Inside the First Boom and Its Forecasts

Kyoto University's “Phonetic Typewriter” (a machine that recognizes Japanese one sound at a time)

Toshiyuki Sakai and colleagues at Kyoto University studied a “Phonetic Typewriter”, which would recognize a person's voice with a machine and turn it into written characters. A 1960 article says they were studying a device that recognizes 100 single syllables (one sound at a time) of Japanese, judging vowels from the peaks in the sound's resonance (formants) and consonants by combining things such as how many times the waveform crosses zero. A later commentary says it was a circuit of vacuum tubes and transistors as big as a large computer, and that it recognized one sound at a time on the spot.

The Japanese term for “Artificial Intelligence”: in papers from 1962, in the National Diet from 1983

A search of J-STAGE (Japan's database of academic papers) turned up, in a December 1962 paper by Moriya Oda of Nagoya University, the phrase “research on aspects mainly of computer applications, beginning with artificial intelligence” (translated) (a sentence introducing a survey by Minsky and others). Meanwhile, at a 1966 symposium at Tohoku University, the theme was written in English as “Artificial Intelligence”. In the National Diet proceedings, the Japanese term for “Artificial Intelligence” had zero hits through 1982. Its first appearance was in February 1983, when an expert witness brought it up in the phrase “artificial intelligence, fifth-generation computers” (translated).

What we still don’t know:J-STAGE's full-text search cannot search old scanned papers that have no text data, so there may be uses from before 1962. Newspapers (paid article databases such as the Asahi and Yomiuri) and books of the time have not been checked.

A symposium on “Artificial Intelligence” at Tohoku University (more than 200 people)

The Research Institute of Electrical Communication at Tohoku University chose “Artificial Intelligence” as the theme of its 3rd symposium and co-hosted it with the Information Processing Society of Japan and others. There were talks on pattern recognition (Taizo Iijima of the Electrotechnical Laboratory), one of the “recognition machines”, which reads characters (Toshiyuki Sakai and Makoto Nagao of Kyoto University), speech information processing (Osamu Fujimura of the University of Tokyo), “self-organizing systems” that change their own criteria for judgment, like the perceptron (Shoichi Noguchi of Tohoku University), and learning control (Kahei Nakamura of Nagoya University), among others, and more than 200 people took part. The theme was kept in English, and the Japanese term for “Artificial Intelligence” is not used even in the report.

What we still don’t know:The report does not say why the theme was left in English (it is not known whether the Japanese term had not yet been settled or whether English was chosen on purpose).

The first robot to work in a factory, the “Unimate” — in 1961 it takes hot castings out of machines at a GM plant

The patent “Programmed Article Transfer”, which the inventor George Devol applied for in 1954 and which was granted on June 13, 1961, described a mechanical arm usable for many purposes. Instead of cams that could only make fixed motions, or manual operation by people, the sequence of motions (the program) can be recorded and swapped. The patent called this “Universal automation, or Unimation”. The first “Unimate” robot of Unimation, Inc., the company founded by Devol and Joseph Engelberger, was installed in 1961 at a General Motors (GM) plant in Trenton, New Jersey, USA. Its job was to take hot parts out of a die-casting machine (casting by pouring metal into a mold) and stack them. It weighed about 3,000 pounds (about 1.4 tonnes).

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 TV anime “Astro Boy” (Tetsuwan Atom) begins broadcasting — a robot boy with a heart appears on TV every week

Osamu Tezuka's manga “Astro Boy” (serialized from 1952; a separate record) was made into a TV anime, produced by Tezuka's Mushi Production. Broadcasting began on Fuji Television on Tuesday, January 1, 1963 at 6:15 p.m. It was the first Japanese-made series of 30-minute weekly TV anime, and it aired in black and white for 193 episodes until December 31, 1966. Episode 1, “The Birth of Atom”, tells how Dr. Tenma, who has lost his son in a traffic accident, builds a robot that looks just like him, but gets angry that it does not grow and sells it. It recorded a peak audience rating of 40.7% (Tezuka Productions) and was also broadcast in the United States through NBC FILMS (Mushi Production).

What “beat a checkers master” really meant (1962, Samuel's program vs. Nealey)

In 1962, Arthur Samuel's program beat Robert Nealey, a blind checkers player from Connecticut, in one game. As a result, checkers was widely reported as “solved”. According to Schaeffer and colleagues, who later solved checkers completely, Nealey called himself a master, but his actual highest rank was one below master. The game he lost contained a simple mistake, and in a match a year later Nealey won without losing a single game. In 1965, the program played four games each against two men who were fighting for the world championship, and lost all eight. Samuel himself is said to have apologized in a letter for the mistaken belief, which had spread, that checkers had been solved.

