John McCarthy's Mistake
The man who coined 'artificial intelligence' invented the most elegant language ever designed for it. It was the wrong tool — because it was the wrong category.
By Geordie Everitt
Interleaf, the desktop publishing system I spent my GTEDS years keeping alive, had a secret weapon: an embedded LISP engine. You could open a document and program it from the inside — extend the editor, automate the layout, make the software do things its designers hadn't imagined. For a certain kind of person this was intoxicating. I was that kind of person.
LISP carried a pedigree. It was invented by John McCarthy, the man who coined the term artificial intelligence and convened the 1956 Dartmouth workshop that founded the field. For decades, LISP was not just a language for AI — it was the language, the official chisel of the project, the thing entire machines were built to run. If you believed intelligence could be constructed, LISP was what you'd construct it with.
It is among the most elegant artifacts in the history of computing. And as a tool for creating intelligence, it failed completely.
The Premise Underneath the Language
The failure wasn't in the engineering. It was in the premise the engineering served — a premise so natural that for fifty years almost nobody thought to question it: intelligence is something you program.
It made sense. Everything else a computer did, you programmed. You analyzed the task, decomposed it into rules, and expressed the rules in code. Chess fell to this approach. Theorem proving fell. Symbolic algebra fell. The founders of AI reasonably extrapolated: reasoning itself would fall, because reasoning was — surely — rules. Find the rules of thought, write them down, and the machine thinks.
So the field spent decades on exactly that. Expert systems: interview the experts, extract their rules, encode them in LISP. Knowledge representation: formalize what the world contains so a program can reason about it. Each effort worked a little, in a narrow domain, and then hit the same wall — the rules never ran out. Every rule needed exceptions, every exception needed context, every context needed more rules. Common sense, it turned out, was not a compact rulebook waiting to be transcribed. It was bottomless.
The Answer Came From the Other Direction
The AI that finally worked — the one currently rewriting the economics of every industry, including the documentation mines I wrote about last time — contains, in the sense McCarthy meant, almost no programming at all.
Nobody wrote rules for grammar into a large language model. Nobody encoded the knowledge that water is wet or that grief follows loss or that a sentence about a courtroom should not suddenly be about a wedding unless something interesting is happening. There is programming in the system, certainly — the training infrastructure, the architecture, the orchestration is conventional code, much of it mundane. But the intelligence isn't in that code. You could read every line of it and learn nothing about why the model can write a sonnet.
The capability was grown. Show the system an enormous amount of human output, let an optimization process adjust billions of parameters, and intelligence — or something doing a disconcertingly good impression of it — condenses out. Not constructed, rule by rule, the way McCarthy's generation assumed. Cultivated, the way you'd raise a crop: prepare the conditions, supply the inputs, and wait.
This is a category error of the deepest kind, and it took half a century to surface: the founders mistook a growth process for an engineering process. They built ever-better chisels for a material that cannot be carved.
What the Mistake Teaches
I don't tell this story to diminish McCarthy, who was a giant, and whose mistake was the productive kind — the field that grew from his wrong premise eventually generated the evidence that overturned it. That is how it's supposed to work.
I tell it because the distinction it surfaces — some things are built; some things are grown — turns out to be the load-bearing idea of this whole series. We now have machines whose abilities were grown rather than written, and the lesson generalizes in an unexpected direction.
Because if the engineers were wrong that intelligence could be programmed into machines, there's a mirror-image error on the human side: the belief that knowledge can be installed in people. That learning is a transfer problem — find the right format, the right compression, the right delivery mechanism, and the skill goes in. Every generation of educational technology has been built on some version of that premise.
I've been spending my days lately acquiring a human language, with the most powerful AI in history at my elbow, and I can report that the premise is just as wrong the second time.
But first, a detour through what all that abundant bandwidth from Part 1 actually got spent on — and what happened when the only scarce thing left was attention.