AMI Labs: Ян Лекун наконец-то строит тот самый ИИ, о котором спорил годами

Yann LeCun has long remained the loudest and most authoritative dissident in the world of large language models. While all of Silicon Valley marveled at ChatGPT's ability to write poetry and code, LeCun would melancholically remind people on his social networks that these systems don't understand elementary things. They don't know how gravity works, why you can't put a cat in a microwave, and what would happen if you pushed a glass off a table.

Now Meta's chief AI scientist has decided that public criticism on social media is not enough. The emergence of AMI Labs is not just another startup in an overheated niche—it's a direct attempt to prove that the AI industry has been on a dead-end path for the last five years.

The name AMI Labs stands for Autonomous Machine Intelligence. This is a direct reference to the architecture that LeCun has been promoting in his scientific papers in recent years. His obsession is creating so-called "world models" (World Models). Instead of feeding neural networks endless terabytes of text from the internet, he wants to teach them to observe reality and build internal representations of physical laws. At Meta, he attempted to implement this through the V-JEPA project, but apparently corporate constraints and the need to maintain Llama development prevented true radicalism. In his own startup, his hands are untied, and this frightens competitors.

Who stands behind LeCun in this ambitious endeavor? Rumors and the first personnel changes point to a powerful contingent from FAIR (Fundamental AI Research)—that very Meta division that LeCun built and nurtured over a decade. This is a classic Silicon Valley scenario: when a visionary of such magnitude breaks free, the best minds follow—those tired of optimizing advertising algorithms or endlessly "aligning" chatbot behavior to meet censorship requirements. Among the potential project participants are names of researchers specializing in computer vision and reinforcement learning, which clearly hints at the physical embodiment of their future AI.

Investor interest in AMI Labs borders on mild madness. Despite the fact that the company doesn't even have a closed demo version yet, its preliminary valuation is already measured in billions of dollars. This is easy to explain: there is an acute shortage of alternative ideas on the market. Almost all current players build variations on the transformer architecture theme, but no one knows what to do when quality training data on the internet runs out. LeCun proposes a fundamentally different solution—training on video and sensory data, exactly as children or animals do. If his approach works, AMI Labs will become the foundation for the next generation of robotics.

LeCun's transition to founder status also calls into question Mark Zuckerberg's long-term AI strategy. Meta has long been a unique "safe haven" for fundamental science, allowing researchers to publish papers and share code without worrying about immediate profit. The departure of its figurehead could mean a shift in the company's direction toward more applied and commercial solutions. Zuckerberg wants to sell augmented reality glasses and AI assistants today, while LeCun is focused on solving the fundamental mystery of intelligence, which could take years.

Of course, the risk of AMI Labs' failure is enormous. The concept of "world models" looks flawless on paper and in conference presentations, but it hasn't yet shown the same explosive scalability as simple language models. LeCun will have to face incredible computational costs and the need to invent completely new methods for optimizing neural networks. However, if anyone can make a machine truly understand the physics of our world, it's the man who taught computers to "see" back in the late eighties.

The bottom line: AMI Labs is a global bet that text is not the pinnacle of AI evolution. If LeCun succeeds in creating a working "world model," today's chatbots will instantly transform into amusing but useless digital artifacts from the past.