"What is superintelligence? I don't know. It's not a useful word"

LeBrun's stance is rare in today's AI industry. OpenAI defines AGI as part of its mission, Meta renamed its research division to Superintelligence Labs last year, and nearly every major player uses these terms to attract investors and customers. LeBrun sees it differently: "There is no good definition. What is superintelligence? I don't know. It's not a very useful word," he said in an interview with TechCrunch during the ICML (International Conference on Machine Learning) in Seoul.

This is a deliberate stance, not modesty born of necessity. AMI Labs sits right at the center of the industry's newest race — the development of world models. And its CEO already predicted with a smile back in March: "World models will be the next buzzword. In six months, every company will call itself a world model company to get funding."

What is a world model and why ChatGPT isn't enough

LeBrun explains the difference between a large language model (LLM) and a world model simply: An LLM predicts the next word, a world model predicts the next state of the world. When you push a glass near the edge of a table, you know in advance it will tip over and spill. That exact intuition — a physical understanding of reality — is what a world model is meant to capture.

Importantly, LeBrun does not claim that world models will replace language models. According to him, they are "complementary, not interchangeable." He likens it to the human brain, where language and reasoning are handled by different functions: LLMs will remain the most efficient tool for working with language, while world models add context and understanding of the real world.

The technical foundation of AMI Labs is the JEPA (Joint Embedding Predictive Architecture), which LeCun designed back in 2022 — while he was still leading AI research at Meta. LeCun has long argued that language models alone will never reach true intelligence because they learn only from text, not from reality.

Robots have muscles, but they lack a brain

The biggest impact of world models is expected in robotics. According to LeBrun, today's robots merely repeat rigidly programmed routines — they are "completely static" and AI is "really stupid" in the physical world. He cites an incident where a robot demonstrating dance and kung-fu at a public event approached a child and kicked them. An AI aware of its surrounding context would have prevented such a situation.

"The hardware is very advanced, the progress in recent months is incredible — but the brain is missing," LeBrun sums up the state of humanoid robotics. A robot repeating the same movement in a factory works well already today. The problems start the moment you take it "outside into a more open environment, into a home or onto the street," where it must understand its surroundings and operate safely. "Robots aren't safe right now. There is no solution for this today," he says bluntly.

Healthcare: language models cover only 1%

LeBrun has a personal connection to medicine — before AMI Labs, he led the healthcare AI startup Nabla, where he remains chairman of the board. He compares today's AI systems to a doctor who learned only from textbooks and never did a hospital internship. Language models can be useful in medicine, but according to him they cover "only 1% of healthcare" — the rest is built on real-world experience. Nabla is also the first announced partner to gain access to AMI Labs' early models.

A billion dollars and a European footprint

AMI Labs was founded after Yann LeCun left Meta at the end of 2025. In March 2026, the startup raised $1.03 billion (roughly €890 million) at a pre-money valuation of $3.5 billion — though it had originally been seeking "only" €500 million. Investors include Bezos Expeditions, Nvidia, Samsung, Toyota Ventures, as well as Eric Schmidt and Mark Cuban; European names include Xavier Niel, Bpifrance, and the Dassault Group.

For Europe, this is a significant bet: the company is headquartered in Paris and AMI Labs promises open research — publishing scientific papers and open-source code. "We believe things move faster when they are open," says LeBrun. At a time when American labs are increasingly closing off their research, this is good news for the European AI ecosystem — including Czech researchers and developers.

Why Seoul, not Silicon Valley

According to LeBrun, a world model cannot be built in a lab — it needs real environments and industrial partners. "We need access to the real world," he explains why he sought partners in robotics, manufacturing, and electronics in Seoul. South Korea is appealing due to its combination of advanced industry (robotics, semiconductors) and speed of adoption — in June, the government announced a plan to mobilize roughly $880 billion into chips, AI data centers, and physical AI. "Korea was the fastest adopter of the internet 25 years ago," LeBrun recalls.

AMI Labs has neither a product nor a firm date for when it will introduce one. "When we're ready, we'll make a surprise," LeBrun concludes. In an industry where terms like AGI are promised at every press conference, even this is refreshing in its own way.