The Turing Award winner and former Meta AI chief argues that scaling LLMs won't produce human-level intelligence and predicts their obsolescence across most applications.
Yann LeCun, arguably the most credentialed skeptic in artificial intelligence, went on Bloomberg’s “The Close” on May 21 and said what he’s been saying for years, only louder: large language models are not the path to real intelligence. They’re a detour.
The Turing Award winner and former Chief AI Scientist at Meta sat down with Jean-Philippe Vert to lay out a case that LLMs, the technology underpinning ChatGPT and its many imitators, are fundamentally limited by design. His core argument is deceptively simple. Language is just one thin slice of how humans understand the world, and building intelligence on text tokens alone is like trying to learn to swim by reading about water.
The sensory gap #
LeCun’s critique centers on what you might call the sensory bottleneck. Humans process a continuous flood of visual, tactile, auditory, and spatial information every second. We build mental models of how objects behave, how gravity works, how a glass will shatter if you knock it off a table. LLMs process none of that. They process discrete chunks of language, sequences of tokens that represent words, and they do it extraordinarily well.
LeCun has made this point before, many times. What made the Bloomberg appearance notable is the specificity of his timeline. He predicted that LLMs will become “largely obsolete” across most applications within five years. That’s not a vague “someday” dismissal. That’s a clock ticking.
He also advised against using LLMs as the focal point for doctoral research, characterizing them as a diversion from the real work of achieving human-like intelligence.
From Meta to AMI Labs #
LeCun left Meta at the end of 2025 after years as one of the most prominent figures in corporate AI research. His departure wasn’t a retirement. He co-founded Advanced Machine Intelligence Labs, known as AMI Labs, a startup valued at $3.5 billion.
The company’s mission is built around the exact alternative architectures LeCun has been advocating: world models and something called Joint Embedding Predictive Architecture, or JEPA. Think of world models as AI systems that don’t just predict the next word in a sentence but instead build internal simulations of how the physical world works. JEPA takes that further by learning to predict abstract representations of data rather than raw pixel-by-pixel or token-by-token outputs.
Why the scaling argument doesn’t convince him #
LeCun’s argument is that no amount of scaling will overcome the architectural limitations of a system that only processes language. You can make an LLM bigger, but you can’t make it see. You can’t make it feel the weight of an object or predict what happens when you push a ball off a ledge, at least not in a way that reflects genuine understanding rather than statistical correlation from training data.
It’s worth noting that LeCun has been right before in ways that took decades to prove out. He pioneered convolutional neural networks in the late 1980s, an idea that was largely ignored until it became the backbone of modern computer vision around 2012.
The Bloomberg interview contained no references to crypto or blockchain, which is itself informative. LeCun’s focus remains firmly on the science of intelligence itself, not on its financialization.
Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our