When Yann LeCun quietly left Meta in November 2025, where he had spent over ten years building the FAIR (Fundamental AI Research) laboratory, few suspected what he was up to. The answer came in January 2026: together with entrepreneur Alexandre LeBrun, he founded AMI Labs — Advanced Machine Intelligence Labs — in Paris. And in March 2026, the whole world learned that the new company was not just another AI startup. It is a direct bet against the dominant trend of the entire industry.

A Billion for Doubting ChatGPT

On March 9, 2026, AMI Labs announced the closing of a seed round totaling 1.03 billion dollars at a pre-investment valuation of 3.5 billion dollars. This is historically the largest seed investment in Europe. The round was led by Cathay Innovation, Greycroft, Hiro Capital, HV Capital, and Bezos Expeditions — Jeff Bezos's personal investment fund. Individual investors include names like Eric Schmidt (former Google CEO), Mark Cuban, Jim Breyer, Xavier Niel, and Tim Berners-Lee, the inventor of the web.

This constellation is no coincidence. These are people who believe that behind the current facade of chatbots and language models lies a fundamental limitation — and that LeCun knows how to circumvent it.

LeCun, after all, is not expressing his stance for the first time. For years, he has publicly argued that large language models (LLMs) like GPT-4 or Claude cannot lead to artificial general intelligence (AGI). According to him, these are systems that have learned statistical patterns in vast amounts of text but do not understand the physical world, causes and effects, or how things work. "A cat knows more about the physical world than any LLM," says LeCun.

What are world models and why does LeCun consider them key?

AMI Labs is working on so-called world models. While a language model predicts the next word in text, a world model learns to predict what will happen in a real environment if I take a certain action. It thinks in terms of physics, causality, and abstraction — not in terms of tokens.

The technical foundation is the JEPA (Joint Embedding Predictive Architecture) architecture, which LeCun proposed back in 2022 and published within the scientific community. Unlike classical models that try to reconstruct input data (every pixel, every token), JEPA operates in a compressed abstract space. It learns representations, not raw reconstruction. This has fundamental impacts on both efficiency and capabilities.

Specific advantages over LLMs according to AMI Labs:

• Hierarchical planning — the system can plan in abstract space, not just react to prompts
• Causal understanding — the model understands what causes what, not just what statistically follows what
• Actionability — world models are designed for environments where AI must act, not just generate text
• Efficiency — JEPA does not need to predict unnecessary details, it focuses on what is relevant

AMI Labs' target application areas are robotics, healthcare, industrial automation, and safety-critical systems — places where an LLM error can have real physical consequences.

A European Startup in an Era of American Dominance

AMI Labs is a Parisian startup — and that's not just a geographical note. At a time when the AI industry is dominated by American corporations (OpenAI, Anthropic, Google DeepMind, Meta AI) and Chinese laboratories (DeepSeek, Zhipu AI), the emergence of a competitive European AI company with a billion-dollar valuation is exceptional.

The company's leadership reflects its ambition: alongside LeCun as Executive Chairman, CEO Alexandre LeBrun, who previously led the AI division at Facebook, is at the helm. COO is Laurent Solly, former Vice President of Meta for Europe. The Chief Scientist is Professor Saining Xie from NYU, the Director of Research is Pascale Fung from HKUST, and the Vice President for World Models is Michael Rabbat, who led research teams at Meta.

For the Czech and European AI scene, the emergence of AMI Labs has a direct impact: Europe finally has skin in the game in the top league of AI research. Furthermore, AMI Labs benefits from the European regulatory framework — research is conducted under GDPR and in compliance with the EU AI Act, which can be an advantage when deploying in European companies and institutions.

Is it a Bet, or a Breakthrough?

Critics argue that world models and JEPA are fascinating theories, but so far without a product, without a public benchmark, and without a demonstration of capabilities comparable to GPT or Claude. LeCun himself admitted that the path would be long — AMI Labs is described as a "long-term scientific bet."

On the other hand: LeCun is no newcomer. His work on convolutional neural networks (CNNs) in the 80s and 90s laid the foundations for modern deep learning. The Turing Award in 2018 (along with Geoffrey Hinton and Yoshua Bengio) is an accolade in the AI world equivalent to the Nobel Prize. And 1.03 billion dollars from the world's elite investors is not just blind faith — it's a conviction that the paradigm can change.

The timing is also interesting. AMI Labs emerges at a time when LLM scaling was hitting limits — capability gains are getting smaller with ever-increasing costs. Many researchers quietly admit that "next token prediction" might not be the ultimate answer. LeCun says the same out loud and builds a company on it.

What does this mean for users and businesses today?

In the short term — nothing. AMI Labs has no product, API, or waiting list. It is a research organization with a long horizon. But in the medium and long term, it is one of the most significant experiments in AI: if world models work as LeCun claims, it will open the way to AI that truly understands the world — and which could be deployed in robotics, medical diagnostics, or industrial production in ways that current LLMs do not allow.

For Czech companies following AI development: AMI Labs is a reason to also follow non-American research. If Europe can build a globally competitive AI ecosystem — and AMI Labs is the first serious proof that it's not impossible — it will have a direct impact on how AI will be regulated, licensed, and deployed in the EU.