Bellman's prediction: “within ten years, checkers will be a completely decidable game”

Richard Bellman, who had criticized Simon and Newell's predictions in 1958, also stated a prediction in a 1965 paper. He wrote that within ten years checkers would become a game in which the best move is completely known (Schaeffer and colleagues present this as a forecast made in response to the reports of Samuel's one win in 1962).

What happened next:It was announced in September 2007 that checkers had been completely solved and that, if both sides play their best, the game ends in a draw. The team that solved it had been computing since 1989.

Checkers “is now solved”: computers prove it is always a draw if neither side makes a mistake

Jonathan Schaeffer and his colleagues at the University of Alberta kept dozens of computers running almost nonstop from 1989 and proved that “Perfect play by both sides leads to a draw.” Checkers has about 5×10²⁰ (500 quintillion) board positions, making it the hardest well-known game solved so far. One of the main tools in the proof was the search method (df-pn) from Ayumu Nagai's 2002 doctoral thesis at the University of Tokyo. Akihiro Kishimoto, who by the time of the paper had moved to a university in Hakodate, Hokkaido, is also among the co-authors.

STUDENT, a program that solves algebra word problems written in English (Bobrow, 1964)

In his doctoral thesis, MIT graduate student Daniel Bobrow developed STUDENT, a program that solves algebra word problems written in English (written in LISP and run on an IBM 7094; his advisor was Marvin Minsky). STUDENT builds equations from sentences written in restricted English and solves them. For example, given a problem asking for the list price of a radio that costs $69.70, which is 15 percent less than the list price, it answered “82 DOLLARS”. When there were not enough equations, it used general knowledge not stated in the problem, such as “(EQUAL (DISTANCE) (TIMES (SPEED) (TIME)))” (distance = speed × time), and told the user the assumptions it had made. First-year high-school textbooks contain many word problems, and Bobrow tested some of them, rewritten into English STUDENT could handle. Bobrow wrote that it solved most of the problems it could handle as fast as a human or faster. He also used the phrase “the computer understands English” with a clear definition: it “accepts input sentences ... and answers questions based on information contained in the input” within that range of sentences (an operational definition, in the paper's words).

What we still don’t know:Bobrow used “understanding” as defined by accepting sentences in the target range and answering questions. Whether a correct answer is the same thing as knowing what it means is a question that is still unsettled even for today's text AI (the fourth boom) (this site's view; it connects to the part about what we still don't know in u13).

ELIZA, a program that converses in text. One aim of its author's paper was to explain away a magic that seems impressive

Joseph Weizenbaum of MIT presented in a paper ELIZA, a program that can converse with a computer in text. It finds keywords in a sentence typed by a person and rearranges them according to rules to produce a reply. The role of the conversation partner can be changed by swapping a “script”, and as of 1966 there were scripts in Welsh and German as well as English. At the start of the paper, the author writes, in effect, that if the inside of the program is explained in plain words the magic crumbles and it is seen to be just a collection of procedures, and that this paper is meant for that purpose.

Is ELIZA's year 1964 or 1966? — the national teaching material's table says “Development of the artificial dialogue system ELIZA (1964)”, but the original paper was received in September 1965 and published in the January 1966 issue. Work began sometime between July 1964 and July 1965

The 2020 teaching material for teachers of the high-school course “Information II” (the advanced high-school informatics course) from Japan's Ministry of Education, Culture, Sports, Science and Technology carries the table “History of Artificial Intelligence (AI)”, reprinted from the Ministry of Internal Affairs and Communications' 2016 White Paper on Information and Communications. There, ELIZA (a dialogue program that answers in text to sentences a person types) is written as “Artificial dialogue system ELIZA developed (1964)” (translated). The source of the white paper's table is the report of a March 2016 survey that the Ministry commissioned from Nomura Research Institute. It has the same table and also a sentence saying ELIZA was “developed at MIT as early as 1964” (translated). The report says this section is based on Hitoshi Matsubara's “The Future Opened by the Third AI Boom” and Yutaka Matsuo's book “Will Artificial Intelligence Surpass Humans?”, but it does not give a source for each row of the table. Meanwhile, the paper by Joseph Weizenbaum, who developed it, appeared in the January 1966 issue of Communications of the ACM (pages 36–45), and its last page says it was received in September 1965. The annual report of MIT's research program Project MAC at the time (for July 1964 to July 1965) says, in the Weizenbaum section, that a major effort on a system in which people and machines converse in natural language began and advanced during this period, and that it can now converse on limited topics. The researchers who found ELIZA's original program in an MIT archive also put the development period at about 1964–1966 and say the version they found dates from 1965. 1964 is possible as a year when the work began, but it is not the year of completion or announcement. The table and the paper point to different moments.

