Ask the person next to you what intelligence is. Then ask the next one. Two different answers, both delivered with total confidence. Ask five engineers and you get five definitions — same as with story points.
That’s the entire AI debate: an industry arguing about a thing nobody in the argument can define. So let’s define it.
Intelligence is the ability to handle a situation nobody prepared you for.
Not knowledge. Not memory. Four thousand stored facts are a database. Intelligence starts when the situation isn’t in the database and you have to do something anyway.
For the purpose of this argument, there are two kinds
Natural intelligence. Every living thing has some. A crow. An elephant, big brain and all. Humans are in there somewhere. The colleague who fixed prod at 2 a.m. with no documentation and no clear idea what he was doing. We point at neurons the way you point at a server rack and say the website is in there somewhere.
Artificial intelligence. Judgment moved out of a mind and into an artifact. Artificial and artifact share the same ancient roots: art and making. A made intelligence. And a made thing has limits. The early version applied rules. The modern one infers them from examples. Both work inside a frame somebody drew. It may stretch or redraw parts of that frame, but it never operates without one. That’s not an insult. That’s the spec.
Artificial intelligence isn’t seventy years old. The term is. The thing is about four thousand.
Ur, around 2100 BC
The oldest surviving law code we have is the Code of Ur-Nammu. Clay, Sumerian, roughly three centuries before Hammurabi got famous for the same idea.
It’s written like this: if this, then that. If a man commits robbery, he is put to death. Scholars call it casuistic form. You’d call it a switch statement.
Judgment was pulled out of a human head, frozen into procedure, handed to somebody else to execute. The clerk in Ur didn’t need to be wise. He needed to be literate.
That may be the oldest surviving artificial intelligence we have. It ran on clay, and its processor was a guy.
The clause they wrote for us
A few centuries later, Hammurabi scaled the architecture and shipped the first engineering SLA.
Law 229: a builder builds a house, the house collapses, the owner dies — the builder is put to death. Law 232: destroyed property gets replaced, and the builder rebuilds the house at his own expense.
Someone looked at the person who ships things for a living and wrote down: you own the outcome. Incident, root cause, remediation at your own cost. Ours is called a postmortem, and it’s blameless. Theirs was extremely clear on who built the thing.
Everything after that is the same machine
Rome. The census. The tax office. The ticket that can’t move to Done until three fields nobody reads are filled in.
Four thousand years of building bigger artificial intelligence, running on the only hardware available: people.
Anyone who has stood in a municipal queue has watched it run. Paper through the window, a rule the person didn’t write and can’t change, a stamp. That’s inference on wetware, and it has a waiting number.
The clerk may be thinking. The process treats that as an implementation detail. We gave it a name — the process — and everyone agreed to treat it as a job.
You’ve worked inside artificial intelligence your entire career. It just wasn’t plugged in.
Then we plugged it in
Strip any program down to assembly and one move sits under everything: the conditional branch. The oldest instruction in computing is the oldest sentence in law, running at gigahertz.
The first era of digital AI was that: predict the situations, write down the verdicts. Expert systems were ten thousand if-statements with a marketing budget. Then prediction ran out of road, so we built the kind that infers the rules from examples. The large language model is the current peak of it — a better prediction mechanism, nothing more. Underneath, it’s still the conditional branch. Assembly. Binary. Ones and zeros. The third substrate. Clay, clerks, silicon.
Follow that through: every program you’ve ever shipped is artificial intelligence — predicted situations, frozen. A user hits an edge case on a Tuesday, and your judgment from eight months ago answers without you. You didn’t join the AI industry in 2023. You were never in any other industry.
Digital AI isn’t coming for you. It’s coming for the clerk.
Digital AI doesn’t replace natural intelligence. It replaces the analog kind — the layer that was already an algorithm wearing a lanyard. Same species, better hardware, no lunch break.
Plenty of engineering jobs quietly became clerk jobs. Ticket in, known fix, next column, repeat. It’s been an algorithm for years; we kept it in a person because a person was cheaper than the integration project.
The honest test: in the last month, how many times did you face something nobody had prepared you for, and do something about it anyway? If the number is zero, you weren’t replaced by AI. You were already running it.
The frame
Art, music, writing. The machine gets further than people are comfortable admitting — always inside a frame drawn from everything already made. Its ceiling is set a priori — in advance, in the Latin of Roman law.
Natural intelligence occasionally draws a frame that wasn’t there.
My seven-year-old came in saying she had seen an orange sun. I asked her what colour the sun is. Yellow, she said — the answer every child is issued. Try again.
She made a telescope out of her hands, looked up, and said: white.
She was right. And she had been right the first time too: the light is white, and the atmosphere turns it yellow, orange or red on the way down.
But the colour wasn’t the interesting part. The move was. When the stored answer failed, she didn’t reach for another stored answer. She changed the method. She looked.
We still don’t have a complete account of how that happens. Hard to replace a thing when nobody can write the requirements for it.
The confession
Yes, I use AI to write these. Obviously. It’s a tool. The alternative was cuneiform. It reviews, it trims, it argues, it catches the paragraph where I’ve gone boring.
But you don’t seriously think it finds the jokes funny. Or that it ever opened a conversation with: artificial intelligence was invented in Sumer.
I walked in with that one. It can produce a joke. It can improve an idea. What I cannot find anywhere in it is the moment of surprise — the private collision between two things that were not supposed to meet. A clay tablet hit a language model, and it happened in a head, not in a data center.
And it will get there, by the way. The timing. The jokes. All of it. It has the frames. It has the data. It will learn exactly where the laugh belongs. But my mind will still go somewhere that wasn’t in the frames or the data. It won’t know how I got there. Neither will I.
We invented artificial intelligence in Sumer, ran it on people for forty centuries, and are finally migrating it to hardware built for the job. Everything the machine takes from you was a rule someone else wrote down. You were the runtime.
Let it have that. It was never yours.
Keep the part nobody knows how to build.
By order of thebeanengineer.com — automate the clerk before you become one. Be an engineer. 🚀