Language, experience, biological specialization, and the challenges awaiting artificial intelligence
A recent BBC article about Yann LeCun, one of the pioneers of modern artificial intelligence, raises a provocative question: if today’s AI is “not smart,” what must come next?
LeCun argues that large language models, or LLMs, are extremely capable in areas such as writing, coding, and answering questions, but remain poorly equipped for robotics. His criticism can be summarized in one sentence:
“LLMs are largely hopeless for robotics.”
At first, this statement may sound too pessimistic. ChatGPT and other language models appear remarkably intelligent. They can explain difficult ideas, translate between languages, write computer programs, analyze arguments, and converse naturally with human beings.
Yet LeCun’s criticism becomes clearer when we ask a more fundamental question:
What do we actually mean by intelligence?
The Language House in the AI City
I like to imagine artificial intelligence as a city made up of specialized houses.
The LLM lives in the Language House. It understands human speech and writing. It translates human intentions into forms that computers can process, and it translates machine information back into natural language.
Consider a passenger sitting in a self-driving car who says:
“I would like to go to Publix.”
Before language models, the passenger might have needed to use a touchscreen, select a navigation menu, type an address, and press a confirmation button. The human had to learn how to communicate in the computer’s preferred format.
With an LLM, the passenger can simply speak.
The Language House interprets the request and converts it into a structured command:
Intent: Travel Destination: Publix Passenger preference: Nearest suitable location
That command is then passed to the car’s command-and-control system. The navigation system identifies the correct Publix and plans a route. The vision system watches traffic lights, pedestrians, lane markings, bicycles, and other vehicles. The driving controller operates the steering, brakes, and electric motors.
When the car arrives, the navigation system reports that the destination has been reached. The LLM converts that technical status into a human message:
“You have arrived at Publix, sir.”
The LLM has not driven the car. It has served as the communication bridge between the human and the machine.
This suggests an important principle:
Before LLMs, humans had to learn the language of computers. With LLMs, computers are learning the language of humans.
The Brain Is Not the Whole Body
Biology gives us a useful model for understanding why the Language House cannot operate the entire AI City.
The human body is not made from one universal type of cell that performs every task. During development, stem cells specialize into neurons, muscle cells, bone cells, blood cells, skin cells, and many other forms.
These cells cooperate, but they do not all perform the same function.
The brain acts partly as a command-and-control system, but even the brain itself contains specialized regions. Vision, hearing, memory, language, planning, balance, and movement are handled through different but interconnected systems.
When I think:
“I want to pick up my coffee mug,”
my conscious mind does not calculate the exact force of every muscle, the angle of my wrist, the pressure of each finger, or the corrections required to prevent the mug from slipping.
My vision identifies the mug. My brain forms an intention. Motor-planning systems prepare the movement. The cerebellum helps refine it. The spinal cord and nerves coordinate muscles. Touch and vision continuously report whether the movement is succeeding.
Intelligence emerges from the coordination of many specialized systems.
Future AGI may therefore resemble a biological organism more than one gigantic language model. It may contain an executive system coordinating specialized capabilities:
- language and communication,
- vision and hearing,
- memory,
- planning,
- navigation,
- physical world models,
- motor control,
- learning from experience,
- and safety monitoring.
The LLM will remain important, but it will be one specialized organ within a larger intelligent organism.
Intelligence Is Trained Through Experience
Even a task as simple as picking up a coffee mug is not truly simple.
A human being has practiced reaching, touching, grasping, lifting, and balancing objects countless times since infancy. A baby reaches for an object and misses. The baby knocks it over, grips it too weakly, grips it too tightly, or drops it.
Through repetition, the brain, nerves, muscles, eyes, and sense of touch learn to cooperate. What later appears effortless is the result of years of physical training.
The same is true at a higher level of skill.
A violinist may spend thousands of hours practicing finger placement, bow pressure, rhythm, timing, and listening. An LLM can explain how to play a violin, describe musical theory, and analyze a concerto. But knowing the words is not the same as possessing the motor skill.
Erling Haaland offers another example. When he scores a brilliant goal, viewers see only a few seconds of action. Behind those seconds lie years of childhood practice, millions of touches of the ball, missed shots, sprint training, balance exercises, physical conditioning, and repeated corrections.
Haaland does not consciously calculate every joint angle before striking the ball. His body has learned through experience.
This reveals a major difference between language models and embodied intelligence:
| Language Model | Embodied Intelligence |
|---|---|
| Reads and talks about the world | Acts and learns within the world |
| Learns mainly from recorded information | Learns from physical success, failure, and feedback |
| Can explain how to grasp a cup | Can actually grasp cups of different shapes, weights, and materials |
Language teaches us what others have learned. Experience teaches us what we ourselves can do.
Human intelligence depends on both.
Why Language Changed Human Evolution
Although language is not the whole of intelligence, it has been crucial to the extraordinary development of human civilization.
A young chimpanzee may learn to use a tool by watching its parent. A human child can also learn through observation, but the parent can add verbal instructions:
“Hold it this way. Be careful with the sharp edge. Do not touch that part because it is hot.”
Language allows human beings to transfer experiences without requiring every individual to repeat the same experiment.
Writing made this transfer even more powerful. Spoken words disappear, but written language preserves knowledge across distance and time. A person can learn from someone living on another continent or from someone who died thousands of years ago.
Each generation can begin with the accumulated knowledge of previous generations rather than starting again from zero.
This cumulative transfer of knowledge helped produce science, engineering, medicine, philosophy, law, and modern technology.
In this sense, language is a form of intellectual inheritance.
