The Winter After the First Boom

From which year to which year was the “First AI Winter”? The answer differs by source. Some historians say “There Was No ‘First AI Winter’”

Japan's white paper from the Ministry of Internal Affairs and Communications (2016) calls the 1970s “the winter years” (translated), and Wikipedia gives 1974 to 1980. On the other hand, the researcher Nilsson (2010), who wrote a history of AI, calls about 1975 to 1985 the “boomtimes” of AI and says the “AI winter” came in the late 1980s. And the historian Haigh (2023) argued “There Was No ‘First AI Winter’,” pointing out among other things that the membership of SIGART, the AI special interest group of the computing society ACM, nearly tripled from 1973 to 1978. In fact, an early example of the term “AI Winter” is its use in 1984 as a word expressing worry about a winter that might be coming.

What we still don’t know:We have not found figures that line up total AI research funding in the 1970s by country and by year. Whether it was an “AI winter” changes with what you use as the yardstick (the budgets of large labs／the membership of the society／how the words were used)

Do Japanese researchers from the “AI winter” show up in English-language sources? The Nobel explanation's main text names only Fukushima; Nakano and Amari appear only in the references

In the explanation of the 2024 Nobel Prize in Physics (12 pages), Kunihiko Fukushima's neocognitron appears by name in the main text. But Kaoru Nakano (1972) and Shun'ichi Amari (1972) are only included among the reference numbers (in the range [5-10]) and are not named in the main text. In Nilsson (2010, 707 pages), a leading English-language book on the history of AI, Nakano and Amari are mentioned 0 times, and Fukushima appears just once, in a note (with a page-number error). In the same book, the research of Takeo Kanade and Hitachi's Ejiri is introduced in the main text.

The Japanese Diet during the “AI winter”: ‘Artificial Intelligence’ zero times. What was talked about was ‘artificial brain’, ‘electronic brain’ and ‘pattern information processing’

Looking at statements in the Diet from 1966 to 1980, the Japanese term for “Artificial Intelligence” never appears (0 hits from 1947 to February 22, 1983). Instead, computers were called an “artificial brain” (6 hits) or an “electronic brain” (8 hits). The Diet also discussed the large national project the “Pattern Information Processing System” (12 statements contain “pattern information processing”) and “pattern recognition” (10). In October 1980 it was explained that, as a research result of NHK, Kunihiko Fukushima's theory of the “cognitron” was highly regarded in Japan and abroad.

What we still don’t know:The number of times a word was used in the Diet changes depending on variations in how the proceedings spell it (for example the long and short forms of the word for “computer”, “konpyuutaa” and “konpyuuta”). Here only exact-match search results were counted.

A U.S. committee reports that “there is no immediate or predictable prospect of useful machine translation” (the ALPAC report)

At the request of three U.S. government funding agencies (the Department of Defense, the CIA and the National Science Foundation), a committee of the National Academy of Sciences (ALPAC, chaired by John R. Pierce of Bell Laboratories) examined research on machine translation of documents such as Russian into English. The report wrote that, meaning machine translation that does not depend on human translation or editing, there is “no machine translation of general scientific text, and none is in immediate prospect” (p.19). It recommended directing the budget to two things: “computational linguistics” (the study of how language works) and “improvement of translation” (ways to make human translation faster and cheaper, including help from machines) (p.34).

Makoto Nagao's group at Kyoto University in the 1970s — from reading Japanese sentences and answering questions to TITRAN, machine translation of paper titles

For a question-answering system, Makoto Nagao and Jun-ichi Tsujii of Kyoto University made a program that analyzes Japanese sentences from a chemistry textbook and turns their content into a semantic network (a web of connections of meaning). It was published in 1976 in a US journal of computational linguistics. Its features are that it uses the content of earlier sentences (context) to decide what a pronoun refers to and to fill in omitted words. Unlike many AI systems of the time that relied only on logical formulas, it took the position that human intuitive reasoning is based on the working of language itself. According to Nagao's own 1993 summary, machine translation research at Kyoto University began in 1962. In 1973 the group developed the grammar description language PLATON, and in 1979 TITRAN, a system that translates titles of science and engineering papers from English to Japanese, which was used to translate the literature database INSPEC. TITRAN sorted the structure of titles into a dozen or so “skeleton structures” and rearranged word order, and its error rate was under 5% (Nagao 1982).