What we still don’t know:The earliest source that put 1964 in a timeline is not known. The 2016 report says this section is based on Hitoshi Matsubara's (President of the Japanese Society for Artificial Intelligence, according to the report's note) “The Future Opened by the Third AI Boom” and Yutaka Matsuo's book “Will Artificial Intelligence Surpass Humans?” (2015), but this site has not been able to open either. The month the work began is also unknown (somewhere between July 1964 and July 1965, going by the annual report). What year Weizenbaum himself gave in later works (1976 and so on) has not been checked by this site either.

DENDRAL, a program with chemical expertise — from general cleverness toward specialist knowledge

Joshua Lederberg, a Stanford University geneticist (a Nobel laureate), had been building DENDRAL, a procedure that uses a tree structure to enumerate the molecular structures possible for a given chemical formula (a 1964 NASA report). Edward Feigenbaum and colleagues, who moved from the University of California, Berkeley to Stanford in 1965, added to it the experience-based knowledge chemists use when reading mass spectrometry data (a method that breaks molecules apart and measures the weights of the fragments). They built that knowledge in as rules of the form “IF ... THEN”, creating Heuristic DENDRAL, which narrows down the candidates. For one compound, more than 14.71 million candidates were reduced to about 1.28 million by the mass spectrometry data, and to one when other measurements were added (Nilsson). In their 1970 paper the three separated the breadth of problems that can be handled (generality), the quality of the answers, and whether they can be produced at a reasonable cost (specialization). They wrote “Generality ... is not often associated with superior performance. The experts usually are specialists.” and chose the path of valuing knowledge of a specific field. This idea led to the expert systems of the 1980s (AI with built-in specialist knowledge).

Zadeh's “Fuzzy Sets” — expressing collections with blurry edges, like “the class of tall men”, with numbers

Lotfi Zadeh of the University of California, Berkeley published the paper “Fuzzy Sets”. A collection with no clear line between being in and out, such as “the class of tall men” or the “class of all real numbers which are much greater than 1”, is not an ordinary mathematical set. But Zadeh wrote that such collections play an important role in human thinking, particularly in recognizing patterns and communicating information. He then devised “Fuzzy Sets”, which express how much an object belongs to the collection as a number from 0 to 1 (the degree of membership). Looking back in 1980, he said the idea came to him in New York in July 1964. He said he knew from the start that the word “fuzzy” has a belittling ring and is unscientific, which would cause problems, but no other word fit. In the late 1980s in Japan the idea spread as “fuzzy” in automatic subway operation and home appliances (a separate record).

What we still don’t know:In 1980 Zadeh himself wrote that criticisms include that it is “merely a disguised form of the theory of subjective probability” and “the lack of an operational definition of the grade of membership”. Who decides the degree of membership, and how, becomes a matter of agreement in each application.

Simon's prediction “machines will be capable, within twenty years, of doing any work” (1965 book／originally a 1960 paper)

In his 1965 book, Herbert Simon wrote that within 20 years machines would be able to do any work a human can do. In the 1960 paper it was based on, he wrote that in a period much shorter than 25 years machines would have the “technical capability” to replace any human function in organizations. Immediately afterward, however, he went on to say that what is technically possible is separate from how it is actually used. He then assumed that, in the long run, nearly everyone who wants to work can work (full employment), and argued that the division of labor between people and machines is decided by “comparative advantage” (each side takes on what it is relatively better at).

What happened next:No source was found saying that by 1985 machines could do any human work. Nilsson (2010) says that, even as of 2010, Simon's similar statement from 1957 (that the range of problems machines can handle would become as wide as the human mind) was still far from coming true. Simon himself, on the other hand, thought of technical capability and employment separately and expected that, over the long run, unemployment would not rise (in his 1966 reply he wrote that the unemployment rate was below 4% and falling).

