A Billion for 12 People: Why Yann LeCun Rejects Language Models

One billion dollars in seed investment for a company with just twelve employees. Even by Silicon Valley standards, where massive funding rounds have become commonplace, this figure seems anomalous. However, the situation becomes clear when the founder's name emerges. Behind the AMI Labs startup stands Yann LeCun — one of the pioneers of deep learning, a Turing Prize laureate, and a man who spent years shaping artificial intelligence strategy at Meta. This colossal funding round serves not merely as yet another testament to venture capital's boundless faith in AI, but stands as a resounding declaration that the industry's current development trajectory may have reached a conceptual dead end.

To understand the scale of AMI Labs' ambitions, one must examine the fundamental limitations of contemporary technologies. In recent years, the entire technology world has been mesmerized by the magic of large language models. Systems built on transformer architecture have learned to brilliantly mimic human speech, write code, and pass complex academic exams.

Yet behind this impressive facade lurks an algorithm that, at its core, simply predicts the next word in a sentence. LeCun has long and consistently criticized autoregressive models, pointing out that they lack genuine understanding of the physical world, are utterly devoid of common sense, and are incapable of long-term planning. Their knowledge is superficial, and their fundamental tendency toward hallucination cannot be remedied by simply increasing the volume of training data or scaling raw computational power.

The concept that AMI Labs proposes to develop is radically different from the paradigm that dominates today. Rather than forcing algorithms to consume trillions of text tokens from the internet, LeCun is betting on the creation of so-called "world models." The idea is to teach artificial intelligence to perceive reality the way biological organisms do — through observation, understanding of physical laws, causal relationships, and hierarchical planning.

This approach is founded on an architecture that focuses on extracting abstract representations from sensory data, rather than pixel-by-pixel or character-by-character information recovery. This allows the system to ignore unimportant details and concentrate on what truly matters. For example, when a person watches a glass fall from a table, they don't need to calculate the exact trajectory of each potential shard to understand that the glass will break.

Modern language models lack such intuitive physics, which makes their application in real robotics or autonomous vehicles severely limited. AMI Labs intends to overcome this barrier by creating an architecture oriented toward achieving specific goals in complex and unpredictable physical environments.

The fact that investors are prepared to entrust a billion dollars to a microscopic team for developing an alternative architecture represents a crucial signal for the entire industry. It means venture capital is beginning to doubt the concept of scaling for scaling's sake. For years, it was believed that the path to strong AI lay exclusively through creating ever-larger clusters of graphics processors.

Now, however, the focus is shifting from raw computational power to fundamental scientific breakthroughs. If new architectures prove to be more energy-efficient and less demanding of large training datasets, the industry's massive dependence on enormous data centers could diminish. Investors understand that the current development trajectory faces not only algorithmic limitations but also an energy barrier, as well as a growing shortage of quality data.

Finding a workaround becomes a matter of survival for the entire sector.

The success or failure of AMI Labs will determine the vector of artificial intelligence development for the next decade. If LeCun's team can prove the viability of world models and create an effective alternative to the transformers that dominate today, it will lead to a complete overhaul of the industry's technological stack. The market will witness a transition from passive systems that simply respond to text queries in a browser window to active autonomous agents capable of safely and predictably interacting with the real world.

Ultimately, a billion dollars for a team of twelve visionaries may turn out not to be a display of investment madness, but the most forward-thinking bet in the history of the technology market, opening the door to truly thinking machines.