The book “Perceptrons” by Minsky and Papert shows mathematically the limits of simple neural networks

A book by Marvin Minsky and Seymour Papert of MIT that examined, with mathematics, the power of the “perceptron,” the learning machine Rosenblatt built in the 1950s (a way of doing pattern recognition with a neural network). They proved that if the number of points each part can see is limited, the machine cannot recognize things more complicated than “one or the other but not both” (XOR), such as “parity” (whether the number of points is even or odd) and “connectedness” (whether a figure is a single connected piece). About multilayer networks, they wrote their “intuitive judgement that the extension is sterile,” while also saying that confirming (or dismissing) this is an important research problem (p.231, as cited by Pollack).

Did “Perceptrons” stop brain-inspired AI research? Opinions are divided

A story often told is that a 1969 book stopped research on machines that learn by imitating the nerves of the brain (neural networks) and brought on a winter of more than 10 years. The Nobel Prize in Physics explanation (2024) writes that the book led to a gap in research funding. On the other hand, Nilsson (2010), who summarized the history, wrote: “Some have claimed that ... but I doubt this”. A sociology-of-science paper (Olazaran 1996) argued the following (summary). The book's proofs were taken to show that neural networks could make no further progress. That was the official interpretation, formed while symbolic AI was taking root as the mainstream and was being accepted as correct. After the revival of the late 1980s came a “rewriting of the official history of the debate”. The authors also replied in 1988 to the common view that the book was to blame: “progress had already come to a virtual halt ... lack of adequate basic theories”.

What we still don’t know:We have not found a source that shows, in numbers limited to neural network research, how much research funding fell. The size of the book's influence is told through people's testimony and interpretation, not through numbers.

The Japanese translation of “Perceptrons” comes out two years after the original (University of Tokyo Press)

Minsky and Papert's book was translated into Japanese in 1971 as “Perceptron: The Path to a Theory of Pattern Recognition” (translated by Saito, University of Tokyo Press, 292 pages). The 1988 expanded edition was also published, in 1993, in a translation by Kaoru Nakano and Sakaguchi (Personal Media). In the National Diet Library catalog, the 1993 translator Kaoru Nakano has the same name as the author of the book on “Associatron” (1979, Shokodo), and the catalog also lists a paper co-written with his co-translator Sakaguchi (1992) (they are thought to be the same person, but the catalog does not say so explicitly).

Shun-ichi Amari's “A Theory of Adaptive Pattern Classifiers” (a learning theory that updates weights bit by bit from errors)

Shun-ichi Amari mathematically analyzed a learning method that, each time it sees one data point, updates the weights (how much importance to give each feature) a little according to the error. The abstract of the English paper describes a proof that the weights approach the optimum even for data that cannot be classified cleanly, the trade-off between learning speed and accuracy, and an extension to piecewise-linear (broken-line-like) discrimination. A 1968 Japanese commentary calls this learning method the “stochastic descent method” (translated). Schmidhuber (2022) writes that it is probably the first paper to train a multilayer neural network by stochastic gradient descent (the basic learning method of today's deep learning).

What we still don’t know:How much a 5-layer network could actually learn (the 1968 book and H. Saito's master's thesis) is not known, because the original sources have not been opened.

The UK's Lighthill Report: “In no part of the field ... the major impact that was then promised”

A report by James Lighthill, a leading figure in fluid dynamics, who surveyed AI research alone at the request of a British research funding body (the Science Research Council, SRC). He sorted AI into A (automation), B (the “bridge” work such as building robots) and C (studies of the brain and nerves using computers), and while he rated A and C, he was harsh on B. As a major reason for the lack of progress, he pointed to the failure to notice the phenomenon in which the number of combinations to be searched grows explosively (the combinatorial explosion). After this report, trust in AI research in Britain and its funding fell sharply.