“The Triple Revolution” memorandum — warning the President that computers and automatic machines will make more and more human labor unnecessary

Thirty-five people calling themselves “The Ad Hoc Committee on the Triple Revolution” sent the memorandum “The Triple Revolution” to President Johnson in a letter dated March 22, 1964 (also to the leaders of the Senate and House of Representatives and to the Secretary of Labor). The three revolutions are “The Cybernation Revolution”, formed by the combination of computers and automatic control machines, the weaponry revolution, and the human rights revolution. The memorandum pointed out that cybernation would create nearly unlimited productive capacity while the amount of human labor needed would keep shrinking. It argued that the root cause of excessive unemployment is that the power of machines is growing faster than the power of many people to keep up. It then proposed that, because the link between jobs and income is breaking, every person and family should be guaranteed a sufficient income as a right. Signers included Nobel Peace Prize winner Linus Pauling, the economist Robert Heilbroner, and Gerard Piel, publisher of “Scientific American”.

What happened next:The national commission's report that came out two years later (February 1966, → The U.S. national commission's report T…) concluded that the high unemployment after the Korean War was caused not by accelerating technological progress but by insufficient growth in demand. Unemployment, which was 5.1% in August 1964 when the law was enacted, had fallen to 4.0% when the report came out. Over the long run, the economist Autor (2015) writes that in 20th-century America the employment rate relative to population rose, and no long-term upward trend in unemployment is seen (this long-term part rests on one summary and is weakly confirmed).

What we still don’t know:Autor (2015) writes that computers, AI and robots may replace human labor on a scale never seen before, and that this cannot be judged by past experience alone. Whether the 1964 warning came too early or was wrong has still not been settled in the age of AI.

The U.S. national commission's report “Technology and the American Economy” — “technology eliminates jobs, not work”

The “National Commission on Technology, Automation, and Economic Progress” (14 members, including chairman Howard Bowen), created by a 1964 law, issued its report to the President and Congress in February 1966. The report first said that the pace of technological change had increased over recent decades but that there had been no sudden break and none was likely in the next ten years (page 1). It also concluded that the cause of high unemployment after the Korean War was not accelerating technological progress but demand (the power to buy things) growing too slowly compared with productivity and the growth in the number of workers (page 9). And it wrote that “technology eliminates jobs, not work”. The unemployment rate, 5.1% when the law was made, had fallen to 4.0% by the time of the report (page 15). At the same time, the report admitted that technology is a major factor in taking away particular people's jobs. It then recommended public jobs with the government as employer of last resort, a minimum guaranteed household income, and a study of a “negative income tax” (pages 40 and 110).

Moore's prediction: the number of components on an integrated circuit roughly doubles every year, reaching 65,000 in 1975 (1965) — later called “Moore’s Law”

Gordon Moore, head of research and development at Fairchild Semiconductor, wrote the article “Cramming More Components onto Integrated Circuits” in the magazine Electronics. The number of components integrated on one circuit (the number at which the cost per component is lowest) had been roughly doubling every year. He forecast that roughly the same pace would continue for at least ten years and that the number would reach 65,000 in 1975. In the same article he also said that integrated circuits would make possible home computers (at least terminals connected to a central computer), automatic controls for cars, portable communications equipment and more. This forecast later came to be called “Moore's Law”, a yardstick for talking about the growth of computing power, including for AI.

What happened next:In 1975, Moore himself said in a talk that the complexity of integrated circuits had roughly doubled every year since their appearance. For the 1975 point he placed a 16,384-bit memory device, alongside the calculation that a yearly doubling would give about 65,000 times. He then revised his forecast, saying growth by the end of the 1970s might be more like “a doubling every two years, rather than every year”. The Computer History Museum writes that this prediction became a goal for the semiconductor industry and turned into “a self-fulfilling prophecy” (a prediction that becomes people's goal and so comes true).

The philosopher Dreyfus's criticism, “Alchemy and Artificial Intelligence”

The philosopher Hubert Dreyfus wrote a survey of AI research as a RAND Corporation report (based on what he said at RAND in August 1964). He pointed out that after the early successes, the research was running into a bigger wall than expected. Humans think using cues at the edge of their field of view (the fringe of awareness), the ability to tell what is important from what is not, and tolerance of ambiguity. He argued that these drop out in computer languages. In his conclusion he wrote that it was time to ask where we stand, so as not to be like the alchemists who kept trying to turn lead into gold. He also wrote that a combination of people and computers is strong, in which people decide where to focus and machines handle the detailed calculation.

