A Child Who Drops a Spoon Understands More Physics Than a Chatbot
The column says LeCun believes chatbots manipulate language without ever experiencing the physical events they describe.
davidwsilva.substack.com · I'm Sorry to Burst Your Bubble: You Are Being Fooled About AI, and You Will Soon Feel Really StupidAt a moment when large technology companies are spending billions to build bigger chatbots, a columnist says Yann LeCun has been repeating an unwelcome warning about that bet.
LeCun, who founded and led Meta's AI research lab for more than a decade, is presented in the column as a sober technical expert. According to the column's account, LeCun has repeatedly said: "Large Language Models are a dead end on the path to human-level intelligence."
The column explains what it takes that claim to mean. In this view, chatbots are skilled at handling sentences but do not understand what those sentences describe. They can manage language without having any direct grasp of objects, actions, or why one thing causes another.
To make that concrete, the column uses the image of a child in a high chair. The child tries the same small experiment over and over and gets to see and hear the result, while a chatbot meets falling objects only through written descriptions.
According to the column's account of LeCun, simply making systems larger and feeding them more text and calculation will not close that gap. The missing part is not more reading, the argument goes, but a different kind of learning based on watching, moving, and touching things in the world, including video and spatial experience.
The column treats that position as a break from what most large companies are doing, and the columnist argues that this is why LeCun's view gets little attention. As evidence of the industry's direction, the column says Meta put $14.3 billion into Scale AI, brought in that company's founder Alexandr Wang to lead a new Superintelligence division, and placed FAIR, the research lab LeCun had led, under Wang.
The point is the difference between describing the world in sentences and learning about it through sight and touch.
Questions people ask
What does it mean to call a chatbot a text prediction engine?
A language model predicts the next word in a sentence based on patterns in enormous collections of text. Given a beginning, it calculates which words are most likely to come next, then generates an answer that sounds fluent, even when it has no direct experience of what it is describing.
If text is not enough, what kind of learning does LeCun think AI needs?
LeCun argues for systems that learn from video, spatial information, and interaction with the physical world, not just text. The goal is a model that builds an internal representation of how objects move, fall, and cause effects, closer to how a child learns by watching and touching.
Do other AI researchers disagree that scaling has limits?
Scaling has produced much more fluent and useful systems, and many researchers believe larger models keep improving. LeCun's point is narrower: scaling improves language manipulation, but it does not supply the causal and physical understanding that comes from sensory experience.