Yann LeCun, the former chief AI scientist at Meta, has launched a new startup, AMI Labs, with over $1 billion in seed funding to develop artificial intelligence that understands the physical world. Supported by investors including Nvidia and Jeff Bezos’s private wealth fund, the Paris-based company aims to move beyond the limitations of Large Language Models (LLMs) to create systems capable of complex real-world reasoning.
Why do current AI models struggle with the physical world?
Current leading AI systems, such as ChatGPT, Claude, and Gemini, are built on Large Language Models (LLMs) that excel at processing text, code, and mathematics. However, LeCun argues these models lack a fundamental understanding of physical reality.
Speaking at the VivaTech technology conference in France, LeCun stated that LLMs are not a path toward human-like or even animal-like intelligence. He explained that these systems cannot deal with real-world data because they are not built for it. According to LeCun, LLMs essentially accumulate and “regurgitate” knowledge based on statistical patterns rather than possessing an underlying understanding of how the world functions.
While LLMs can solve well-defined and predictable problems, they struggle when faced with the unpredictable outcomes of the physical world. This limitation prevents current AI from performing tasks that require physical intuition, such as a robot managing household chores.
The seed funding round for AMI Labs, which exceeded $1 billion, is one of the largest of its kind in Europe’s startup history.
How does JEPA architecture solve the reasoning problem?
To overcome the limitations of LLMs, AMI Labs is developing a new architecture called Joint Embedding Predictive Architecture (JEPA). Unlike LLMs, which predict the next word or token in a sequence, JEPA is designed to create abstractions of the real world.
These abstractions allow the AI to assess the potential outcomes of actions by filtering out irrelevant information. This process involves complex mathematics to ensure the system focuses only on useful data points about its environment.
LeCun illustrated this distinction using the example of a pen held upright on its tip. He noted that while a human or a toddler understands a pen will topple, an LLM might attempt to predict the specific direction of the fall based on statistical patterns from its training data. Because the LLM is not reasoning about physical reality, its prediction would likely be wrong. JEPA, however, would recognize that predicting the exact direction is unnecessary and would instead focus on the broader physical context.
Comparing AI Approaches: LLM vs. JEPA
| Feature | Large Language Models (LLMs) | JEPA (AMI Labs) |
|---|---|---|
| Primary Function | Statistical pattern regurgitation | Predicting world abstractions |
| Data Type | Text, code, and math | Physical world/sensory data |
| Real-World Goal | Information generation | Physical reasoning and robotics |
What is the scale of investment in AMI Labs?
Investors have signaled significant confidence in LeCun’s vision for a new type of intelligence. Earlier this year, AMI Labs announced it had raised more than $1 billion in a seed funding round. This investment makes it one of the most heavily capitalized early-stage startups in the European technology sector.
Key participants in the funding round include:
- Nvidia: The US-based computer chip giant.
- Jeff Bezos’s wealth fund: The fund managing the private assets of the Amazon founder.
This capital is intended to support the development of the JEPA architecture, moving the company toward the goal of creating AI that can navigate and interact with the physical world as effectively as biological organisms.
When evaluating AI startups, distinguish between “Generative AI” (content creation) and “Physical AI” (robotics and reasoning). The latter requires fundamentally different architectural approaches like JEPA.
Can AI achieve animal-like intelligence?
A central theme of LeCun’s work is the gap between current machine intelligence and biological intelligence. He has frequently pointed out that even a rat possesses a better understanding of the physical world than current AI models.
The goal of AMI Labs is to close this gap. By building systems that don’t just predict the next word but actually model the consequences of physical actions, the company hopes to unlock the next frontier of automation: robots that can handle the complexities of human environments, such as kitchens and living rooms, without constant reprogramming.
Frequently Asked Questions
What is AMI Labs?
AMI Labs is a Paris-based artificial intelligence startup founded by Yann LeCun to develop new AI architectures that understand physical reality.
What is the difference between LLMs and JEPA?
LLMs use statistical patterns to predict text or code, whereas JEPA (Joint Embedding Predictive Architecture) creates abstractions to reason about physical outcomes and world models.
Who is Yann LeCun?
Yann LeCun is a prominent AI scientist and the former chief AI scientist at Meta, known for his work in deep learning.
Why is $1 billion considered a large seed round?
A seed round is typically the earliest stage of startup funding. A $1 billion injection at this stage is exceptionally rare and indicates massive investor interest in the underlying technology.
How can we engage with this topic?
Do you believe AI will ever truly master the physical world, or are we limited by the data we provide? Share your thoughts in the comments below or subscribe to our newsletter for more updates on the future of artificial intelligence.
