From HAL to Jarvis, part 3 of 7. A history of artificial intelligence, told as a story, for people who will use it. It began as a talk.

You taught me language, and my profit on't is, I know how to curse.

Caliban, in The Tempest, I.ii

Caliban, the island’s native creature, to Prospero, who taught him to speak. Four hundred years early, it is the most precise description of a language model ever written. Teach a creature language, and it learns everything that is in the language, including the parts you did not mean to teach.

The history of the field falls into three eras, and each era believed something different about where the hard part was. In the first, humans write the rules. In the second, the machine finds the rules. In the third, the machine generates, and then it acts. The eras do not replace each other. They stack.

A summer project, 1956

In 1955, four men write a proposal. John McCarthy, Marvin Minsky, Nathaniel Rochester, and Claude Shannon. They propose “a 2 month, 10 man study of artificial intelligence” at Dartmouth College, and they write that “a significant advance can be made in one or more of these problems if a carefully selected group of scientists work on it together for a summer.”

They named the field. They budgeted one summer.

Seventy years later, the field is still working on the summer’s list: language, abstraction, creativity, self-improvement. I say this with affection. These were the rule-writers. They were earnest, they were brilliant, and they gave us everything that came after.

Claude Shannon, the father of information theory, is on that proposal. You may recognise the first name. It is widely assumed to be where a certain assistant got its own, although its makers have never confirmed it.

Every rule typed by a human

Start in 1966, with the first chatbot. ELIZA. Joseph Weizenbaum wrote a short program that answered by turning your sentence back on you. “My boyfriend made me come here.” “Your boyfriend made you come here.” People told it their secrets. Weizenbaum’s own secretary, who had watched him write it and knew exactly what it was, asked him to leave the room so she could talk to it in private. The first chatbot, and the first people to trust one more than it deserved.

ELIZA was a trick, and Weizenbaum said so. What the rule-writers built in earnest, for thirty years, was logic. And when logic ran out, expert systems: the knowledge of a specialist, interviewed out of her and typed into the machine as rules.

MYCIN, at Stanford in the 1970s, held about six hundred rules interviewed out of physicians for diagnosing blood infections. If the infection is primary bacteremia, and the site is sterile, and the entry point is the gut, then the organism is likely bacteroides, with confidence 0.7. It matched the specialists.

XCON, at Digital Equipment from 1980, held thousands of rules for configuring computer orders, growing every year, each one typed by hand and each one argued over.

It worked. All of it worked. And then it drowned, and the reason it drowned is arithmetic.

The bottleneck

Take one clause of one contract. “The vehicle is covered if it is a passenger car or a light van, registered in Switzerland or in a neighbouring country, driven by the policyholder or by a named driver, on a public road or on private property with access.” Every or in that sentence doubles the number of cases. Four ors, sixteen cases, for one clause. A thirty-five-page motor contract, written out as rules, is not thousands of rules. It is billions.

That is what the rule-writers ran into, and it has a name. Edward Feigenbaum, who built the first expert systems, called it the knowledge acquisition bottleneck. The expert knows more than she can say. What she can say takes longer to type than the world takes to change. And what she types multiplies faster than anyone can read it. There are simply too many rules for a human to write down.

Era one died of specification. Every rule had to come out of a person, and the rules bred faster than the people.

Era one is not dead

I have to tell you something here, because it is true.

In 2018 I built one of these, at a Fortune 500 insurer. Not a chatbot: a logic representation of legal text. Policy wording, modelled so that a machine can evaluate it, and it still answers questions in call centres today.

Why build a rule system in 2018, thirty years after the winter? Because a claim has to be right a hundred percent of the time, and nothing stochastic came anywhere near that. It was engineering to solve a problem, not research.

And we did not drown in the ors, because we did not write rules. We drew the text. A motor contract that would be billions of rules written out is about two hundred nodes in our graph. The machine answers one question only: is what you are telling me inside the text? Yes. No. Or: I need to ask you something. “What happened?” “My car broke down.” Inside the text. “My car flew away.” Not inside the text. And when the machine does not know, it does not decide. It asks.

HAL had only yes and no. That third answer is the fence.

I will not say more about it here; it is a story of its own. I mention it because the modern shape, the one I think we will all end up with, is exactly that: a small layer of rules that never bend, an era-three machine doing everything else, and a human at the seam.

The winters

In 1973 a report by Sir James Lighthill for Britain’s Science Research Council says the grand promises have failed, and most of the funding stops. The field will later call it a winter.

Between 1987 and 1993 the expert-systems market collapses under its own maintenance cost. The companies that built dedicated AI hardware die. Researchers learn to call their work anything but artificial intelligence if they want to be funded.

In the autumn of 2026 the White House ordered the same two words out of its official language. Not out of shame, this time. Artificial, the President said, sounded fake.

Twice, the field was declared dead. Both times, a few people kept working. To understand what they were working on, I have to go back to 1958, because the idea that finally won was declared dead first.

Next: Part 4: A Machine with Millions of Dials

The series