Yann LeCun Says AI-Powered Robots Are Not As Good As Rats. So What’s Next for Advanced Technocrats? 

Is an evolutionarily modeled rat more intelligent than probabilistically modeled artificial intelligence? Yann LeCun, a pioneering figure in artificial intelligence, knows the answer. He believes that current AI-powered robots cannot understand the physical world as well as a rat. After serving as Meta’s chief AI scientist for a decade, LeCun left in 2025 to found Paris-based Advanced Machine Intelligence Labs, known as AMI Labs. His new venture aims to move AI beyond existing platforms like ChatGPT, Claude, and Gemini. He argues that while large language models, or famously called LLMs, excel at predictable tasks like coding or generating text, they lack an underlying understanding of physical reality and cannot handle real-world data effectively. 

To solve this, AMI Labs is developing Joint Embedding Predictive Architecture, or JEPA. This system filters out useless information to create math-based abstractions of the real world, allowing the AI to predict physical outcomes accurately. Investors see massive potential in this approach. AMI Labs recently raised over one billion dollars in a historic seed funding round, drawing capital from chipmaker Nvidia and Jeff Bezos’s private wealth fund. 

Many computer scientists agree that existing language models are largely hopeless for robotics because they rely on statistical patterns rather than logical reasoning. For instance, if you release an upright pen, a toddler knows it will fall, but a language model might make a specific, incorrect prediction about the exact direction based purely on training data statistics. 

Consequently, the robotics industry is actively exploring alternative systems known as World Models. Ingmar Posner, a professor of Applied Artificial Intelligence at Oxford University, directs a research team developing a mechanistic world model to help machines organize, recall, and modify knowledge efficiently. This methodology builds on earlier concepts, such as Google’s Dreamer model, which successfully gathered diamonds in the video game Minecraft by imagining future scenarios. Other groups, including DeepMind, Wayve, and Fei-Fei Li’s World Labs, are also engineering competitive systems. AMI Labs plans to test its model in industrial settings next year, aiming to eventually deploy generic intelligence systems across various global platforms.