Seventy years in one page

The name is older than the transistor radio

In 1950 Alan Turing published a paper asking whether a machine could think, and proposed a test: could a person, typing questions, tell the machine from a human? He called it the imitation game. He did not use the words “artificial intelligence”. Nobody had yet.

Those words were written in 1955, in a funding proposal for a summer workshop at Dartmouth College in the United States. The proposal, by John McCarthy and three colleagues, said that ten people working for two months could make “a significant advance” on getting machines to use language and improve themselves. The workshop ran in 1956. The advance did not arrive in two months. The name stuck.

That proposal is worth remembering for one reason: the gap between what was promised and what was delivered is the oldest feature of the field, older than any of its methods.

Two families, taking turns

From the start there were two ways to build a thinking machine, and you met the distinction in the first module of this course. You can write the rules, or you can learn them from examples.

The rule-writers dominated early. The Logic Theorist in 1956 proved theorems from a book of mathematics. ELIZA, in 1966, held a conversation by matching patterns and reflecting them back, and its author was disturbed by how many people confided in it.

The learners started at the same time and were sidelined for decades. In 1958 Frank Rosenblatt built the perceptron, a machine that adjusted its own connections from examples. The New York Times reported that the US Navy expected it to walk, talk, see, write and be conscious of its existence. In 1969 Marvin Minsky and Seymour Papert published a book showing what a single layer of such units could not learn, and funding for the learning approach dried up for a generation.

Both sides were right about something. The rule-writers were right that you could build a working system in a week. The learners were right about where it would end.

The pattern

Put the whole seventy years on one line and it repeats:

• 1956 to 1973. Symbolic programs, big claims, government money. Then a British government report in 1973 concluded the promises had not been met, and funding on both sides of the Atlantic fell.
• 1980 to 1987. Expert systems: hand-written rules for narrow jobs like configuring computers or suggesting antibiotics. Real companies, real revenue. Then the cost of maintaining thousands of hand-written rules caught up, and the market collapsed.
• 2012 onward. Learned systems, first in image recognition, then speech, then language. This is the boom you are living in.

Each boom began with a demonstration, attracted money on the strength of a promise, and cooled when the promise met the long tail of cases the demonstration had not covered. The next two lessons look inside those winters, because the mechanism of a winter is more useful to know than the dates.

A game a machine won without learning

One event confuses the pattern and is worth placing correctly. In 1997 IBM’s Deep Blue beat Garry Kasparov, the world chess champion. It was reported as the arrival of machine intelligence.

Deep Blue did not learn. It examined about two hundred million board positions a second and scored them with rules that chess experts had written by hand. It belongs to the rule-writing family, and it was that family’s high-water mark. The learning family’s equivalent came nineteen years later. In 2016 AlphaGo, trained on human games and then on millions of games against itself, beat Lee Sedol, one of the greatest Go players of his generation, by four games to one, at a game where the positions are too many to search. A year later it beat the world number one, Ke Jie.

The lesson to carry from this: “a machine beat a human at X” tells you nothing about how, and how is the only thing that predicts what it will do next.

A timeline to keep

1950 Turing’s paper. 1956 the Dartmouth workshop. 1958 the perceptron. 1966 ELIZA. 1973 the first winter begins. 1980s expert systems. 1986 the method for training multi-layer networks is popularised. 1997 Deep Blue. 2012 a learned system wins the ImageNet competition by a wide margin. 2016 a learned system wins at Go. 2017 the transformer design. 2020 GPT-3. 2022 a chatbot on a public web page.

What the pattern does and does not tell you

It tells you that a demonstration is not a deployment, and that money arrives before proof. It does not tell you when, or whether, the current boom cools. Two things are new this time and pull in opposite directions: real revenue exists, and so does capital spending on a scale that requires returns nobody has yet demonstrated. Anyone who tells you confidently which way it goes is pattern-matching on a sample of two.