Speaking at VivaTech in Paris, Yann LeCun told the BBC that ChatGPT, Claude and Gemini are “not a path towards human level or human-like intelligence” and that current software cannot understand the physical world as well as a rat. LeCun is not a peripheral figure making a contrarian bet. He shared the 2018 Turing Award for inventing the deep learning techniques that make every major software tool today possible. He left Meta in late 2025 and has since raised $1.03 billion to build something different.
The argument LeCun makes is not that ChatGPT and similar tools are useless. He uses them himself. The argument is more specific and more damaging: that the way they are built makes it structurally impossible for them to understand the physical world, and that without physical understanding, they will never reach the level of general intelligence that Silicon Valley has been promising is just around the corner.
To understand why, you need to understand what these systems actually do. A large language model, which is the category that includes ChatGPT, Claude, Gemini and the majority of the most widely used software tools today, is trained on enormous quantities of text. It learns to predict what word is most likely to come next in a sequence, based on patterns it has seen across billions of documents. It becomes extraordinarily good at this. Good enough that its outputs look, to most human readers, like genuine understanding.
The problem, LeCun argues, is that text is a very thin slice of reality. The word “apple” appears in millions of documents alongside words like “red,” “fruit,” “Newton” and “tree.” A language model learns these associations perfectly. What it never learns is the weight of an apple, the way it gives slightly when pressed, the arc it follows when knocked off a table, or the sound it makes when it hits the floor. Those things do not live in text. They live in the physical world.
This is why, LeCun argues, a toddler who has dropped a spoon three times understands gravity better than any software system trained on all the physics textbooks ever written. The toddler has built a model of the world from direct experience. The software has built a model of how humans describe the world in language. These are not the same thing.
The consequences of this gap are most visible in robotics. Despite years of investment and genuine technical progress, there is still no robot capable of reliably loading a dishwasher in an unfamiliar kitchen, folding laundry from an unsorted pile or navigating a construction site. These tasks, trivial for any adult human, require continuous real-time prediction about how objects will behave, what forces are at play and what will happen if a specific action is taken. A system that can only predict the next word in a sentence has no machinery for any of this. LeCun calls this hopeless for robotics specifically, and he means the architectural limitation rather than a criticism of the engineers working on it.
He calls the underlying phenomenon the Moravec Paradox, named after the roboticist Hans Moravec who described it in 1988. The tasks that feel intellectually demanding to humans, solving equations, writing essays, translating languages, turn out to be relatively easy to replicate in software. The tasks that feel trivially easy to humans, catching a ball, walking across uneven ground, identifying whether a surface is safe to put weight on, remain stubbornly difficult. This is because human cognition is built on millions of years of physical experience. Language came much later. Trying to build general intelligence from language alone is, in LeCun’s view, building from the wrong foundation.
Whether he is right about where the current approach fails is one of the most consequential open questions in modern computing.
About the Author
Marcel Chidozie is a tech analyst and writer covering foreign news, fintech, and emerging technologies at TechRegard. Based in Nigeria, He's passionate about translating complex tech developments into compelling, accessible stories for diverse audiences. His work focuses on how technology shapes innovation across Africa and globally.
