Breaking News: Yann LeCun Leaves Meta for World Models
AI pioneer Yann LeCun left Meta in November 2025 to launch Advanced Machine Intelligence Labs (AMI Labs), a world-models-focused startup targeting a $3+ billion valuation. This is the biggest signal yet that the AI industry is moving beyond LLMs.
Large language models hit a wall. That's the controversial claim from one of AI's founding fathers - and he's betting billions on an alternative called "world models."
If you're a founder watching the AI space, this is the most important shift happening right now. While everyone is focused on GPT-5 and Claude 4, the smartest researchers believe the real path to truly intelligent AI runs through world models - systems that understand physics, space, and causality like humans do.
This guide explains what world models are, why they matter, and what it means for founders building with AI in 2026.
What Are World Models?
World models are AI systems that understand how the physical world works - concepts like gravity, inertia, object permanence, and cause-and-effect relationships. Unlike LLMs that predict the next word, world models predict the next state of a physical environment given actions taken within it.
As Yann LeCun explains: "You can imagine a sequence of actions you might take, and your world model will allow you to predict what the effect of the sequence of actions will be on the world."
The Core Difference
LLMs predict text: "Given these words, what word comes next?"
World models predict reality: "Given this situation and these actions, what happens next in the physical world?"
World models learn by watching videos and processing spatial data, building internal representations of how objects move, interact, and behave. They understand that balls fall when dropped, that cars need roads, and that pouring water fills a glass.
Why LLMs Aren't Enough
LeCun has been vocal about LLMs' fundamental limitations:
1. No True Understanding of Reality
LLMs learn statistical patterns in text. They can describe how a ball bounces, but they don't actually understand the physics. Ask ChatGPT to predict a complex physical scenario, and it often fails in ways a child wouldn't.
2. Hallucination Is Structural
Because LLMs predict what text typically follows, not what's actually true, hallucination isn't a bug - it's how they work. World models, trained on actual physical interactions, have a ground truth to check against.
3. Can't Learn from Few Examples
Humans and animals build world models that let them generalize from minimal experience. A child who sees one ball bounce can predict how any ball will bounce. LLMs need millions of examples and still struggle with novel situations.
4. No Real Planning
True planning requires simulating futures - imagining "what if I do X, then Y happens, then I could do Z." LLMs can describe plans but can't actually simulate outcomes. World models enable this kind of mental simulation.
| Capability | LLMs (GPT-4, Claude) | World Models |
|---|---|---|
| Text generation | Excellent | Not the focus |
| Physical reasoning | Poor | Excellent |
| Planning/simulation | Describes plans | Simulates outcomes |
| Generalization | Needs many examples | Few-shot learning |
| Robotics/embodiment | Cannot control robots | Core application |
| Video understanding | Limited | Native capability |
The Story Behind AMI Labs
Understanding why LeCun left Meta reveals a lot about where AI is heading.
The Meta Fallout
After more than a decade leading Meta's Fundamental AI Research (FAIR) lab, LeCun clashed with Mark Zuckerberg's new direction. When Meta CEO pivoted to rapid product development to compete with OpenAI and Google, he reorganized under "Superintelligence Labs" led by Alexandr Wang (former Scale AI CEO). LeCun was forced to report to Wang - nearly four decades his junior.
The breaking point came with Meta's Llama 4 release in April 2025, which LeCun described as "dead on arrival." He later acknowledged that benchmark results were manipulated. Following the admission, Zuckerberg lost confidence in LeCun's team, and the AI godfather decided to leave.
The AMI Labs Vision
LeCun's new company, Advanced Machine Intelligence Labs, is betting everything on world models. Key details:
- Valuation: Targeting $3+ billion
- Technology: Built around V-JEPA (Video Joint-Embedding Predictive Architecture), an architecture LeCun created at Meta
- Leadership: Alex LeBrun as CEO, LeCun as Executive Chair
- Location: Headquarters in Paris, deliberately outside Silicon Valley
LeCun believes "the groupthink around LLMs is so strong in the Bay Area that genuine innovation needs geographic distance."
V-JEPA Explained
V-JEPA learns from video instead of text. Instead of learning from text alone, these models train on video and spatial data to understand how the physical world works. They can plan, reason, and hold onto information over time - capabilities that text-trained LLMs fundamentally lack.
Major Players in the World Models Race
AMI Labs isn't alone. Here are the key players building world models in 2026:
AMI Labs (Yann LeCun)
$3B+ valuation target | Paris-based
The biggest bet on the thesis that LLMs will never achieve AGI. Building V-JEPA-based world models for understanding physical reality and enabling robot intelligence.
Google DeepMind - Genie 3
Released August 2025
The first real-time interactive general-purpose world model. Genie 3 produces navigable 3D worlds at 24 frames per second - unlike previous systems that generated static environments or required significant processing time.
NVIDIA Cosmos
Launched CES 2025 | 2M+ downloads by Jan 2026
Platform for physical AI development, specifically targeting autonomous vehicles and robotics. Cosmos world foundation models generate synthetic training data for robots and self-driving cars.
World Labs (Fei-Fei Li)
$1B valuation | Founded late 2024
Stanford AI legend Fei-Fei Li's startup focused on "spatial intelligence" - AI that understands 3D space and can interact with the physical world.
OpenAI Sora
Video generation model
While primarily marketed for video generation, Sora's underlying architecture gives it world-modeling capabilities - it understands physics and spatial relationships to generate realistic video.
