The French Engineer Who Refuses to Chase the Ghost of Superintelligence

Alexandre LeBrun sits in a quiet Parisian cafe, worlds away from the neon-lit hype cycles of San Francisco. While tech executives across the Atlantic compete to see who can promise the end of human labor first, LeBrun is busy with something much more mundane, and yet infinitely more difficult. He is trying to teach a computer how a coffee cup falls.

As the leader of AMI Labs, a startup spun out of Meta’s artificial intelligence research division, LeBrun works closely with Turing Award winner Yann LeCun. Together, they are building what they call world models. It is a quiet pursuit, far removed from the theatrical product launches that dominate the evening news. And unlike his peers, LeBrun has no interest in selling you a digital deity.

The Vocab War of Silicon Valley

Walk through the offices of any major AI lab today, and you will hear a specific vocabulary spoken with religious fervor. Founders talk of artificial general intelligence as if it is a train pulled into the station just behind schedule. They speak of superintelligence with a mix of fear and greed, painting pictures of systems that can solve cancer by lunchtime and rewrite the laws of physics by dinner.

LeBrun refuses to use these words. To him, they are marketing terms masquerading as science, designed to thrill venture capitalists rather than solve engineering hurdles. He believes that labeling a software system as "super" before it can even comprehend basic physical cause and effect is putting the cart before a horse that hasn't even been born yet.

"We are building tools that understand the friction of the real world, not digital gods that live in a vacuum of text."

The current crop of large language models can write beautiful poetry and pass legal exams. Yet, they remain fundamentally detached from reality. They do not know that if you pull a tablecloth, the glasses on top might shatter. They merely know which words usually follow other words.

Building a Map of the Physical World

AMI Labs is taking a different path. Instead of feeding billions of pages of internet text into a massive cluster of servers, LeBrun’s team is focused on teaching machines to observe and predict the physical environment. They want to build systems that possess common sense, the kind of intuitive knowledge a toddler develops just by dropping toys from a high chair.

This approach requires a shift in how we think about machine learning. Instead of predicting the next word in a sentence, these systems must predict the next frame in a video of the real world. Under the hood, this means developing algorithms that can handle uncertainty and understand gravity, friction, and human intent without explicit instructions.

For developers and founders, this distinction matters. A text-based assistant can write a marketing email, but it cannot navigate a robot through a crowded hospital corridor or manage a complex supply chain during a storm. By focusing on world models, AMI Labs aims to create software that can safely interact with the physical world we actually inhabit.

The Danger of the Magic Trick

The danger of the current hype, LeBrun argues, is that it creates a false sense of capability. When a chatbot speaks with perfect grammar, our brains naturally assume it possesses human-level comprehension. We attribute intelligence to what is essentially a highly sophisticated mirror of our own writing.

This illusion can lead to costly mistakes for businesses trying to implement these systems. A startup founder might deploy an AI agent to handle customer logistics, only to find the system hallucinating delivery routes that defy geography. By stripping away the mystical vocabulary, LeBrun hopes to ground the industry in what is actually possible today.

Back in Paris, the afternoon light fades as LeBrun finishes his espresso. The road ahead for world models is long, filled with complex mathematical hurdles that cannot be solved by simply adding more data centers. But as the noise of the AI gold rush continues to grow, a quiet, methodical focus on the laws of physics might be exactly what the industry needs to find its footing.