Harari and the Power of Shared Fiction
In Sapiens, Yuval Noah Harari adds another important dimension. Human language does not merely describe physical objects and practical experiences. It also enables us to create and share stories about things that have no independent physical existence.
Humans create money, religions, companies, nations, laws, institutions, brands, and political systems through collective belief.
A dollar bill is only paper, and digital money is merely information stored in computer systems. Yet money works because people collectively accept the story that it has value.
A company such as Apple, Tesla, or SpaceX is not identical to any particular building, employee, factory, or product. It is a legal and social entity created by shared agreements.
A country such as Norway exists physically as land, mountains, fjords, roads, and cities. But its borders, laws, citizenship, constitution, flag, and national identity depend on shared human concepts.
Harari calls these structures imagined or intersubjective realities. Calling them imagined does not mean that they are unimportant or simply false. Their consequences are very real because large numbers of people organize their behavior around them.
This capacity for shared imagination enables millions of strangers to cooperate.
A chimpanzee group can cooperate in limited numbers through personal relationships. Human beings can build nations, religions, corporations, universities, armies, global markets, and international scientific communities because they share stories, rules, symbols, and institutions.
Human intelligence therefore has at least three great dimensions:
- Physical intelligence: learning through interaction with the real world.
- Linguistic intelligence: transmitting knowledge through speech and writing.
- Social and cultural intelligence: creating shared realities that allow large-scale cooperation.
AI Has Cracked the Language Code
The achievement of modern generative AI should not be underestimated.
AI has, in an important sense, cracked the language code of humanity.
Large language models can process the written inheritance of human civilization and communicate through the same linguistic channel that parents, teachers, writers, scientists, governments, and institutions use to transmit knowledge.
This is why LLMs feel so powerful. Language is not merely one application among many. It is the principal medium through which humanity stores and transfers its accumulated intellectual and cultural experience.
By mastering language patterns, AI has gained access to our libraries, technical manuals, laws, histories, stories, scientific discussions, business institutions, religions, and imagined realities.
However, cracking the language code is not the same as solving intelligence itself.
An LLM may know thousands of descriptions of coffee mugs without ever lifting one. It may explain football tactics without possessing Haaland’s trained body. It may analyze a violin concerto without having fingers, ears, muscles, or years of physical practice.
From Generative AI to AGI
The race to make LLMs larger has produced remarkable progress. More data and computing power have improved language understanding, reasoning, coding, and communication.
But it is reasonable to question whether simply training ever-larger language models will be sufficient to reach artificial general intelligence.
The next stage may not be another Language House containing more books. It may be the construction of the rest of the AI City.
AGI may require the integration of:
- a language system that communicates with humans,
- a world model that understands objects, space, time, and physical causality,
- vision and hearing systems that perceive the environment,
- memory that connects past experience to present decisions,
- planning systems that can pursue long-term goals,
- robotic control systems that coordinate physical movement,
- the ability to learn from mistakes and unfamiliar situations,
- and safety systems that keep actions within human intentions and values.
In such an architecture, the LLM would not disappear. It might become the universal human interface, interpreter, teacher, cultural library, and diplomatic ambassador of the machine.
But it would not be the entire brain, body, and civilization.
The Challenges Ahead
Moving from generative AI to AGI will require solving several difficult problems.
1. Learning the physical world
Robots must understand weight, friction, balance, force, distance, uncertainty, and cause and effect. They must handle objects they have never encountered and operate safely in unpredictable environments.
2. Learning efficiently from experience
Humans and animals often learn from relatively few experiences. Robots may currently require enormous amounts of data and repetition. Building systems that learn more efficiently remains a major challenge.
3. Connecting high-level intentions to low-level action
“Clean the kitchen” is a simple linguistic command but an extremely complicated physical mission. The system must identify objects, choose priorities, avoid hazards, manipulate delicate items, detect failure, and know when the task is complete.
4. Coordinating specialized systems
Vision, language, memory, planning, navigation, and motor control must communicate reliably. Intelligence may depend as much on their coordination as on the power of any individual component.
5. Understanding human society
An intelligent machine operating among people must understand not only physical reality but also social reality: ownership, permission, money, laws, institutions, promises, customs, privacy, and responsibility.
6. Safety and alignment
A machine that can act physically must understand limits. It must distinguish between a suggestion, a joke, a dangerous command, and an authorized instruction. The more capable the machine becomes, the more important these safeguards will be.
Conclusion: Building an AI Organism
Yann LeCun is probably right that LLMs alone are not sufficient for robotics or AGI.
But this does not make language models unimportant. On the contrary, language may be one of the most valuable capabilities an intelligent machine can possess because language provides access to humanity’s accumulated knowledge, culture, institutions, and collective imagination.
The mistake is not investing in language models. The mistake would be believing that the Language House is the entire city.
Biology suggests another path. Evolution created organisms composed of specialized cells, organs, senses, nervous systems, and control mechanisms. Human intelligence emerged not from one uniform component but from cooperation among many specialized systems, all shaped by experience.
The path from generative AI to AGI may follow a comparable pattern.
The future may belong neither to the LLM alone nor to the robot alone, but to an integrated artificial organism that can:
- communicate through language,
- learn through experience,
- understand the physical world,
- navigate human society,
- coordinate specialized capabilities,
- and act safely toward meaningful goals.
AI has cracked the language code of humanity. The next great challenge is teaching it how to live, learn, cooperate, and act within the world that language describes.
Reference
Ben Morris, “AI is ‘not smart’ so what’s next in artificial intelligence?” , BBC News, July 2, 2026.
Yuval Noah Harari, Sapiens: A Brief History of Humankind.
No comments:
Post a Comment