The Chisel-Construct Cycle · AI Series · 13

Who First Said “AI Winter”?

Han Qin · About 16 minutes · Published

1. The Last Session

August 6 to 10, 1984, at the University of Texas at Austin. The American Association for Artificial Intelligence was holding its annual meeting, formally the National Conference on Artificial Intelligence. The previous essay described the exhibition hall. Sales demonstrations played through loudspeakers, companies were recruiting, and a brochure said artificial intelligence had finally come of age. One company's tool displayed a web of lines and boxes on a terminal screen.

The last session at the same meeting was a panel. George Johnson, the reporter from the previous essay who was there, called it the closing session. The subject was “The Dark Ages of AI: Can We Avoid Them or Survive Them?”

Here is who was onstage, as the published account in the fall of 1985 records it. Drew McDermott chaired. The speakers included M. Mitchell Waldrop of Science, Ron Ohlander of the Defense Advanced Research Projects Agency, Roger Schank of Yale, B. Chandrasekaran of Ohio State, and John McDermott of Carnegie Mellon.

John McDermott wrote the configuration program in the previous essay. It was in use on DEC's production line that year. He sat on a panel about dark ages. Minsky isn't on the list.

At the same conference, the exhibition hall was selling and the last session was worrying. Both were happening in the same place in the same week.

2. Some People Called It That

The chair said straight away that people expected too much of artificial intelligence. The panel would talk about what three groups expected: businesses, the military, and the press. What those groups hoped for didn't line up with what was coming out of the labs.

Then he sketched the worst case, around five years in the future. In it, the Strategic Computing program of the Defense Advanced Research Projects Agency, DARPA, had failed as its driverless vehicles failed to roll; Japan's fifth-generation project hadn't yielded results; the new companies had folded; large companies such as Texas Instruments and Schlumberger had lost interest; funding for artificial intelligence faced a backlash. Then research projects would be renamed in a hurry.

At this point he said that some people were already calling this scenario an AI winter. Some people: the name was in use before he said it. Here is how winter first appears in this session: a name borrowed for a future that hasn't happened.

The reporter used the word too. He explained it this way: businesses and the military would grow disappointed, see artificial intelligence as a hole they could never fill, and stop giving it money.

The chair went straight on to say that he didn't think the scenario was likely, even in a milder form.

The worst case, the name winter, and the words I don't think are next to one another in the same passage. Over the next forty years the first two would be quoted countless times. Almost nobody quotes the third.

3. Fifty Years

Schank told a story of his own at the same session. About ten years before, he'd told a man named Eugene Charniak that the hard problems would be solved in ten years. Ten years had passed and they weren't solved. Some, he now thought, might not be solved in fifty.

He told of another conversation, this one with a real estate developer about expert systems. The developer took an expert system to be a way to put everything a person knew into a machine so the machine would be like a person. Schank told him that nobody yet knew how to do that.

Someone outside the field took an expert system for a person. Someone inside the field told him they couldn't do it yet. This was in 1984, in the summer when expert systems were hottest.

Putting everything a person knew into the machine was what people in the previous essay had spent twenty years trying to do. What that essay found was that somebody had to stay beside the machine to look after whatever had gone into it. As the outsider imagined it, once the work was done the machine would be the person. The person who had to stay beside it was missing from that picture.

Minsky, as we saw, wasn't on the list. In a common later account, he and Schank came up with the term winter at the 1984 panel. Minsky isn't listed in the published proceedings. What the chair actually said was that some people already called it that.

4. The Disk Outside the Room

Later histories put the start of the second winter around 1987. One item in the worst-case scenario was that the new companies would fold.

The best-known company was Symbolics. It sold computers made specifically to run a list-processing language. At the time people called them Lisp machines. Another company sold similar machines, under the name Lisp Machines, Inc.

The noisy disks and processors of these machines were deliberately placed outside the programmer's office. Only a display and keyboard stayed on the desk, connected by a network cable. Later the company provided very good displays and a mouse of its own design. Keeping the noise down was an explicit goal. The person at the desk couldn't hear the machine. The machine was outside. A high-end one cost around a hundred thousand dollars.

