What did the Fifth Generation Computer national project aim for, and what did it leave behind? (1979–1995)
In 1982, Japan launched a national plan for a new kind of computer that would speed up inference (reaching answers from rules) by running it on many computers at once. The whole world reacted. Open the sources and you find that depending on what you count as the goal, the project can look like a success or a failure. We run the mechanism and check it for ourselves.
011946–1981 · The name
Why "Fifth Generation"?
What were computer "generations" counted by?
Computers were sometimes counted in generations by the electronic parts they used. A booklet published in 1981 by the Japan Information Processing Development Association (JIPDEC) sorted them into the first generation (vacuum tubes), second (transistors), third (ICs), 3.5th (LSI) and fourth (VLSI). It said that the world was then at the dawn of the fourth generation, and that nearly all of these machines followed the von Neumann architecture that John von Neumann described in 1946.
A von Neumann machine keeps instructions and data in the same memory and fetches the instructions one at a time, running them in order. The booklet argued that a fifth generation, (translated), was needed to get past the weakness of this structure. Press the generation boxes in the figure on the right. Only for the fifth generation are the parts described as (translated), and its contents as (translated), meaning inference, knowledge bases and intelligent interfaces. It was a change of generation in ways of thinking, not in parts.
The survey for the project began in fiscal 1979 (the chair was Tohru Moto-oka of the University of Tokyo). In August of the same year, 1979, the 6th International Joint Conference on Artificial Intelligence (IJCAI) was held in Tokyo. The five earlier meetings had been in Washington, London, Stanford, Tbilisi and Cambridge.
Who chose the name
The 1993 final evaluation says the name was chosen to convey technology that was creative and pioneering by international standards. An explanation in the ICOT archive (Shun'ichi Uchida, 2005) says it was Masataka Nakano of the Ministry of International Trade and Industry (MITI) who chose it. We have not confirmed this in contemporary documents.
Sources count generations differently. Nilsson (2010) has no 3.5th generation: he makes the third generation small-scale integration (SSI) and the fourth VLSI. You can switch to this with "Nilsson 2010" in the figure on the right.
Sources & confidence (2)Why "Fifth Generation"?
1981Verified in the originalSources disagreeJapan
Why “The Fifth Generation”? ── the name that comes after vacuum tubes, transistors, ICs, LSI, and VLSI
Computers are sometimes counted in generations by the electronic element they use. A 1981 JIPDEC booklet sorted them into the first generation (vacuum tubes), the second (transistors), the third (integrated circuits, ICs), the 3.5th (large-scale integrated circuits, LSI), and the fourth (very-large-scale integration, VLSI). It explains that at the time the fourth generation was just beginning, and that most of those computers were von Neumann machines (the design idea von Neumann presented in 1946). The same booklet describes the fifth generation as computers (translated), and in its diagram of generations it wrote the element only as (translated). It then gives their content as (translated) (inference processing, knowledge-base management, and intelligent-interface management). The 1993 final evaluation report says the name “Fifth Generation Computer” was chosen to mean a technology that was new and a pioneer even by international standards. The ICOT Archive's explanation says the person who named it was Masataka Nakano of the Ministry of International Trade and Industry (MITI).
What we still don’t know:That Masataka Nakano gave the name has been confirmed only from the ICOT Archive's explanation (a later look back, with a note at the end that Shunichi Uchida wrote it on December 1, 2005), not from documents of the time. The year and month the name first appeared in a document has not been confirmed either.
4 sources and how they were checked
Where sources disagree:How generations are counted: the JIPDEC booklet puts in a 3.5th generation (LSI) and makes the fourth generation VLSI. Nilsson makes the third generation small-scale integration (SSI) and the fourth VLSI, with no 3.5th generation.