The UK's Lighthill Report: “In no part of the field ... the major impact that was then promised”

A report by James Lighthill, a leading figure in fluid dynamics, who surveyed AI research alone at the request of a British research funding body (the Science Research Council, SRC). He sorted AI into A (automation), B (the “bridge” work such as building robots) and C (studies of the brain and nerves using computers), and while he rated A and C, he was harsh on B. As a major reason for the lack of progress, he pointed to the failure to notice the phenomenon in which the number of combinations to be searched grows explosively (the combinatorial explosion). After this report, trust in AI research in Britain and its funding fell sharply.

Lighthill's forecast: over the next 25 years the field of AI will largely split apart, and “Such a general-purpose program ... seems as remote as ever”

In Section 4 of his report, “The next twenty-five years”, Lighthill wrote that research tied to applications (such as factory automation and brain research) would advance in close connection with each specialist field. On the other hand, he predicted that “bridge” work such as robot development would lose esteem and research funding, and that AI as a single body of research would break up. In Section 3, he reported the researchers' conclusion that it is unrealistic to expect, within the twentieth century, a general-purpose mechanism that uses a large amount of knowledge and learns by itself, and described a program that can do everything this way: “Such a general-purpose program ... seems as remote as ever.”

What happened next:(1) In the 1970s and 80s, specialization in application areas such as language processing, expert systems and image processing did advance (Nilsson p.265). On the other hand, the body called “AI” did not disappear, and Nilsson calls the period from about 1975 to 1985 the “boomtimes” of AI (p.343). (2) Within the twentieth century, no general-purpose mechanism that uses a large amount of knowledge and learns by itself was built. Training deep neural networks was seen as hard to achieve by many researchers in the 1990s, and the situation changed in the 2000s (Nobel Committee 2024).

Predictions by unnamed scientists in 1972: in the 1980s “an all-purpose intelligence on a human-scale knowledge base”, and by 2000 the possibility of “machine intelligence exceeding human intelligence by the year 2000”

According to Section 3 of the Lighthill Report, “able and respected scientists” had passed predictions to Lighthill in letters. AI is the next step in evolution, they said; in the 1980s a general-purpose intelligence with human-level knowledge might be realized, and by 2000 there might be the “awe-inspiring possibilities” of machine intelligence exceeding that of humans. Lighthill gave the fact that such predictions were still being made as late as 1972 as a reason why past predictions should be compared with actual results.

What happened next:1980s: general-purpose intelligence was not realized, and falling short of expectations was one cause of the “AI winter” of the mid to late 1980s (Nilsson p.408). During that winter, researchers turned to more modest goals and bold predictions became fewer (p.429). 2000: by the stated deadline, machine intelligence did not surpass humans overall (it was from the 2000s onward that deep neural networks could be trained in earnest; Nobel Committee 2024 p.7).

What we still don’t know:The report does not tell us who the scientists who wrote the letters were.

A public debate on television: Lighthill against Michie, McCarthy and Gregory (BBC “Controversy”, Royal Institution)

On July 4, 1973, a public debate on the content of the report was held at the Royal Institution in London. The BBC recorded it as an episode of its TV program “Controversy” and broadcast it on BBC2 on August 30. The title of the debate was “The general purpose robot is a mirage”, and members of the public filled the seats. Lighthill explained his report and spoke about the difference between automation and the general-purpose robot, which he himself called “the mirage”. Donald Michie of the University of Edinburgh, John McCarthy of Stanford University and Richard Gregory of the University of Bristol argued against him. Michie showed film of the robot Freddy II assembling a model. The recording is about 81 minutes long and is made public by the University of Edinburgh.

The U.S. Department of Defense's DARPA five-year program for a “speech understanding system”: a machine met the goals, yet the follow-up program was cancelled

For five years from 1971, DARPA put about $3 million a year of research funding into research on machines that understand connected speech (SUR). The goals included a vocabulary of 1,000 words, many speakers, and a semantic error rate under 10%. In 1976, Harpy of Carnegie Mellon University (CMU) achieved a 5% semantic error rate with 1,011 words, clearing the goals. But DARPA did not fund the follow-up program that had been planned, and it was not until 1984 that it restarted speech research on a large scale.

What happened next:In September 1976, CMU's Harpy achieved 1,011 words and a 5% semantic error rate, meeting the lowered goals (CMU report). The original concept of 10,000 words over the telephone was not realized within this program. DARPA did not fund a follow-up program.

