Language is not the world

The last post was about Yann LeCun’s doubts about LLMs.

His core warning is:

An AI that can talk doesn’t necessarily understand the world.

That view really resonates with me.

Because language itself has limits.

Many things can’t be fully expressed in language.

You can describe why a space is comfortable, but the description isn’t how your body feels being in that space. You can explain why a piece of music is moving, but the explanation isn’t the tone, rhythm and pauses themselves. You can call a design “refined,” but the real judgment often comes from proportion, material, white space, rhythm and experience. You can write “I’m lonely,” but that sentence can never fully hold loneliness itself.

Language is powerful.

It’s one of human civilization’s most important compression tools.

But language isn’t experience itself.

That’s why a group of directions in AI research today are all trying to answer the same question:

If language isn’t enough, what else does AI need?

I’ve come across a few key terms lately that I think are worth understanding on one map.

The first is symbol grounding.

Simply put: how does a word really connect to the world?

Say an AI knows the word “cup.” Does it really understand cups?

If all it learned from text is:

Cups hold water. Cups break. Cups sit on tables. Cups can fall.

That’s certainly a kind of knowledge.

But for “cup” to truly connect to the real world, it also takes sight, shape, weight, material, size, how you grip it, whether it’s full, what happens when it falls.

If a system only explains words with words and symbols with symbols, it may form a beautiful network of language, but the meaning never really connects to the world.

That’s the question symbol grounding cares about:

How do linguistic symbols get their meaning from the world?

The second is embodied AI.

Embodied AI holds that intelligence doesn’t happen only in language, or only in an abstract brain.

Humans understand the world through the body, the senses, action, environment and feedback.

We understand “heavy” not just because we know the word, but because our bodies have lifted heavy things. We understand “far” not just because we know units of distance, but because our bodies have walked the road. We understand “nervous” not just because we know the definition, but because our bodies have actually felt the heartbeat, the breath, the pressure.

A lot of knowledge isn’t explicit language but bodily experience.

That’s also why playing piano, sports, design judgment and the subtle shifts in a close relationship are so hard to put fully into words.

An AI that truly enters the real world may not be able to just read text.

It needs to see, hear, move, try and fail, and take in feedback.

The third is world models.

You can think of a world model as:

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

Not just predicting the next word, but predicting what will happen next.

Take the cup at the edge of a table.

An ordinary recognition model might say: there’s a cup here. A language model might say: the cup is near the edge and might fall. A stronger world model should be able to judge: if someone bumps the table, the cup might fall; if it’s glass, it might break; if the goal is to avoid damage, move it somewhere safer first.

In other words, a world model cares not about “how to say it,” but about “how things will unfold.”

That’s very important for robots, self-driving, industrial systems and spatial computing.

The fourth is spatial intelligence.

It’s one of the directions Fei-Fei Li is especially focused on now.

Spatial intelligence is about whether AI can understand a three-dimensional world.

Where are things? How do they relate to each other? What changes when you look from different angles? How does a person move through space? How will an action affect the surroundings?

Many AIs today are very good with text and two-dimensional images, but the real world is three-dimensional, continuous and dynamic.

If AI is to enter AR/VR, robots, self-driving, home spaces, architectural design and game worlds, it can’t just “see a picture”; it has to understand spatial structure.

The fifth is VLA, the vision-language-action model.

The term breaks down like this:

vision: seeing. language: understanding spoken or written instructions. action: taking action.

That is, AI doesn’t just look at images and answer questions; it links vision, language and action together.

Say you tell a robot:

“Take the red cup on the table to the side of the sink.”

It needs to see the table, recognize the red cup, understand your words, plan a path, control its arm, grasp the cup, avoid obstacles and put it in the right place.

That isn’t a purely language task.

It’s a combined task of vision, space, action and feedback.

The sixth is human-compatible AI, or AI that works better alongside people.

This line of work is concerned with:

AI not just completing tasks, but understanding people’s goals, constraints, preferences and ways of working together.

Because real-life tasks are rarely completely clear.

When a user says “help me tidy this up,” how far should it go? When a user says “make it look more refined,” what does refined mean? When a user says “handle this for me,” what can be done automatically, and what must be checked with a person first?

A truly useful AI isn’t only smart; it’s also controllable, trustworthy, explainable and able to hand things over.

It needs to know when to act, when to confirm, when to stop, and when it must hand judgment back to a person.

These directions sound academic, but they’re actually very instructive for building products.

They all point to one thing:

AI’s next step may not be chatting better, but being more constrained by reality.

A good AI product shouldn’t just be a language machine.

It should understand the task, understand the situation, touch the real context, and produce results that can be verified.

Take a piano practice product.

If the user only types “I can’t play it well,” it’s hard for AI to really understand what’s going on.

More valuable context might be:

Audio. Video. Tempo. Position in the score. The teacher’s notes. Practice history. The passages that went wrong. Hand shape and touch.

Because playing piano isn’t a language problem.

It’s a problem of the body, the ear, rhythm, force and feedback.

Or take a tool for organizing your life’s information.

If AI just generates a piece of advice like “you should organize your accounts and documents,” the value is limited.

What really has value is something structured, deliverable and verifiable:

Where are the accounts? Where are the documents? Who can be contacted? What information is missing? What comes first? Can it finally export a document that someone else can actually understand and use?

That’s the move from “able to answer” to “able to deliver.”

So when I judge AI products now, I ask one question less:

Which model does it use?

And a few more:

What real data does it touch? Does it understand language, or the task and situation? Can its output be verified? Can it act? Where are the limits of its actions? Can it correct itself after failing? Where must human judgment be kept?

A lot of the time, the real difference between AI products isn’t the model name but grounding.

That is, whether it’s actually held in check by reality.

Language is powerful, but language isn’t the world. Answers are useful, but answers aren’t action. Generation is stunning, but generation isn’t reliable delivery.

The AI of the future may not just be a model that talks better.

It may be a system that can understand goals, enter a situation, touch reality, take action and correct itself through feedback.

That’s why I think these research directions are worth following.

They won’t necessarily become a new tool right away. They won’t necessarily be the most useful for everyday efficiency right now. But they help us understand:

where AI’s limits are. Where the opportunities in the next phase are. And why a truly good AI product can’t stop at “sounding like it understands.”