On the first page of its 1985 annual report, the company said its products were intended to make artificial intelligence and other symbolic processing techniques easier to use. Artificial intelligence appeared in the company's own annual report.

In January 1988 the company announced it was cutting ten percent of its staff, a little over seventy out of around 730 people. The quarter reported on February 2 brought sales of $23.8 million, ten percent down on the same quarter a year earlier. It lost $15.4 million that quarter and nearly $20 million over the first six months of the fiscal year. On March 22 the Los Angeles Times ran a headline calling Symbolics a fallen star of the computer world.

The account that came to be repeated is that in 1987 a five-hundred-million-dollar industry vanished overnight. The figures that survive from the time are those above. They are from one company.

People who were there, and people writing later, give several explanations for why it fell. One is that other companies' machines got cheaper. By the mid-1980s, general-purpose workstations made by Sun and DEC were cheap enough to run Lisp on.

Another is that the buyers had bought enough. Francis Feeney of the company told a business magazine in September 1988 that while it was growing fast nobody had seriously asked whether there would be a wider market once that small group of high-end Lisp programmers had what they needed.

Another is that an indirect source of money dried up. The Strategic Defense Initiative was not the Strategic Computing program mentioned earlier. Its money went to contractors and research groups, which bought these machines to build expert systems. When that money was cut back, fewer orders came in. The company hadn't received the money directly.

Another is that the company did everything itself. Chips, circuit boards, cabinets, peripherals, operating systems, development tools, mail, networking, applications, even manufacturing. Its machines were difficult to fit into a large company's existing setup. A project for a cheaper platform was dropped when money got tight. David Moon, an engineer, said later that the cost had been a target for the main model, too, but they hadn't got it down. The company put clean design ahead of cost, compatibility, and delivering on time.

There's another explanation to keep separate: expert systems were difficult to maintain. This is about software, not the price of the machines. By 1989 DEC employed about fifty-nine technical staff just to maintain the infrastructure and rule bases of its own expert systems. The maintenance team in the previous essay, growing from two people to more than sixty, was part of this.

The company called Lisp Machines went bankrupt in 1987. Another company that made Lisp software, Lucid, shifted to developing tools for another popular language. It too went bankrupt, in 1994.

5. Three Endings

The worst case had two other items: Japan's fifth-generation project yielded nothing, and Strategic Computing failed.

The Japanese project's official name was Fifth Generation Computer Systems. Its main term ran from fiscal 1982 to fiscal 1992, roughly spring 1982 to spring 1993. On June 4, 1992, a German expert invited to assess the project, Wolfgang Bibel, turned in his evaluation. In his view it was an outstanding success. His reasons were its work in logic and parallel computing, and its software for logic programming.

A scaled-back follow-on project came afterward. The number of researchers dropped from about ninety to about forty; the research groups went from seven to two. The research center closed at the end of March 1995. Since August 1992 it had been putting out the software it made for people to use.

Britain's Alvey Programme ran principally from 1983 to 1987. Its AI component was called Intelligent Knowledge-Based Systems from the start. A formal government-commissioned evaluation in 1991 assessed different things separately. Many of the scientific and technical goals had been met. Commercial exploitation was markedly weaker. Capacity and cooperation between universities and companies had been built up.

America's Strategic Computing program didn't have a closing date like Japan's. In late 1987 DARPA dropped the high-profile goal of an autonomous vehicle and shifted the focus of its computing research to a set of grand challenges in high-performance computing. On the ground at Carnegie Mellon, work on autonomous driving carried on.

Three government programs, three endings, three different years. Japan's ran its term, was judged outstanding by one reviewer, and eventually wound down and closed. Britain's finished its main term, with evaluators reporting substantial scientific achievement but weak commercial results. America's changed direction in 1987. The vehicle itself kept going.

