Why LeCun says today's LLMs aren't really intelligent

Lately I’ve been trying to understand Yann LeCun.

He’s one of the three “godfathers” of deep learning, and shared the Turing Award with Geoffrey Hinton and Yoshua Bengio.

Interestingly, he isn’t “anti-AI,” nor “anti-deep learning.” Quite the opposite: he’s one of deep learning’s most important champions.

But he has stayed skeptical of the current LLM craze.

His core point isn’t:

AI is useless.

It’s:

Today’s large language models are very useful, but they shouldn’t be mistaken for the whole answer on the road to real intelligence.

That distinction matters.

Because the ChatGPT, Claude, Gemini and DeepSeek we know today are all built on the large-language-model approach. They learn the knowledge humans have expressed from huge amounts of text, code, images, video and other data, then get better and better at answering questions, writing code, doing research and completing tasks through post-training, reinforcement learning, tool use, multimodality, long context and agent workflows.

This approach has been enormously successful.

You could even say it’s the most practically productive AI approach we have right now.

But LeCun’s doubt is this:

A system that can say a lot doesn’t necessarily understand the world.

It knows “a cup at the edge of a table might fall” possibly only because it has seen similar phrases in lots of text.

So what should real understanding look like?

It should be able to see the cup’s position, material and weight, the edge of the table, the movements of the people nearby. It should be able to predict: if someone bumps the table, the cup might fall. It should be able to judge: if the cup is glass, it might break when it falls. It should be able to plan: if the goal is to avoid damage, move the cup to the middle of the table. And it should be able to act, then correct its next step based on the result.

That’s what LeCun means by a world model.

Roughly speaking, a world model means:

The AI has an internal model of “how the world changes.”

Not just predicting the next word, but predicting how the world will change next.

This is also the core of his disagreement with the mainstream LLM approach.

The mainstream LLM camp roughly believes:

Language/multimodal data + scaling + post-training + tool use + agent workflows will bring AI closer and closer to general intelligence.

LeCun puts more weight on:

Real intelligence needs perception, memory, goals, prediction, planning and feedback from action. Language is just one surface output of intelligence, not its root.

This isn’t a purely technical detail; it’s a big difference in how you see things.

If you understand intelligence as “can it generate high-quality answers,” then LLMs are already very smart.

But if you understand intelligence as “can it understand the state of a complex environment, predict consequences, make a plan, take action, and correct itself based on feedback,” then today’s LLMs are still far from enough.

So the real question here is:

When we say AI is smart, what exactly do we mean?

Being able to talk is one ability. Being able to summarize is one ability. Being able to write code is one ability. Being able to call tools is also one ability.

But these abilities don’t add up to complete intelligence.

Being able to talk isn’t understanding. Being able to answer isn’t being able to act. Being able to generate isn’t being able to judge. Sounding like human expression isn’t having human experience.

This is also where I resonate a lot.

I often feel that language itself is limited.

Many things can’t be fully expressed in language.

Why a space feels comfortable. Why a piece of music moves you. Why a certain moment in a relationship disappoints you. Why a design “feels right.” The subtle differences of touch on a piano key. How your body finds its balance when you move.

These can of course be described, but the description isn’t the experience itself.

Language is a highly compressed representation of experience, but it isn’t experience itself.

And LLMs learn precisely from these highly compressed human expressions.

That makes them powerful, because a huge amount of human civilization’s knowledge has been preserved in language.

But it also gives them limits, because language loses a lot of what lives in the body, space, action, situation and feedback.

So LeCun’s warning isn’t asking us to dismiss LLMs.

Quite the opposite: we should acknowledge their real-world value.

They’re already very well suited to writing, programming, summarizing, learning, research, product prototypes and content creation.

But we also need to know where their limits are.

LLMs are strong at language, as an interface to knowledge, at generating patterns and at coordinating tools.

But if AI is to enter the real world more reliably — robots, spatial computing, self-driving, industrial systems, long-term personal assistants — it may need another kind of ability:

Understanding how the world changes. Understanding what consequences an action will have. Understanding goals, environments, constraints and feedback.

The future most likely isn’t a choice between the LLM approach and LeCun’s approach.

It’s more likely a fusion.

LLMs handle language, knowledge and communication. World models handle understanding the environment and predicting change. Agent systems handle tools, permissions, memory, action and feedback.

In other words, a truly powerful AI may not be a model that chats better, but a system that can understand goals, understand its environment, predict consequences, carry out tasks and correct its mistakes.

That’s very instructive for building products, too.

Don’t only ask:

Does this AI answer well?

Also ask:

Does it understand the situation the user is in? What real context has it touched? Can its output be verified? Can it act? Are its actions bounded? Can it correct itself after failing?

The problem with many AI products today isn’t that they “can’t talk well enough,” but that they talk too well.

They seem to understand everything, but haven’t necessarily touched reality. They answer fluently, but can’t necessarily deliver results. They generate beautifully, but aren’t necessarily reliable.

So understanding LeCun’s view isn’t about picking a side.

It’s about reminding myself:

Don’t mistake “able to express” for “able to understand.” Don’t mistake “able to answer” for “able to act.” Don’t mistake “fluent language” for “reliability in the real world.”

Language is powerful. But language isn’t the world.

LLMs are powerful. But LLMs aren’t all of intelligence.