What we still don’t know:Dreyfus wrote that big progress in areas such as understanding ambiguous situations “must await computers of an entirely different sort” (the only model available today being the poorly understood human brain). How far today's learning AI has achieved this is judged differently by different people and has not been settled.

The time Dreyfus lost to MIT's chess program (Mac Hack)

Seymour Papert of MIT arranged a game between Hubert Dreyfus and Richard Greenblatt's chess program (Mac Hack VI), and the program won. The newsletter SIGART printed the game record with only one line added, from Dreyfus's 1965 report: “... no chess program can play even amateur chess”. Dreyfus replied that it was not a prediction but a statement of the facts of the time (according to Papert's memo). The program achieved a mid-level amateur result at a 1967 tournament and was made an honorary member of the United States Chess Federation.

SRI's robot “Shakey” — a wheeled robot that sees with a TV camera, makes plans and moves (1966–72)

The Artificial Intelligence Center at the Stanford Research Institute (SRI) studied a wheeled robot, “Shakey”, from 1966 to 1972 (funded by ARPA, the Advanced Research Projects Agency, and others). The robot carried a TV camera, a range-finding device and whisker-like sensors that could tell when it bumped into something. A large computer connected by radio (an SDS-940 in 1969, a PDP-10 and PDP-15 from 1971) built a map (model) of the room from the images, planned routes to destinations and steps for pushing boxes, and made the robot move. Its body shook when it stopped suddenly, so it was called “Shakey” (the shaky one). This research gave rise to the path-finding algorithm A* (a separate record) and the action-planning program STRIPS.

MIT's “THE SUMMER VISION PROJECT” (1966) — what “had undergraduates solve vision in one summer” really meant

Seymour Papert of MIT wrote a plan to be carried out by people working over the summer. The plan was to build a system that, from images from a vidisector (a device that reads images), separates “likely objects” from “likely background areas”, examines the properties of surfaces, and identifies the names of known objects. July was to deal only with non-overlapping spheres, bricks and cylinders (with uniform color of the surfaces and background), and August to widen the targets to cigarette packs and batteries, and then to tools and cups, in a step-by-step plan, with the work assigned to about 11 people. Nilsson records that the task was far harder than expected but became the occasion for vision research to continue at MIT.

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.

Shun-ichi Amari's forecast (1968): “A device with recognition ability close to a human’s ... is unlikely to be made in the near future” (translated)

At the start of a 1968 explanatory article, Shun-ichi Amari wrote that pattern recognition, which identifies characters and sounds, is extremely difficult, that there is not yet any device with discrimination ability close to that of humans, and that the chance of one appearing in the near future is also slim. He also wrote that people's ability to tell characters and sounds apart is not worked out by logic but picked up unconsciously through many years of learning. What is needed, he said, is not a complete logic that covers everything but a rough, efficient one.

What happened next:No deadline is given for the “near future” (translated). As one rough yardstick, there is a paper that examined the 2014 image-recognition competition (1,000 kinds of photos). The error rate of a well-trained person was 5.1%, while the best AI at the time (GoogLeNet) had 6.8%. On a limited task, AI reached accuracy close to a person's (about 46 years later). No source settles when “discrimination ability close to that of humans” (translated), in the broad sense that includes characters, sounds and medical diagnosis, was achieved.

The A* (A-star) pathfinding method — searching by the distance so far plus an estimate to the goal

Peter Hart, Nils Nilsson and Bertram Raphael published the paper “A Formal Basis for the Heuristic Determination of Minimum Cost Paths”. For the problem of finding the lowest-cost route in a graph of connected points, they built a theory that brings in hints from the problem itself (a heuristic: a shortcut based on experience or estimates) and showed that one search method is optimal (from the paper's abstract). According to Nilsson's recollection, for the path search of the robot “Shakey,” they took the idea of using the “airline distance” to the goal as the estimate (Nilsson) and added to it the idea of “the distance traveled so far” (Raphael). Hart then showed that as long as the estimate never exceeds the actual remaining distance, the shortest route is always found. This search method was named A*, and today it is used in places such as car navigation route guidance and games (Nilsson 2010).