How World Models Work (Technical Overview)
For founders who want to understand the technology:
1. Learning from Video
World models watch massive amounts of video footage - everything from YouTube videos to robotics simulations. They learn to predict what happens next in any given scene, building an understanding of physics and causality.
2. Compressed Representations
Rather than storing raw video, world models create abstract, conceptual representations of scenes. This "latent space" captures what matters for reasoning - object positions, velocities, relationships - without the noise of pixel-level detail.
3. Predictive Simulation
Given a current state and a proposed action, world models simulate the consequences. This allows for planning: "If I move the robot arm here, then there, what happens?" The model runs the simulation internally before acting.
4. Multi-Modal Integration
World models combine vision, spatial understanding, and sometimes language into a unified system. This enables tasks like "put the red block on the blue block" - understanding language, identifying objects, planning movements, and predicting outcomes.
Key Architecture: Joint Embedding Predictive Architectures (JEPA)
LeCun's approach doesn't try to predict every pixel of future video frames. Instead, JEPA predicts abstract representations of the future - what's important for reasoning and planning. This is more efficient and more aligned with how humans think about the world.
Real-World Applications for Founders
World models aren't just research projects - they're enabling real products:
Robotics and Automation
World models are essential for robots that operate in unstructured environments. A warehouse robot with world model understanding can handle novel situations - a fallen box, a new product shape, an unexpected obstacle - without explicit programming.
- Warehouse automation: Robots that understand object physics for picking and placing
- Manufacturing: Assembly systems that adapt to variations
- Service robots: Household helpers that navigate real-world mess
Autonomous Vehicles
Self-driving cars need to predict what other drivers, pedestrians, and objects will do. World models enable this by simulating possible futures and planning safe routes.
Video Generation and Editing
Products like Sora, Runway, and Pika use world model understanding to generate physically realistic video. Understanding physics means generated content looks right - objects fall correctly, liquids pour realistically, reflections behave properly.
Game Development and Simulation
World models can generate infinite game worlds with consistent physics. They can also power NPCs that actually understand their environment rather than following scripted behaviors.
Digital Twins
Creating accurate virtual representations of physical systems - factories, cities, supply chains - requires understanding how the real world works. World models enable predictive digital twins that can forecast issues before they happen.
AR/VR and Spatial Computing
As Apple Vision Pro and Quest devices mature, world models will power the understanding needed to seamlessly blend digital and physical reality. Virtual objects that properly occlude behind real objects, respond to real lighting, and interact with real surfaces.
What This Means for AI Founders
The rise of world models creates new opportunities and challenges:
Opportunities
- Physical AI applications: Robotics, autonomous systems, and embodied AI are finally becoming practical
- Synthetic data generation: World models can create training data for other AI systems
- Video understanding products: Analysis, search, and manipulation of video content
- Simulation and training: Virtual environments for training both humans and AI
- Design tools: Products that understand physical constraints (architecture, product design, engineering)
Challenges
- Compute requirements: World models need even more compute than LLMs
- Data needs: Training requires massive video datasets with physical interactions
- Infrastructure gap: The tooling ecosystem is much less mature than LLMs
- Talent scarcity: Few researchers have world model expertise
Don't Abandon LLMs Yet
World models complement LLMs, they don't replace them. LLMs remain the best choice for text-based tasks, reasoning about language, and general-purpose AI assistants. The future is multi-modal AI that combines LLM language understanding with world model physical understanding.
How to Get Started with World Models
For founders interested in exploring world models:
1. Experiment with NVIDIA Cosmos
NVIDIA's Cosmos platform is the most accessible way to work with world foundation models. It's designed for developers building physical AI applications like robotics and autonomous vehicles.
2. Follow the Research
Key papers and resources:
- LeCun's V-JEPA papers from Meta AI
- Google DeepMind's Genie papers
- NVIDIA's Cosmos technical documentation
- World Labs blog for spatial intelligence updates
3. Consider Adjacent Applications
You don't need to build world models from scratch. Consider applications that will benefit as world models mature:
- Tools for training robots using simulated environments
- Video analytics platforms that understand object physics
- Content creation tools that leverage world model video generation
- Testing and validation platforms for autonomous systems
4. Watch the Infrastructure Layer
As world models grow, so will the supporting infrastructure. Opportunities exist in:
- Video data curation and labeling
- Simulation environments and synthetic data generation
- Inference optimization for real-time world model applications
- Evaluation and benchmarking tools
Timeline: What to Expect
Here's a realistic timeline for world model development:
2026 (Now)
- AMI Labs initial products and research releases
- NVIDIA Cosmos ecosystem expansion
- Google DeepMind Genie improvements
- More startups entering the space
2027
- First commercial robotics products using world models
- World model APIs becoming accessible to developers
- Integration with existing LLM workflows
2028+
- Mainstream adoption in autonomous vehicles, robotics, video
- Consumer products with embedded world model understanding
- Potential breakthroughs in AGI if world models prove transformative
The Bottom Line for Founders
World models represent the most significant shift in AI research since the transformer architecture that enabled ChatGPT. While LLMs will continue to dominate text-based applications, the next wave of AI products - robotics, autonomous systems, video understanding, spatial computing - will be built on world models.
Yann LeCun betting his reputation and billions of dollars on this thesis is a strong signal. Smart founders are paying attention now, building expertise, and identifying opportunities before world models become as mainstream as LLMs are today.
The founders who understand both LLMs and world models will be best positioned to build the next generation of AI products.
Stay Ahead of the AI Curve
World models, agentic AI, and new architectures are reshaping what's possible. Join AI First Founders to get weekly updates on the technologies that matter for builders.