Compare it with the worst case from 1984. The high-profile Strategic Computing goal was dropped within five years. The vehicle wasn't built as originally described, but the work on building it didn't stop. One reviewer called the fifth-generation project outstanding; others later called it a failure. Several of the new companies folded. Part of the worst case came, and part didn't.

6. Changing the Name

One item remained in the worst case: projects hastily changing their names. In the story told later, this is the item that gets told the most. The name artificial intelligence went bad, so people put it away, took out another name, and went on with the work.

Here is what one would look for if that were the story. At some point between 1988 and 1995, a named researcher, in a funding application, a rejection letter, or an internal memo about changing the title, sets down in black and white that the term has such a bad reputation that they're taking it out. People have been telling the story for years. None of those who tell it has produced that document here.

There is a sentence about people avoiding the name. It appeared in 2005 in a piece by New York Times reporter John Markoff. When artificial intelligence was at its most disreputable, he said, some computer scientists and software engineers stayed clear of the term for fear of looking like impractical dreamers. He doesn't name a person, a grant, or a paper title. The sentence has been widely quoted since.

There is a line from the machine company as well. Russell Noftsker, its former president, said in a 2001 interview that they had never been an artificial intelligence company. That was how he defined it more than a decade later. On the first page of its annual report in 1985, artificial intelligence was in print. Both statements exist. One was printed in 1985, and one was said in 2001.

Lucid shifted to making development tools for another language. The documents from the time do not say it did so to hide artificial intelligence from its name.

What about the names said to have replaced it? Machine learning was in the title of the 1959 paper by Samuel, the checkers programmer in essay seven. A journal called Machine Learning started in 1986, before the usual starting date for the second winter. Pattern recognition was in the title of Oliver Selfridge's “Pattern Recognition and Modern Computers” in 1955. Calvin Mooers was using information retrieval in the late 1940s and early 1950s for discussions of how machines might help people find information. Decision support was already used in a management paper by Gorry and Scott Morton in 1971. In 1978 Keen and Scott Morton published Decision Support Systems.

Knowledge engineering was what one company called its environment at the 1984 exhibition. Intelligent Knowledge-Based Systems was in the British Alvey Programme's vocabulary from 1983. Expert systems was a term even a real estate developer used in 1984. Cognitive Systems was the name of Schank's own company, which was already in business before 1984.

None of these names came out of winter. They were there before it, some as much as thirty years before.

7. The Name Was Still on the Door

What about the name itself? Had it been put away? The American Association for Artificial Intelligence still called its 1992 meeting the National Conference on Artificial Intelligence. Its journal, AI Magazine, went on coming out.

In August 1989 a meeting in Windsor, Ontario, published proceedings under the title Proceedings of the Fifth Conference on Uncertainty in Artificial Intelligence. Artificial intelligence was right there in the middle of the title. In 1988 Judea Pearl published Probabilistic Reasoning in Intelligent Systems.

In 1994 the association wrote a report for DARPA, Intelligent Systems for the 21st Century. In 1995 it put out The Role of Intelligent Systems in the National Information Infrastructure. The titles say intelligent systems and not artificial intelligence. They were issued in the name of the American Association for Artificial Intelligence.

There really were new names in those years. The first Neural Information Processing Systems conference took place in 1987, and its proceedings came out the next year. Neural Computation began in 1989. In 1989 Gregory Piatetsky-Shapiro organized an IJCAI workshop on knowledge discovery in databases. A book under that name came out in 1991, with the American Association for Artificial Intelligence as one of its publishers. In 1991 Zadeh promoted soft computing; in 1994 a major meeting had computational intelligence in its name. Pattie Maes wrote about agents for a computer publication in 1994. In 1995 Wooldridge and Jennings published “Intelligent Agents: Theory and Practice.”

The new names have something in common. They were names for new conferences, new journals, or new areas of particular interest, each narrower than artificial intelligence. None comes with a surviving document saying the name was chosen to avoid artificial intelligence.