SRI's robot “Shakey” — a wheeled robot that sees with a TV camera, makes plans and moves (1966–72)

The Artificial Intelligence Center at the Stanford Research Institute (SRI) studied a wheeled robot, “Shakey”, from 1966 to 1972 (funded by ARPA, the Advanced Research Projects Agency, and others). The robot carried a TV camera, a range-finding device and whisker-like sensors that could tell when it bumped into something. A large computer connected by radio (an SDS-940 in 1969, a PDP-10 and PDP-15 from 1971) built a map (model) of the room from the images, planned routes to destinations and steps for pushing boxes, and made the robot move. Its body shook when it stopped suddenly, so it was called “Shakey” (the shaky one). This research gave rise to the path-finding algorithm A* (a separate record) and the action-planning program STRIPS.

Kunihiko Fukushima (NHK), “Feature extraction by an analog multilayer circuit network” (translated) — an element that gives 0 if negative and passes the value through if positive

Kunihiko Fukushima of the NHK Broadcasting Science Research Laboratories took experiments on how the eyes and brains of cats and monkeys work (by Hubel and Wiesel) as his model and, as a first step toward a machine that reads handwritten characters, recreated a multilayer circuit network on a computer. As the input shape moves through the layers, it is broken down into points, then lines in each direction, then the ends of lines. The element he used outputs the input total as it is if that total is positive, and outputs 0 if it is negative. This has the same form as the ReLU that is widely used in today's deep learning. He also wrote that if everything is built from elements that only add, the network is equivalent to two layers no matter how many layers are stacked.

Automatic reader-sorter for handwritten postal codes (Toshiba and the Ministry of Posts, 1968)

On July 1, 1968, the day the postal code system began, the automatic postal code reader-sorter (TR-3 and TR-4), which Toshiba had developed under the guidance of the Ministry of Posts, was shown at the Tokyo Central Post Office. It is said to be the world's first machine to recognize digits that people had written freely by hand and sort mail with them. The developers analyzed about 300,000 handwritten digits collected from across the country to find which quirks in letter shapes appear most easily, and used an optical filter to erase the red frame so that only the strokes of the digits were extracted. It read the first 3 digits at a speed of about 6 items per second, and characters it could not recognize were marked as “reject” and passed to people.

What we still don’t know:The statement that it was “the first in the world” (translated) comes from the accounts of Japanese academic societies, museums and companies. We have not opened any source that compares it with machines from the same period in other countries.

The first international AI conference (IJCAI-69, Washington) — five talks from Japan as well

The first international conference where artificial intelligence researchers gathered across national borders, the “International Joint Conference on Artificial Intelligence” (IJCAI), was held for 3 days in Washington, D.C., in the United States. The proceedings' table of contents lists 63 papers, divided into sections on pattern recognition (identifying shapes), problem solving by search, theorem proving, question-answering systems, self-organizing systems, robots, language, and more. Nilsson of SRI presented a vehicle-type robot carrying a TV camera (later Shakey; see another record). From Japan, five papers were included, by researchers at the Institute of Industrial Science of the University of Tokyo, the Research Institute of Electrical Communication of Tohoku University, the central research laboratories of NEC (Nippon Electric Company) and Tokyo Shibaura Electric (Toshiba), and Chuo University; three of them were studies on recognizing characters (such as handwritten digits). The conference then continued with London in 1971, Stanford in 1973 and so on, and was held in Tokyo in 1979 (another record).

Minsky's prediction (1967 book): “Within a generation ... the problems of creating ‘artificial intelligence’ will be substantially solved”

Marvin Minsky made a prediction on page 2 of his book “Computation: Finite and Infinite Machines”. He wrote that within a generation few areas of the intellect would remain out of reach of machines, and that the problems of creating “Artificial Intelligence” would be nearly solved.

What happened next:Nilsson (2010) writes that even as of 2000, AI surpassed people in many intellectual tasks but was still far from them in most respects. There is no criterion for judging at what point it could be said that the problems were “substantially solved”, so this has not been settled.

What we still don’t know:The book does not say how many years “Within a generation” means.

Words that spread as Minsky's prediction (Life magazine, 1970): “In from three to eight years ... the general intelligence of an average human being”

Brad Darrach, a reporter for Life magazine, covered SRI's robot “Shakey”. In his article the reporter wrote that Minsky said the following. In 3 to 8 years there will be a machine with the general intelligence of an average person. After that, the machine will teach itself and rapidly rise above the level of a genius. In the same paragraph, the reporter writes that other researchers said “give us 15 years”, but agreed that such a machine would be built.

