Key Takeaways

  • LeBrun rejects AGI and superintelligence as meaningless marketing churn that the industry cycles through every few months
  • World models predict physical state transitions, not text probabilities — a fundamentally different architecture for a fundamentally different problem
  • The bottleneck in robotics is no longer hardware; it is the missing brain that lets machines operate safely outside caged factories
  • LLMs and world models are complementary brain halves — language and physics — not competitors racing toward the same finish line

Alexandre LeBrun refuses to play the labeling game. While rivals attach "AGI" or "superintelligence" to every press release, the AMI Labs chief executive calls the terms what they are: empty vessels the industry fills, empties, and refills on a quarterly cycle. "We never used the word AGI," he told me in Seoul. "And I just noticed that nobody is using it anymore; they switched to superintelligence. Next time we'll switch to something else." He paused. "There's no good definition. What is superintelligence? I don't know. It's not a very useful word."

That skepticism is the most honest thing I have heard from a founder in months.

AMI Labs sits at the center of a quieter race: building world models that ingest physics instead of tokens. A large language model predicts the next word. A world model predicts the next state. Nudge a glass off a table and you already know it will tip and spill. That intuition — the causal logic of matter — is what LeBrun's team is trying to codify. The startup is still pre-product, but it is already courting robotics, manufacturing, and electronics giants. They understand something the chatbot builders do not: language models are useless when a robot must decide whether a falling object threatens a child.

The hardware is ready. Actuators, sensors, compute — progress in the last few months is, by LeBrun's account, incredible. But the brain is missing. Today's robots run fixed routines, completely static. They do not know context. A dancing machine at a public event recently approached and kicked a child because its control loop had no concept of "child" or "danger." That is not a hardware failure. It is an intelligence vacuum.

LeBrun does not position world models as LLM killers. He draws a parallel to the human brain: distinct circuits for language and for physical reasoning. LLMs will remain the most efficient tools for processing text. World models will supply the context that lets a machine operate in a kitchen, a construction site, a hospital corridor. They are complementary, not replaceable.

This framing matters because the industry insists on a single winner. Venture capital wants one architecture to rule them all. LeBrun's stance implies a messier, more profitable truth: the physical economy — manufacturing, logistics, healthcare, agriculture — needs a different stack than the knowledge economy. Almost every industry that touches the real world could eventually deploy robots guided by world models. The factory floor is already automated; the loading dock, the home, the street are not. "Robots are not safe right now," LeBrun said. "There's no solution for that today."

He is right. Safety in open environments requires predicting the next state of a chaotic world, not the next token in a sanitized corpus. That prediction demands physics, not probability. AMI Labs is betting its future on that distinction. The hype cycle will churn through another buzzword by winter. LeBrun will still be building the brain that keeps the glass from spilling.