Some people are predicting the imminent arrival of “artificial general intelligence,” or AGI. Depending on whom you ask, AGI is either just around the corner, already hiding in the basement, or waiting for one more GPU cluster before it wakes up and starts filing its own tax returns.
I’m skeptical—not because recent AI systems are unimpressive, but because the phrase “artificial general intelligence” is doing a lot more work than it can bear. Yann LeCun, one of the leading computer scientists responsible for developing AI, recently coauthored a paper titled “AI Must Embrace Specialization via Superhuman Adaptable Intelligence.” One of its strongest points is that AGI is poorly defined. Sometimes it is said to mean intelligence comparable to human intelligence. But as LeCun points out, human intelligence is not really “general” in the first place.
That matters. If we define AGI as “human-level intelligence,” we first have to ask: which human, doing what?
Human beings are intelligent in many extraordinary ways. We use language, make tools, tell stories, plan for the future, design cities, compose symphonies, and argue on the internet with a confidence that generally exceeds our knowledge. But our intelligence is also limited and specialized to meet human-specific bodily needs. Other species have other bodies and other kinds of intelligence. Bats and whales use their brains to navigate through echolocation, a feat humans cannot come close to matching. My cats are stupid critters sometimes, but they are vastly better than I am at running, leaping, balancing, and calculating distances in midair. I may understand more about the Federal Reserve, but if the skill involves landing gracefully on top of a bookshelf, the cat wins.
LeCun puts the matter sharply: “We suspect that the cause for such a widespread conflation of human intelligence with generality stems from the urge for self-flattery, and the difficulty of truly conceiving of our own limitations.”
That strikes me as exactly right. We call human intelligence “general” because we are the ones doing the defining.
When people say AI is close to achieving “general” intelligence, they are usually thinking about the remarkable advances in large language models. Those advances are real. The ability of modern systems to read, summarize, translate, code, explain, imitate styles, and converse fluidly is astonishing. However, language processing is not the only kind of intelligence humans display. It is not even the only kind of intelligence AI systems have developed.
AlphaFold, for example, is not a large language model. It is a specialized AI system that achieved impressive results in predicting protein structures. Its power came not from being general, but from being exquisitely focused. It was built around a specific scientific problem, trained on domain-specific data, and evaluated by domain-specific standards.
LeCun argues that this is not an accident. Specialization is not a defect in intelligence and is often the source of its power. A system that concentrates its energy on one type of task can outperform a system that spreads itself thinly over every possible task. LeCun calls this the “No Free Lunch” rule: no single learning algorithm or optimization strategy works best for every problem. If a system tries to “fold proteins and fold laundry,” it will probably underperform an algorithm that specializes in just folding proteins.
This seems obvious once you stop treating language as the whole of intelligence.
Physical activities such as locomotion are examples of intelligent behavior that humans perform readily but large language models do not perform at all. Robots can now do some physical tasks, but they remain limited compared with the broad, flexible competence of humans in the material world. People can wash dishes, fold clothes, care for hospital patients, put away groceries, climb stairs, open doors, and improvise constantly in messy environments. Humans do this so readily that often we don’t even think of it as intelligent behavior, but these are not trivial tasks. They require perception, balance, timing, prediction, adaptation and control.
Discussions of AI focus excessively on language partly because language is the particular skill that lets us have the discussion. It is also because people often make the category error of assuming that language and intelligence are the same thing.
They are not.
To understand this, consider tennis, which happens to be one of my hobbies. Tennis is a game that demands both athleticism and rapid adaptation. A player must react to spin, speed, angles, wind, footing, fatigue, and the opponent’s intentions. The ball is never exactly where you would like it to be. The court is a laboratory in which small emergencies keep arriving at high speed.
Here is a video of a single point played between Roger Federer and Novak Djokovic at the 2009 U.S. Open:
The players exchange groundstrokes. Then Djokovic hits a drop shot, drawing Federer into the net. Federer gets to the ball on a dead run, but Djokovic anticipates his shot and sends up a lob over Federer’s head, forcing him to scramble back toward the baseline. Running away from the net with the ball dropping behind him, Federer somehow barely reaches it in time and hits a through-the-legs “tweener” for a winner.