What happened next:We found no record that, by the deadline given in the article (1973–1978), a machine with the general intelligence of an average person had been built. Nilsson (2010) writes that even as of 2000, AI surpassed people in many intellectual tasks but was still far from them in most respects.

What we still don’t know:We could not find the text of Minsky's rebuttal itself, which is said to have been handed out to researchers. So we do not know whether he denied the words “In from three to eight years”.

Shun-ichi Amari's forecast (1968): “A device with recognition ability close to a human’s ... is unlikely to be made in the near future” (translated)

At the start of a 1968 explanatory article, Shun-ichi Amari wrote that pattern recognition, which identifies characters and sounds, is extremely difficult, that there is not yet any device with discrimination ability close to that of humans, and that the chance of one appearing in the near future is also slim. He also wrote that people's ability to tell characters and sounds apart is not worked out by logic but picked up unconsciously through many years of learning. What is needed, he said, is not a complete logic that covers everything but a rough, efficient one.

What happened next:No deadline is given for the “near future” (translated). As one rough yardstick, there is a paper that examined the 2014 image-recognition competition (1,000 kinds of photos). The error rate of a well-trained person was 5.1%, while the best AI at the time (GoogLeNet) had 6.8%. On a limited task, AI reached accuracy close to a person's (about 46 years later). No source settles when “discrimination ability close to that of humans” (translated), in the broad sense that includes characters, sounds and medical diagnosis, was achieved.

Minsky's forecast in the press materials for the film “2001: A Space Odyssey” (1968) (cited secondhand)

According to Nilsson (2010), Dreyfus wrote the following. In 1968, in the press materials for the film “2001: A Space Odyssey” (directed by Kubrick), Minsky said that within 30 years there should be machines with intelligence about equal to a human's.

What we still don’t know:We could not confirm in the original sources how Minsky was involved in making the film (whether he gave advice). We have not seen the press materials themselves.

The film “2001: A Space Odyssey” and HAL 9000, the computer that talks and thinks — Minsky was on the set

The film “2001: A Space Odyssey,” which director Stanley Kubrick wrote with the author Arthur C. Clarke, was released (distributed by MGM). HAL 9000, the computer on the spaceship Discovery bound for Jupiter, was shown as a machine that speaks in a human voice (the voice was finally done by Douglas Rain). In the story it predicts the failure of an antenna part but is wrong, and then acts in a way that kills crew members (from the AFI Catalog synopsis). In a 1997 news item, MIT introduced its professor Marvin Minsky as “the scientific expert on the set of the 1968 movie” “2001: A Space Odyssey,” and in 1998 the MIT Press included a chapter titled “Scientist on the Set: An Interview with Marvin Minsky” in a book. On the other hand, the person credited as “Scientific consultant” in the film's record (AFI Catalog) is Frederick Ordway, and the designs of the future, such as the computers, were offered free of charge by companies such as IBM.

What we still don’t know:From the sources gathered for this site, we cannot tell how far Minsky was involved in how HAL was portrayed (the text of the interview is behind a paywall). Nilsson (2010, p.163) cites Dreyfus and writes that Minsky said, in the press materials for the 1968 film, that within 30 years there should be machines with intelligence about equal to a human's (record → Minsky's forecast in the press material…, weakly confirmed).

SHRDLU, which talks in English about a world of blocks (Winograd) — the robot and the blocks were inside the screen

Terry Winograd, a graduate student at MIT, developed in his doctoral thesis a program that can hold a dialogue in English (later called SHRDLU). It can answer questions, carry out commands and take in new information. Its subject is a world in which a simple robot with a hand and an eye moves toy blocks and pyramids on a table. When he typed “pick up a big red block.” the program moved away the green block on top, picked the red one up and answered “OK.” To the instruction “grasp the pyramid.” it asked back “I DON’T UNDERSTAND WHICH PYRAMID YOU MEAN.” because there were three. It could also decide that “it” referred to something that had come up earlier in the conversation, and explain the reasons for its own actions. However, the robot and the blocks were simulated on a computer screen. The scene data was supplied from the start, assuming “some sort of robot ‘eye’”; the dialogue was done by typing text at a terminal. The vocabulary was 200 words, and it took 5 to 20 seconds to process one sentence.

What we still don’t know:Winograd's words from 2004 (according to Nilsson): he thought that the gap between a program like SHRDLU and the brain had not been closed much even by the research of the time. How far today's text-writing AI has crossed this gap is a matter on which evaluations differ (this site's view).