A U.S. law decides that defense funds may be used only for research with “a direct and apparent relationship to a specific military function or operation” (the Mansfield Amendment)

Around the time of the Vietnam War, the U.S. Congress, in Section 203 of a law on the defense budget (Public Law 91-121), stated that funds under that law could be used only for research with “a direct and apparent relationship to a specific military function or operation.” It is called the “Mansfield Amendment” after the name of its proposer. ARPA (later DARPA) of the Department of Defense, which had been paying out much of the funding for AI research, came to be asked to show that its research was useful to the military.

Hiroaki Sakoe and Seibi Chiba of NEC present a way to recognize speech by lining up differences in speaking speed with “dynamic programming” (DP matching)

Even for the same word, the speed at which people speak changes. Hiroaki Sakoe and Seibi Chiba of the NEC Central Research Laboratories showed a method that lines up the stretching and shrinking of speech in time using dynamic programming (a calculation method that efficiently finds the best procedure among many candidates) and compares the result with registered speech. In their 1971 paper, it correctly recognized an average of 99.8% in 2,400 utterances of one-digit numbers spoken by 5 people, and 99.6% in 500 utterances of two-digit numbers spoken continuously. In 1978 they published an improved version in an IEEE journal.

Japan's national large-scale research project, the “Pattern Information Processing System,” begins (fiscal 1971–1980): building computers that recognize patterns

In 1971, when AI research funding was being cut in the United States and the United Kingdom, Japan's Ministry of International Trade and Industry added the “Pattern Information Processing System” to its large national projects paid for with public money. The plan was to develop computers that recognize characters, figures and the like, centered on the Electrotechnical Laboratory in cooperation with private companies, and in the Diet it was also described as “fourth-generation computer” (translated). The explanation at the start was “the state will pay the full amount, about 35 billion yen over eight years” (translated) (about 35 billion yen over 8 years, all paid by the state), but in a 1985 Diet answer the commissioned research costs for fiscal 1971–1980 were given as about 22 billion yen.

What we still don’t know:We have not opened any report on the detailed contents of the plan (what kind of devices were built and what was achieved). The evaluation of its results is unconfirmed.

Hitachi Central Research Laboratory's robot HIVIP: reads drawings from three directions, finds the parts and assembles them by itself

Masakazu Ejiri and colleagues at the Central Research Laboratory of Hitachi (Hitachi Ltd.) developed the robot HIVIP, which has three parts: an eye (EYE), a brain (BRAIN) and a hand (HAND). The first camera looks at an assembly drawing drawn from three directions and reads off the finished shape, the parts and the order of assembly. The second camera recognizes the shape, position and orientation of parts scattered on a table. A computer decides the procedure, and an arm with 7 degrees of freedom (directions of movement) does the assembling. The objects it could handle were limited to simple polyhedra (angular, block-like shapes). The paper states that for assembly and inspection in future factories, it will be necessary to realize the functions of the human eye and hand with machines.

Kaoru Nakano (University of Tokyo)'s “Associatron”: a brain-inspired memory that recalls the whole when shown a part

Kaoru Nakano of the University of Tokyo presented the “Associatron,” a mechanism that stores information as patterns of 0s and 1s spread across a whole network and recalls the whole from a part of it. The larger the clue, the more accurately it recalls, and the smaller the clue, the more vaguely. It is a simplified model of the network of nerves in the brain, and he showed it could be built with elements connected only to their neighbors working at the same time. He ran experiments on a computer and also showed examples of applying it to concept formation and to games.

Shun-ichi Amari (University of Tokyo)'s “self-organizing nets”: memorizing patterns and their sequences as stable states

Shun-ichi Amari of the University of Tokyo studied theoretically how a network of threshold elements modeled on nerve cells “remembers” patterns, and sequences of patterns, shown from outside, as settled states and fixed state transitions. He showed the conditions under which many patterns can be stored at the same time and whether they can be recalled even with noise. The introduction to the Kyoto Prize (2025) states that it is mathematically equivalent to the associative memory model (a network that recalls the whole from a part) that Hopfield published ten years later, in 1982, and made famous.

