Current artificial intelligence systems are fundamentally incapable of understanding the physical world, according to Yann LeCun, a leading figure in the technology industry who spoke on the sidelines of the VivaTech conference in France, BBC News reported. The former chief AI scientist at Meta asserted that modern robots possess significantly less understanding of their physical surroundings than a common rat. To address this severe limitation, LeCun founded a new startup named Advanced Machine Intelligence Labs, aiming to move the industry beyond the restrictive capabilities of popular generative models.
LeCun departed from Facebook parent company Meta in 2025 to launch the Paris-based initiative. His primary objective is to develop a new computational architecture that diverges entirely from the foundational technology powering current market leaders like ChatGPT, Claude, and Gemini. While acknowledging the immense utility of these large language models for specific digital tasks like generating written text or compiling software code, he maintains they are not a viable path toward human-like or even animal-like intelligence. These conversational systems struggle significantly when confronted with complex real-world data because they were never structurally designed to process chaotic physical environments.
Large language models essentially accumulate massive amounts of digitized knowledge and regurgitate it on command, but they lack any underlying comprehension of the physical material, LeCun explained to reporters. He illustrated this critical flaw by describing a simple physics problem involving a pen held upright on its tip on a flat surface. While any human toddler intuitively knows the pen will fall when released, they also instinctively understand that predicting the exact direction of the fall is an impossible task. Conversely, a large language model would likely attempt to generate a precise prediction based entirely on statistical patterns, demonstrating a complete lack of genuine reasoning regarding physical reality.
To overcome these structural hurdles, Advanced Machine Intelligence Labs is aggressively developing a novel system called Joint Embedding Predictive Architecture. This framework creates digital abstractions of the real world, allowing the artificial intelligence to filter out irrelevant visual noise and focus exclusively on useful environmental data. By mathematically assessing the potential outcomes of physical actions rather than just generating words, the system can help machines navigate unpredictable situations without relying on flawed statistical guesses. What remains unclear is exactly how soon this technology can be widely deployed in commercial hardware.
Investors have nonetheless shown immense financial confidence in this alternative approach to machine learning. Earlier this year, the start-up secured more than one billion dollars in a massive seed funding round, making it one of the largest early-stage investments in European technology history, Bloomberg confirmed. High-profile financial backers include the American computer chip manufacturer Nvidia and the private wealth fund managed by Amazon founder Jeff Bezos.
Building this flexible, physically aware type of artificial intelligence has become an urgent, multi-billion-dollar priority for the broader robotics industry. Enormous amounts of capital are currently being poured into the development of bipedal humanoid robots by various tech conglomerates. However, programming these complex machines to safely execute basic household chores, such as ironing clothes or successfully loading a kitchen dishwasher, remains an incredibly difficult and expensive programming challenge. Current language models are largely useless for applied robotics, making the success of systems like the one being developed in Paris absolutely critical for the future of automated physical labor.
