The artificial intelligence industry, currently captivated by the rapid advancement of large language models (LLMs) such as ChatGPT and Gemini, faces a fundamental critique from one of its own architects. Yann LeCun, a former chief AI scientist at Meta and a Turing Award winner, has publicly dismissed these systems as "not particularly smart," arguing they lack the underlying reasoning necessary for real-world interaction. LeCun, who left Meta in 2025 to found Advanced Machine Intelligence Labs (AMI Labs) in Paris, contends that LLMs excel only in narrow, well-defined domains like coding and text generation, but fail to grasp the unpredictable complexity of physical environments. His departure from Meta and the launch of AMI Labs signal a growing schism in the AI community between those pursuing incremental LLM improvements and those seeking a paradigm shift toward embodied intelligence.
At the heart of LeCun's critique is a fundamental limitation of current AI architecture. LLMs operate by predicting the next word or token in a sequence based on statistical patterns from vast training datasets—a process LeCun likens to "regurgitation" rather than genuine comprehension. To illustrate this, he points to a simple physical scenario: holding a pen upright and releasing it. While a toddler intuitively understands the pen will fall, an LLM might attempt to predict the exact direction of the fall based on improbable statistical correlations, yielding an incorrect answer. This inability to model causal relationships and physical dynamics, LeCun argues, prevents LLMs from achieving even animal-level intelligence, let alone human-like reasoning. His perspective underscores a broader industry challenge: how to build AI systems that can navigate the messy, high-dimensional variables of the real world, such as those required for household robotics or autonomous navigation.
AMI Labs is betting on a novel architecture called Joint Embedding Predictive Architecture (JEPA) to overcome these hurdles. Unlike LLMs that process raw data directly, JEPA creates abstract representations of the world, enabling the system to predict outcomes of actions without needing to simulate every possible detail. This approach, which requires advanced mathematical frameworks, aims to model the latent variables that govern real-world events—such as gravity, momentum, or object permanence—allowing AI to reason about cause and effect. LeCun envisions JEPA as a foundation for "world models" that can learn from observation and interaction, much like animals do. If successful, this could unlock breakthroughs in robotics, autonomous systems, and any application demanding robust physical understanding, potentially leapfrogging current LLM-based approaches that remain brittle in dynamic environments.
The financial community has taken note of this ambitious vision. In early 2026, AMI Labs announced a seed funding round exceeding $1 billion, one of the largest of its kind in European history, with backing from Nvidia and Bezos Expeditions, the investment vehicle for Amazon founder Jeff Bezos. This massive capital injection reflects investor appetite for diversifying AI bets beyond the LLM-dominated landscape, particularly as concerns about diminishing returns from scaling larger models grow. However, significant technical hurdles remain: JEPA must prove it can scale to complex tasks, and AMI Labs faces competition from other labs pursuing alternative paradigms, such as reinforcement learning or neuro-symbolic AI. LeCun's critique of LLMs may be controversial, but it has sparked a necessary debate about whether the industry's current trajectory can truly deliver on the promise of general intelligence, or whether a more radical departure is required.