A Japan–US research meeting on learning machines in Nagoya: 28 papers on pattern recognition and learning control

In August 1970, the “Japan—U.S. Seminar on the Learning Process in Control Systems” was held in Nagoya. As a Japan–US science cooperation project, it was supported by the National Science Foundation of the United States and the Japan Society for the Promotion of Science. 28 papers were submitted on pattern recognition (mechanisms for telling things apart) and on control that improves by learning from experience (learning control). From Japan, talks included unsupervised learning, handwritten character recognition with neural networks, Kaoru Nakano's associative memory model, and learning control using fuzzy methods.

Kunihiko Fukushima (NHK Broadcasting Science Research Laboratories)'s “cognitron”: a multilayer network that assembles itself without a teacher

Kunihiko Fukushima of the NHK laboratory proposed a new idea about how the connections between nerve cells in the brain get stronger. Using it, he showed a way for a multilayer neural network to organize itself without a “teacher” (someone who tells each cell in detail how to react). When he named it the “cognitron” and tried it on a computer, deeper layers covered wider areas, and the neurons in the last layer came to respond only to particular patterns.

Kunihiko Fukushima's “Neocognitron”: a multilayer neural network that tolerates shifts in position, regarded as a source of today's image-recognition AI (CNN)

Kunihiko Fukushima of the NHK research laboratories developed the Cognitron further and presented the “Neocognitron,” a multilayer neural network that can identify patterns such as characters even when they are shifted in position or slightly distorted in shape. Drawing on studies of the brain's vision (Hubel and Wiesel), he stacked many pairs of layers: layers of S cells that extract features and layers of C cells that absorb shifts in position. Cells in the same plane hold the same connection weights, just shifted in position. It learns without a teacher, and he tested a 7-layer network on a computer. The explanation for the 2024 Nobel Prize in Physics names this work as a source of today's image-recognition AI (convolutional neural networks, CNN).

Takeo Kanade's doctoral thesis (Kyoto University): a computer automatically detects eyes, nose and mouth in face photos and tells 20 people apart

Takeo Kanade of Kyoto University developed an image-processing system that connected a mid-size computer to a fast minicomputer. He wrote a program that automatically detects feature points such as the eyes, nose and mouth in a face photo. He added a “feedback” device that goes back to an earlier stage and tries again when things do not work out, and tested it on about 800 face photos (688 of them were faces collected at the 1970 Osaka Expo). He then ran an experiment to tell the faces of 20 people apart (2 photos each, taken a month apart in different places) using 16 numbers calculated from the feature points.

The Osaka Expo's “Computer Magnifying Glass”: a show that automatically reads a face captured by a camera and answers with a personality type

At the Sumitomo Fairy Tale Pavilion at the 1970 Osaka Expo, NEC created and ran a show called “Computer Fortune-Teller's Magnifying Glass.” A TV camera took pictures of visitors' faces, and a computer (NEAC 2200 Model 200) extracted lines from the pattern of light and dark and detected the nose, mouth, chin and so on. From three angles of the face, it chose one of 21 “basic personality types” (translated) prepared by the show's makers and told the visitor by voice and by a lamp showing a caricature of a famous person. The mechanism and programs were developed with the cooperation of Professor Toshiyuki Sakai and Assistant Professor Makoto Nagao of Kyoto University. Part of the face data collected here was later used in Takeo Kanade's research on faces.

What we still don’t know:The idea that personality can be told from the face (physiognomy) itself has no scientific backing. The judging rules of this exhibit were prepared by the makers (the producer); what the technology showed was only the step of automatically finding the parts of a face.

Waseda University's WABOT-1: a humanoid robot that walks on two legs, grasps with hands, sees with eyes and moves on spoken commands in Japanese

A research group led by Ichiro Kato at Waseda University defined a robot as “a machine in which working hands, moving legs and a thinking head are one ... that acts autonomously” (translated) (a machine in which hands, legs and head are one, that takes in information from outside and inside and moves by itself), and developed the humanoid robot WABOT-1. It walked on two hydraulically driven legs, grasped objects with both hands, and located objects with artificial eyes. It carried a small computer (a minicomputer) as its brain, recognized Japanese commands spoken with pauses, and answered by voice. A 1973 report states that, carrying a 20 kg load, it could go straight and change direction as ordered by voice.

