Fireside Chat, Interview
Fireside Chat with Yann LeCun, Executive Chairman of AMI Labs | RAISE Summit 2026
- Yann LeCun's Core Thesis: LeCun argues that Large Language Models (LLMs) are fundamentally limited to discrete symbolic sequences and cannot capture the complexity of the physical world, creating a performance gap between intellectual tasks (e.g., passing the bar exam) and physical tasks (e.g., domestic robotics, Level 5 autonomy).
- LLMs rely on autoregressive prediction of the next token, which works for text/code but fails for video because the vast majority of pixel-level details in physical reality are inherently unpredictable.
- Training systems to predict every pixel in a video is an "unsolvable task" because random, unpredictable details (e.g., specific movements of people) cannot be forecasted from prior context.
- AMI Labs and the JEPA Architecture: LeCun's new company, AMI Labs, is developing World Models using an architecture called JEPA (Joint Embedding Predictive Architecture) to bridge the physical intelligence gap.
- JEPA avoids pixel-level prediction by learning abstract representations of signals (video/sensors) and predicting the next state within that abstract space, effectively filtering out unpredictable details.
- Current iterations, such as VJEPA 2.1, can already understand video physics, detect impossible events in footage, and possess a form of common sense with minimal fine-tuning.
- The approach mimics human infant learning (approx. 9 months to grasp gravity and inertia) by training on observational data to predict the consequences of actions for autonomous planning.
- Strategic Departure from Meta: LeCun founded AMI Labs seven months ago after a strategic divergence from Meta regarding the path to Artificial General Intelligence (AGI).
- Although the project "AMI" (Advanced Machine Intelligence) was initially supported by Mark Zuckerberg and internal leadership, Meta pivoted in 2025 to prioritize scaling LLMs to catch up with the industry.
- LeCun publicly opposed the "scaling LLMs" hypothesis, arguing it would not yield systems capable of interacting with the physical world.
- Meta's business model (B2C social connectivity) was incompatible with AMI Labs' intended focus on B2B industrial applications (e.g., jet engine control, pharmaceutical manufacturing, complex system modeling).
- Data and Training Scale Disparities: Visual data offers a more efficient and dense training medium for physical reasoning than text data.
- Publicly available text contains approximately 100 trillion tokens (10^14 bytes), which would take a human 400,000 years to read.
- In contrast, a four-year-old human's visual cortex receives an equivalent amount of data (bytes) simply by watching the world, enabling the learning of 3D space, inertia, and object permanence.
- A cat, with only 800 million neurons (1/100th of a human brain), can master these physical principles, yet they remain a significant challenge for current computer systems.
- AI Sovereignty and Project Tapestry: LeCun advocates for "Project Tapestry," a distributed, open-source initiative to build sovereign foundation models without sharing raw data.
- The project aims to create a global model by aggregating parameters from diverse contributors (nations, universities, private entities) who train locally on their own data without transmitting the data itself.
- This approach addresses the "closed" nature of Western AI (Meta, Google) and the geopolitical risks of relying on Chinese open-source models due to potential political shutdowns or cultural bias.
- Tapestry seeks to preserve linguistic and cultural diversity, which LeCun argues is at risk if global information diets are mediated by a handful of proprietary US or Chinese engines.
- AMI Labs Global Structure: AMI Labs is structured as a neutral, global entity to maintain independence from US-China geopolitical tensions.
- Headquarters are in Paris, with offices in New York, Montreal, and Singapore; the company explicitly has zero presence on the US West Coast or in China.
- Investor distribution is globally balanced: 40% European, 33% US, and 27% Asian (including a fund from the UAE).
- Future Economic and Societal Impact: Success in World Models is predicted to unlock significant sectors currently stalled by the limitations of LLMs.
- Potential breakthroughs include Level 5 self-driving cars, reliable domestic robots, and agentic systems capable of reliable physical planning.
- LeCun asserts that Europe has a competitive advantage in this specific "world model" race, contrasting it with the "trench" where Silicon Valley is focused solely on scaling LLMs.
- The ultimate goal is an AI revolution where machines can perceive, reason, and plan actions in the real world, rather than just processing text.