Yann LeCun argues that large language models (LLMs) are fundamentally limited and cannot achieve true intelligence because they lack real-world understanding and experiential grounding, relying solely on text prediction. He proposes developing AI systems that integrate sensory experiences and world models, enabling deeper reasoning and adaptability beyond what LLMs can offer.
Yann LeCun, a prominent AI researcher, has expressed strong skepticism about the future of large language models (LLMs) as the path to true artificial intelligence. He argues that simply scaling up LLMs will not lead to genuine intelligence, despite the optimistic claims from major tech companies like Google, OpenAI, Meta, and Microsoft. These companies often suggest that artificial general intelligence (AGI) is just around the corner, but LeCun believes this is more about marketing and investor hype than scientific reality.
LeCun’s main critique is that LLMs are fundamentally limited because they only predict the next word in a sequence of text. This narrow focus means they lack a broader understanding of context, intent, and the world at large. Unlike a human teacher, who tailors explanations based on a student’s background and needs, LLMs respond in a generic way, unable to adapt meaningfully to individual users or situations. LeCun points out that real intelligence involves understanding and interacting with the physical world, not just manipulating words.
He further explains that text alone is an insufficient medium for learning and reasoning. Humans learn through a combination of sensory experiences—taste, smell, touch, and sight—not just language. For example, knowing the word “apple” doesn’t convey what an apple tastes or smells like. LLMs, therefore, lack the experiential grounding necessary for true understanding and decision-making. This limitation prevents them from planning or simulating future outcomes in a meaningful way.
LeCun also warns that the apparent intelligence of LLMs is largely an illusion. While they may excel at standardized tests or generate impressive-sounding answers, they often fail in real-world applications where deeper understanding and adaptability are required. When faced with novel problems or the need to go beyond surface-level responses, LLMs tend to repeat mistakes or require explicit guidance, revealing their lack of genuine comprehension and foresight.
To address these shortcomings, LeCun has left Meta and started a new company called Advanced Machine Intelligence. His approach focuses on building AI systems with world models that integrate information beyond text, allowing for richer representations of reality. He advocates for architectures like Joint Embedding Predictive Architecture (JEPA), which enable AI to predict abstract representations and simulate multiple future outcomes. This, he believes, will allow AI to plan proactively and make better decisions, especially in physical and agentic contexts such as robotics, moving closer to real intelligence rather than just the appearance of it.