Masahiro Mori's “Uncanny Valley” — is a robot that looks too much like a person actually creepy?

Masahiro Mori, then a professor of robotics at the Tokyo Institute of Technology, published the essay “The Uncanny Valley” in the magazine “Energy” of Esso (the Japanese company of Standard Oil), in an issue on robots and ideas. He began by saying that people tend to think things behave like a “monotonically increasing function”, the way income grows as effort grows, but that sometimes there is a valley partway, like on a mountain path. Then, with a graph, he showed the idea that as a robot's looks and movements come closer to a person's, it feels more familiar, but just before it becomes a perfect likeness there is a valley where the feeling suddenly turns eerie. As an example, he gave a prosthetic hand that cannot be told from a real one but feels creepy the moment you realize it is artificial. He then advised robot designers to aim for “the first peak” of moderate human-likeness and to deliberately make robots non-human in form, rather than trying to cross the valley to look exactly like a person.

What we still don’t know:Why do people feel that something that looks too much like a person is creepy? There are studies, using brain waves and other measures, that show the valley exists. But in 2012 Mori himself said of the brain-wave studies, “It still doesn’t explain why we feel eerie to begin with”.

Japan's first domestically built industrial robot, the “Kawasaki Unimate 2000” — a big response in a Japan short of workers

In 1966, Kawasaki Heavy Industries invited Engelberger of Unimation in the United States and held a seminar. Against expectations it was a great success and set off the industrial robot boom in Japan (Japan Institute of Invention and Innovation). Kawasaki Heavy Industries formed a technology partnership with the company in 1968 and in 1969 completed the first unit of Japan's first domestically built industrial robot, the “Kawasaki-Unimate 2000”, at its Akashi Works. It drew a big response as something to make up for the labor shortage of the period of rapid economic growth (Kawasaki Heavy Industries). At first it stayed at the evaluation stage, but after spot-welding robots were delivered to Nissan Motor and Toyota Motor in 1973, it began to be used in earnest at car factories (Japan Institute of Invention and Innovation).

What we still don’t know:Even for the same industrial robots, in the United States workers worried that robots would “take people’s jobs” (translated) (Japan Institute of Invention and Innovation). In Japan during the high-growth period, on the other hand, they drew a big response as the arrival of a “robot to make up for the labor shortage” (translated) (Kawasaki Heavy Industries). How the reception of the same technology changes with a country's labor situation is a question that connects to today's AI as well.

The manga “Doraemon” begins serialization — a cat-shaped robot from the 22nd century

Fujiko F. Fujio's manga “Doraemon” began serialization in Shogakukan's grade-by-grade study magazines (gakunen-shi). The first episode, “Coming All the Way from the Land of the Future”, appeared in the January 1970 issue of “Shogaku Yonensei” (the magazine for fourth graders), and other episodes appeared in the same month in magazines such as “Shogaku Sannensei” (for third graders). Later, stories drawn differently for each grade were serialized at the same time in six magazines, from “Shogaku Ichinensei” (for first graders) to “Shogaku Rokunensei” (for sixth graders). The serialization continued for 26 years, mainly in the grade magazines (Shogakukan). Doraemon is a cat-shaped robot who comes from the future to help the lazy boy Nobita. The official site gives his birthday as September 3, 2112.

McCarthy and Hayes name the problem of having to write down everything that does not change (the frame problem)

In AI that represents the world with logical formulas and makes plans, you have to describe not only what changes because of an action but also each fact that stays the same. John McCarthy and Patrick Hayes named this difficulty the “frame problem”. They pointed out that with n actions and m quantities describing the state, you could end up having to write m×n conditions saying that things stay the same.

What we still don’t know:The frame problem in its technical sense is now said to be largely solved (SEP, Shanahan). The philosophical sense (narrowing down what is relevant), on the other hand, is said to be still under debate. We have not yet looked into how today's large-scale AI handles this.