It was a spectacular shot, the kind most club players never even attempt in a lifetime of tennis. But what made it remarkable was not just the shot itself. It was the chain of perception, prediction, movement, emergency adjustment, and improvisation that preceded it. Both players had to read the situation, anticipate possibilities, and adapt instantly. None of this could have been scripted in advance. The entire point unfolded in just 15 seconds.
That was intelligence, but I’m quite sure that Federer was not doing his thinking in words. He did not have a little narrator in his head saying, “Run to the net. Now turn around. Sprint backward. Prepare the tweener. Execute now.” There was no time for that. The intelligence was embodied. It lived in reflexes, trained perception, and what we sometimes call “muscle memory.”
Once the point was over, Federer could talk about what happened, turning the experience into language after the fact. However, the intelligence he used to win the point was not verbal intelligence.
The same is true also of a tennis serve, which seems at first glance like the simplest shot in the game. Unlike other strokes, the serve begins entirely under the player’s control. No opponent has hit the ball. There is no incoming spin or pace to manage. The server simply tosses the ball into the air and hits it into the service box.
Simple—until you try to do it well.
I have worked on my serve for years and am still trying to improve it. An effective serve requires a complicated chain of coordinated actions: holding the ball correctly, tossing it consistently, moving the feet, bending the legs, rotating the torso, lowering the right shoulder, dropping the racquet behind the back, and then uncoiling the body so kinetic energy flows from the legs through the body’s core, shoulder, arm and wrist to maximize racquet head speed during the tiny fraction of a second when the strings meet the ball.
I can describe all of that in language. In fact, I just did. A tennis pro can also use words to explain the mechanics of serving. A book, video, or AI assistant can provide detailed descriptions that help guide attention, clarify concepts, and suggest drills to practice and improve technique.
But no one serves a tennis ball by reciting instructions internally at the moment of contact. There is no time. The purpose of the words is to help train neural circuits that are not themselves linguistic. When learning to serve a tennis ball, language is an assistant to nonverbal intelligence.
The same is true of driving a car, dancing, walking, playing piano, chopping vegetables, or catching a falling glass before it hits the floor. Language can help us learn these skills, analyze them, teach them, and remember them. But the skills themselves are not made of language.
This is why I think language occupies a strange and special place in human intelligence. It is not the whole of intelligence, but it is one of the great amplifiers of intelligence.
Language allows us to coordinate, plan, teach, persuade, document, imagine, and build things no individual brain could manage alone. We cannot echolocate like bats, but language helped us develop radar. We cannot predict protein structures by staring at amino acid sequences, but language, mathematics, science, computation, and engineering helped us build AlphaFold. We cannot individually contain all the knowledge required to run a modern hospital, power grid, or air-traffic control system, but language lets us distribute cognition across people, institutions, books, software, instruments, and machines.
In that sense, language and its artifacts were already a kind of artificial intelligence before the first computers were ever built. They live mostly outside our brains. They extend us. They surround us with specialized abilities we do not individually possess.
But even language is not “general intelligence” in any complete sense. No language system will ever serve a tennis ball. It can describe the serve, critique the serve, diagram the serve, and perhaps help a player understand how to improve. But the serve itself belongs to a different kind of intelligence: embodied, sensory, motor, and deeply specialized.
That does not make language unimportant. Quite the opposite. Language is one of the reasons human beings can build specialized intelligences beyond ourselves. It is the bridge between the intelligence in our bodies and the intelligence in our tools.
The mistake is to confuse the bridge with the whole landscape.
Large language models are impressive because language is powerful. They can manipulate the medium through which humans explain, plan, teach, and reason together. That makes them useful, surprising, and sometimes uncanny. But it does not mean they are close to possessing every kind of intelligence humans have, let alone every kind of intelligence found in nature.
Maybe the future of AI will not be one grand, unified, general mind. Maybe it will be more like the world we already inhabit: a civilization of specialized intelligences, some biological, some mechanical, some linguistic, some mathematical, some embodied, some superhuman in narrow domains, but often helpless outside their domains of specialization.
The fantasy of “artificial general intelligence” imagines a single machine intelligence that can do everything. The reality may be stranger and more interesting: intelligence as an ecosystem.
A cat can leap. A whale can echolocate. AlphaFold can predict protein structures. Federer can hit a tweener on the dead run. A language model can explain why none of these talents should be confused with one another.
And somewhere on a tennis court, I am still trying to toss the ball in the same place twice.
Leave a Reply