So the records show three different things. First, the established names of neighboring disciplines went on being used. Second, new specialist conferences and journals got narrower names. Third, many years later someone said some people had avoided the term at the time. These three things were later made into one story: the name went into hiding.

The work didn't stop in those years. Starting in 1984, Douglas Lenat led a group at a research consortium funded by several companies, typing common sense into a machine one item at a time, all the way to 1994. The project was called Cyc. The expanse of everyday background that Dreyfus had been pointing to since 1965 became an engineering job. Instead of arguing about whether it could be put in, they put it in one item at a time. The method in the previous essay, putting knowledge into a machine, was carried to its limit.

In 1989 Yann LeCun and colleagues at Bell Labs published “Backpropagation Applied to Handwritten Zip Code Recognition” in the first volume of the new journal Neural Computation. The postal codes on the envelopes had been written one by one, by hand.

Pearl's book in 1988. Christopher Watkins's Cambridge dissertation Learning from Delayed Rewards in 1989. The Carnegie Mellon car still driving. From 1992 through 1995, Gerald Tesauro at IBM Research used a method that learned from differences over time to train a backgammon program and published one paper after another. In 1995 Corinna Cortes and Vladimir Vapnik at Bell Labs published “Support-Vector Networks” in volume twenty of Machine Learning. Beginning in 1989, DARPA held annual workshops on speech and natural language.

The money for this work came from different places: a consortium of companies, a telephone-company lab, universities, a computer company's research center, a special Defense Department program. Look only at sales of expert systems and the machine company's financial reports and you might think everything stopped. Look somewhere else and you see another stretch of ground.

8. If

In August 1984, at the last session, someone laid out the worst case. Six things, the last of which was research projects changing their names in a hurry. He said some people called it winter.

Then he said he didn't think the scenario was likely.

Over the next few years some of the six things happened, some only happened in part, and some leave no trace in the records of the time used here. Several companies folded. One machine company lost a great deal of money. The American program abandoned its most public goal, but the people working on the car went on working. An outside reviewer called the Japanese project outstanding. As for renaming, this essay hasn't found a person who changed the name of a particular project in those years to avoid the term. Later still, renaming got written into the history.

Winter is used here in two ways, and they should be kept apart. In the material essay ten examined, the events from 1966 to 1976 weren't called winter then. The word was laid back over them afterward. This time it points the other way. In 1984 it pointed forward, a name for a feared future that hadn't happened yet. A name given to a future carries a script: this comes first, this comes after, this is who changes the name at the end.

Part of that future came. The name was used for that part. The scenario in 1984 includes renaming, and the later retellings include renaming too. The first is a conjecture and the second a recollection. No record of a name being changed joins the two. We can't say whether the second came from the first.

Read only the later account in the past tense and two things are easy to miss: the if at the beginning and the I don't think at the end. Both are qualifications.

More than two phrases were cut. The session asked whether people could avoid it or survive it. The title has two verbs. Both are things people have to do.

A warning is addressed to people who still have time to do something. A history is addressed to people for whom it's too late to act. Cut the if, and a discussion of whether something could be avoided is remembered as a record of its not being avoided. The people onstage go from giving a warning to having got the prediction right.

From the request for money in essay six to this essay, something has been cut out each time. In essay six it was conjecture before the sentence. In essay seven it was the six games after the one game. In essay eight it was or refute in the parentheses. In essay nine it was a whole sentence: of course she knew it was a machine. In essay ten it was the boundaries between four events. In essay eleven it was still too early. In essay twelve it was the scale of one to ten behind 0.9. Here it is an if.

The things removed are large or small, a couple of words or six games. What remains afterward is shorter, tidier, and easier to pass on.

Back, finally, to the session. A man stands at the front of the room and sets out the worst case point by point. The program fails, the car isn't built, Japan has no results, the companies close, the money is withdrawn, and the research projects change their names in a hurry.

Some people call that winter, he says.

Then he says he doesn't think the scenario is likely, even in a milder form.

The people sitting there that day heard two sentences. The people who came later heard one.

The machines are still being built. Each generation says again where the boundary lies.