From which year to which year was the “First AI Winter”? The answer differs by source. Some historians say “There Was No ‘First AI Winter’”

Japan's white paper from the Ministry of Internal Affairs and Communications (2016) calls the 1970s “the winter years” (translated), and Wikipedia gives 1974 to 1980. On the other hand, the researcher Nilsson (2010), who wrote a history of AI, calls about 1975 to 1985 the “boomtimes” of AI and says the “AI winter” came in the late 1980s. And the historian Haigh (2023) argued “There Was No ‘First AI Winter’,” pointing out among other things that the membership of SIGART, the AI special interest group of the computing society ACM, nearly tripled from 1973 to 1978. In fact, an early example of the term “AI Winter” is its use in 1984 as a word expressing worry about a winter that might be coming.

What we still don’t know:We have not found figures that line up total AI research funding in the 1970s by country and by year. Whether it was an “AI winter” changes with what you use as the yardstick (the budgets of large labs／the membership of the society／how the words were used)

The computer program BKG 9.8 beats the world backgammon champion 7 to 1, though its author wrote that this time the program was “the beneficiary of the element of luck”

On July 15, 1979, in Monte Carlo, BKG 9.8, a backgammon program (backgammon is a board game played with dice) developed by Hans Berliner of Carnegie Mellon University, played a match to 7 points against Luigi Villa, who had just become world champion, and won 7 to 1. The author wrote that this was the first time a computer had beaten the officially recognized world champion of a game. He also wrote, however, that the champion clearly played better in the match and that the program was favored by the luck of the dice this time.

The Stanford Cart: avoids obstacles using only camera images, but takes 10–15 minutes to move 1 m

At Stanford University's Artificial Intelligence Laboratory, Hans Moravec had a cart carrying a television camera move through a space with chairs and other obstacles, avoiding them using only the camera images. From pictures taken from several angles, it worked out the positions of objects in three dimensions and planned a path around them. It moved a little, stopped to take pictures, computed for a long time, and then moved forward again. It took 10–15 minutes to move 1 m, and a 20 m course with 3–4 obstacles to dodge took about 5 hours.

The first computer shogi (Japanese chess) program: its early strength was “amateur 20-kyu” (translated)

At the request of NEC, in 1974 a research group at Waseda University made up of Professor Takahashi, the student champion Hoshino and the graduate student Takenobu Takizawa began developing a shogi program. Toshio Ochi of Hitachi and others had already developed software that solves tsume-shogi (problems of checkmating the king), but this was something different. At first it had only the most basic method of choosing moves by looking ahead (the minimax method: reading ahead on the assumption that you pick the move with the highest score and your opponent picks the move with the lowest). Its strength was “complete beginner level” (translated), at best amateur 20-kyu. After that, groups at Osaka University and the Tokyo University of Agriculture and Technology also joined the development. Takizawa improved the way it looked ahead, but when it played the elementary-school champion in a 1984 TV program, it lost badly and was rated 5-kyu (Takizawa and Kotani 2009).

Searle's “Chinese Room”: if you answer correctly by manipulating symbols, can you be said to understand the meaning?

The philosopher John Searle proposed a thought experiment. Imagine that he, who does not know Chinese, shuts himself in a room and, using only a rulebook written in English, answers Chinese questions passed in to him, in Chinese. From the outside he cannot be told apart from someone who understands Chinese, but inside the room he understands nothing. So, he argued, a computer that merely manipulates symbols by following a program cannot be said to understand meaning either. The paper was published in a journal together with 27 replies and comments.

What we still don’t know:The question of whether a machine that merely manipulates symbols correctly can be said to understand meaning has been debated from 1980 to the present with no agreement (SEP 2024). It is a question that applies just as it is to today's text-generating AI, and it can serve as a leading example of what we still don't know.

Weizenbaum, the creator of ELIZA, warns in a book: “tasks which computers ought not be made to do, independent of whether” they can be made to do them

Joseph Weizenbaum of MIT, who made the program ELIZA that seemed to be talking with a person, was surprised that many people took ELIZA seriously and confided in it, and he took a position critical of AI. In his 1976 book “Computer Power and Human Reason: From Judgment to Calculation”, he wrote that humans and machines are different and that truly human choices and judgments must not be left to machines.