-JEPA 2 is the difference between memorizing every possible chess position and understanding the rules well enough to play a new game. The former approach — task-specific training — works but doesn’t scale. The latter — learning a world model — could unlock robots that adapt to novel situations without constant retraining. That’s the theory, anyway.

But 62 hours of training data also raises questions. What kind of data? How diverse were the scenarios? How well does the model transfer beyond the distribution it saw during training? The devil lives in those details. I’d want to see benchmarks on tasks the model has never encountered, in environments with lighting, clutter, and physics edge cases that weren’t in the training set. Until then, zero-shot planning is a claim worth watching, not a solved problem.

LeCun vs. Li and the AGI Research Frenzy

The funding war between AMI Labs and World Labs isn’t just about robotics. Both teams are positioning world models as a path toward AGI — systems that reason about the world the way humans do. LeCun has argued for years that language models hit a ceiling because they don’t ground their predictions in physical reality. World models, in his view, provide that grounding.

Li’s spatial intelligence angle attacks the same problem from a different direction. Language models fail at tasks like ‘put the book on the shelf behind the plant’ because they don’t understand behind or shelf in a geometric sense. Spatial intelligence aims to fix that by teaching models to reason about 3D structure, occlusion, and physical relationships. It’s less about predicting the next word and more about predicting the next state of a scene.

The competitive context here matters because both teams are blurring the line between research labs and infrastructure companies. AMI Labs isn’t just publishing papers — it’s building models that could power physical AI products. World Labs reportedly plans to license its spatial reasoning tech to robotics and AR companies. The funding reflects a bet that world models become the new foundation layer, the way transformers became the foundation for language models.

And the stakes are enormous. Whoever cracks world models first doesn’t just win robotics — they potentially unlock autonomous vehicles, embodied AI assistants, and simulation engines for scientific research. The $1 billion-plus raised between these two companies is a down payment on that vision.

World Models and the Shift Beyond Language

World models simulate environments so AI systems can plan, predict, and reason about consequences before acting. They’re a departure from the language-model-first approach that dominated AI research from 2020 through 2025. Language models excel at predicting text, but they stumble when tasks require understanding physics, causality, or spatial relationships.

The recent funding underscores a broader shift. Investors are pouring capital into companies that treat the physical world as the training ground, not just text corpora scraped from the internet. That shift makes sense if you believe the next wave of AI applications lives in robotics, autonomous systems, and embodied agents. A chatbot doesn’t need to know that objects fall when you drop them. A robot does.

World models also promise better sample efficiency. Language models need trillions of tokens to learn anything useful. World models, in theory, can learn from far less data because they compress observations into reusable representations. V-JEPA 2’s 62-hour training time is an early proof point. If that efficiency holds across domains, world models could train on datasets orders of magnitude smaller than what GPT-5 or Gemini required.

What to Watch as World Models Scale

The first thing to monitor is whether V-JEPA 2’s zero-shot planning generalizes beyond controlled lab settings. Real-world robotics is messy — lighting changes, objects move, sensors fail. If AMI Labs can demonstrate robust performance in unstructured environments, the 62-hour training claim becomes a genuine breakthrough. If it only works in narrow scenarios, it’s an impressive demo but not a paradigm shift.

Second, watch how World Labs defines and benchmarks spatial intelligence. The concept sounds intuitive, but measuring it is tricky. Does the model understand occlusion? Can it reason about objects it can’t see? Can it predict how a scene changes when an object moves? Clear benchmarks will separate real progress from vaporware. Li’s team has the credibility to set those standards, but they need to publish results that others can replicate.

Third, keep an eye on how quickly these models move from research prototypes to commercial products. AMI Labs and World Labs raised over $1 billion, which means investors expect revenue, not just papers. The timeline from V-JEPA 2 to a shipping robotics product will reveal how far world models still have to go. If we see real deployments in 2026, the hype is justified. If we’re still watching demos in 2028, the funding was premature.

Source: radicaldatascience.wordpress.com