Yann LeCun is not a critic from the outside. He is one of three people who invented the deep learning algorithms that power every AI system you use today. He won the Turing Award — the Nobel Prize of computer science — for this work. He is currently Chief AI Scientist at Meta, one of the largest AI research organizations on Earth.
In April 2026, speaking to a capacity crowd at Brown University, he said something that should terrify every person who has invested in, built on, or staked their career on large language models:
“There’s literally hundreds of billions invested in an industry that basically is counting on the fact that LLMs are going to reach human-level intelligence.”
He does not believe they will. He believes the entire industry is heading toward a wall.
What LeCun Actually Argues
LeCun’s position is not that AI is useless or overhyped in general. His position is specific and technical: large language models — the architecture underlying ChatGPT, Claude, Gemini, and Grok — are fundamentally incapable of achieving human-like intelligence because of what they are and how they work.
LLMs predict the next token in a sequence. They are extraordinarily good at this. They have been trained on essentially all human text ever written, and they have learned patterns so deep and complex that they appear to understand language, reasoning, and knowledge.
They appear to. LeCun argues they don’t — and cannot — because understanding requires something LLMs entirely lack: a model of the world. Humans understand language because we have bodies that experience physical reality. We know that a glass falls when you drop it not because we read about gravity, but because we have dropped things. We know that a person crying is in pain not because we read about sadness, but because we have been sad.
LLMs have read about all of this. They have not experienced any of it. And LeCun argues that no amount of reading — no matter how much data, no matter how many parameters — can substitute for the embodied, experiential understanding that produces genuine intelligence.
The $100 Billion Problem
OpenAI, Google, Microsoft, Anthropic, and Meta have collectively committed hundreds of billions of dollars to scaling LLMs. The entire investment thesis depends on a single assumption: that making models bigger, training them on more data, and running them on more compute will eventually produce human-level intelligence.
LeCun says this is wrong. Not directionally wrong in a “we’ll get there eventually” way. Fundamentally wrong in a “this architecture cannot get you there no matter how much you scale it” way.
If he’s right — and he is one of the most credentialed people alive to make this assessment — then the entire LLM investment wave is not building toward AGI. It is building toward increasingly impressive autocomplete. Valuable, commercially useful, genuinely impressive autocomplete. But not intelligence.
The Counterarguments Are Weaker Than They Sound
The standard response from the LLM camp is: the models keep getting better. GPT-2 couldn’t do what GPT-4 can. GPT-4 can’t do what GPT-5 can. Therefore, the trajectory points toward human-level intelligence.
LeCun’s response: trajectory within an architecture does not tell you whether the architecture has a ceiling. A car going faster and faster down a road does not mean it can fly. The road might end. The architecture might plateau. And the evidence from GPT-5 — which launched to disappointment, which couldn’t beat its predecessor on user preference tests, which faked mathematical achievements — suggests the plateau may already be visible.
What Comes After LLMs
LeCun is not arguing AI is finished. He is arguing that the path to human-level AI requires a fundamentally different approach — one that incorporates world models, embodied learning, and reasoning that goes beyond statistical pattern matching.
Meta is working on this. DeepMind is working on this. Various academic labs are working on this. The work is less flashy than releasing a new chatbot. It doesn’t generate revenue yet. It doesn’t produce demos that go viral. But if LeCun is right, it is the actual path to the thing everyone keeps claiming LLMs will eventually become.
The uncomfortable question for investors, for companies, and for the millions of people who have restructured their careers around AI: what happens to the $100 billion when the wall arrives?
LeCun has been saying this for years. Every year, the industry says he’s wrong. Every year, the models get more impressive. And every year, the gap between “impressive language model” and “human-level intelligence” stays exactly as wide as it always was.
[…] that every AI company on the planet claims to be outperforming within six months of each release. The world’s top AI scientist has already called the entire LLM approach a dead end, and he has not been proven wrong […]
[…] Yann LeCun, one of the foundational figures in modern AI research, has publicly stated that the enti… for achieving genuine machine intelligence. His argument is not that LLMs are useless but that scaling them endlessly will not produce the kind of reasoning and understanding that people assume comes with raw size. The trillion-parameter race continues anyway, partly because the companies running it have already committed enormous capital to it, and partly because “ten trillion parameters” is an easier investor story than “we made architectural improvements that are difficult to quantify.” […]