What LLM’s are really good at is manipulating language and producing language. And as it turns out, producing and manipulating language is simple. Because…. We think of human language as the epitome of human intelligence, and it turns out to be simple… It’s simple because it’s discrete. What we’re not able to do with AI systems today is dealing with the real world, understanding the real world, not even as well as a cat, or even a rat. So, in some directions, our AI systems are as smart as a lawyer passing the bar exam. It can answer all the questions, or solve some math problems, if they fall into some standard template. But on the other hand, they can’t do what a cat can do, or do what a ten year old can learn in one shot, they can’t learn in 20 hours how to drive a car, which any 17 year old can do.
--Yann LeCun, the godfather of AI, from The Economist podcast "Babbage", Feb 5, 2025.
Artificial Intelligence systems have proven, apparently, that rats do far more difficult things than lawyers. Who woulda thunk?
Aside from this tired lawyer joke proven true, however, there's actually a curious insight here -- that the mastery of words may not be humanity's crowning achievement, but instead a relatively simple game. Theories on why we developed language are diverse, for no other animal apparently needs it, and its purpose is slippery. One idea is that we use it for conscious thinking, evident in that little voice in our head. But so much of our cognition is wordless and emotional-- we feel our way through decisions, choosing friends and lovers, and so much more. Another idea is that we use language for complex communication, evident in the obvious -- the existence of teachers and books. And while linguistic communication is particularly useful, we can live without it. People who take vows of silence seem to manage all right.
So human language might just be more of a pastime than the defining and essential tool of being human... for apparently, there are far more complex things we do, such as the navigation of material reality, the understanding of it, and its manipulation into other things. All of which we can do without words.
Then again, we can interpret LeCun's observations as utterly obvious -- that if you train a computer on language, language will be easy for it, and other things, which it hasn't trained on, will be hard. We could surmise that we just haven't tried to teach machines to understand the real world yet... but this isn't true.
Since the 1960's, we've tried to teach cars to drive themselves. Shouldn't it be simple to tell a computer "Don't hit the child running after a ball! Ignore the tree's shadow!" And we keep getting almost there, but not quite. Apparently, telling the difference between a tree's reflection and a child running after a ball is beyond even the best computer systems. We have no true self-driving cars in 2025, just ones that do OK, mainly when in a closed circuit.
While it might seem obvious that the manipulation of reality is far more complex than the manipulation of language, nearly all human cultures give higher status to people who can manipulate language well over those who can manipulate reality well. It's the white collar versus blue collar divide. The ability to persuade another person through argument (the lawyer, the politician, the actor, even the salesman) has higher social and economic value than the ability to persuade a raw material into a new form (the craftsman, the workman, the tradesman).
I teach white collar professionals how to use their hands every day in my woodworking classes. They are nearly all surprised to find that, in fact, the work is challenging to learn. You can't cut fine dovetails on day one, or even Day 500. Happily, the learning curve is enjoyable, so they stick with it, slowly developing new skills, from a knowledge of the non-negotiable properties of materials and tools, to the capacity to see details of objects both natural and man made to understand their construction, to a capacity for three-dimensional or spatial thinking.
How to each a computer to visualize a cabinet's constituent parts such that you imagine different joinery, at each intersection, and make decisions about the capacity of each to resist the stresses the finished piece will encounter? A whole lot harder than teaching a computer what a writ of habeas corpus is, and when they are used.
We've long known this.
Reading The Clouds, I felt Aristophanes arguing the same point. The feckless, lazy kid takes a few short lessons from philosophers in how to argue, and thereby steals his father's fortune in court. So easy. To learn that kid how to make a Philadelphia Highboy? That would take much, much longer.
The lesson I take from Yann LeCun?
People who can use words powerfully are not doing anything particularly difficult. They really are dumber than rats. Let's stop respecting those who merely persuade us, especially as they often have no problem misleading us. Let's be smarter, and think more about what we do, and what others do, than what people say.
Comments