Computers of the Fifth Generation (booklet) (Japan Information Processing Development Center (JIPDEC) (incorporated foundation) · 1981 · PDF page 3 Introduction · Original source)
The AI world's big conference, the International Joint Conference on Artificial Intelligence (IJCAI), is held in Asia for the first time, in Tokyo
The sixth IJCAI, the international AI conference held every two years, took place in Tokyo on August 20–23, 1979. The previous five were held in Washington, London, Stanford, Tbilisi, and Cambridge, so this was the first in Asia. Kokichi Tanaka of Osaka University served as General Co-Chairman, and Saburo Tsuji of the same university served as Program Co-Chairman (for Asia) (names as romanized in the proceedings). The presentations included AI that builds in experts' knowledge (such as PROSPECTOR, which searches for minerals), and a panel discussion on Inexact Reasoning (four speakers: R. Davis, S. Kaihara, E. Shortliffe, and L. Zadeh).
What we still don’t know:We have not confirmed which venue in Tokyo it was held at, or how many people took part.
What the figure shows:Computer generations 1–4 were counted by the components used, but generation 5 was named for what the machine would do, reason with knowledge. It was a change in the way of thinking.
021981 · Goals
A computer that sees, hears, understands
In 1981, what kind of computer was the Fifth Generation said to aim for?
The November 1981 JIPDEC booklet said the fifth-generation computer was (translated), and that research had to start early in order to (translated), that is, to build a trial machine. It lists 8 functions as the intended picture. They include a (translated) that handles speech and images directly, a (translated) that can converse in natural language and make (translated), and a computer that has common sense and solves problems by inference.
The figure on the right is a small working example that splits the into 3 stages (① turn a phrasing into a logical question, ② infer, ③ look up knowledge). If the phrasing changes only a little, it stops at ①. And for the question “Older?”, it cannot answer unless the common sense that a parent is older than a child has been registered.
The 1993 final evaluation report records that, during the research, it became clear that handling speech and images as knowledge requires accumulating and organizing a huge amount of background knowledge.
What you think you knowThe popular story:The Fifth Generation promised to build a computer that could talk like a person in 10 years The original source says:The 1981 document does say (translated) and (translated). But the director, Kazuhiro Fuchi, said in 1992 that the plan that actually began in 1982 had dropped the development of application systems such as machine translation from its goals and narrowed its research to inference machines. The goals that were announced differ depending on which year's document you look at.
The 1981 survey report from the same year
It describes machine translation as a field that had (translated), but whose research then declined once the difficulty was understood, and it analyzed insufficient computer performance as the main cause. The idea was to get over that wall with a new kind of computer.
We have not found the 1981 (translated title, 1981) that is said to have set numerical targets such as numbers of translated words.
Sources & confidence (2)A computer that sees, hears, understands
March 1981Verified in the originalJapan
The 1981 forecast: it is said that in the early 1990s computers with eyes, ears, and a mouth, that understand speech and can do automatic translation and interpreting, will appear
The report of the survey before the plan (1979–1981; chaired by Tohru Moto-oka of the University of Tokyo) stated that the goal was to complete a prototype in the year 1990. A booklet explaining it for the general public wrote that fifth-generation computers are (translated) and that research must start early to (translated) (to build a trial machine). As the pictures of what they aimed for, it listed eight functions, including (translated) that can handle sound, figures, images, and characters as they are; (translated) that can be spoken to in everyday language (and so can also do automatic translation and interpreting); computers that have common sense and solve unfamiliar problems by reasoning; and computers that can be used without writing programs.
The popular story:It is said that the Fifth Generation promised to build, in 10 years, a computer that can talk like a human The original source says:The 1981 survey documents do say “computers that understand speech” (translated), (translated), and (translated) (a trial machine). However, in 1992 the director, Kazuhiro Fuchi, said that in the project that actually began in 1982, building applications such as machine translation had been dropped from the goals and the project was narrowed to machines for reasoning. The story changes depending on which year's, and which document's, goals you mean.
What happened next:The project ran in fiscal years 1982–1992. According to the 1993 final evaluation report, the goal for inference speed (100 million per second) was met and exceeded, at over 150 million per second. On the other hand, the report says that systems that interact with people by voice or images turned out to need a huge amount of background knowledge, so partway through the project narrowed its focus to a base for language research and did not get as far as building concrete systems. In 1992 the director, Kazuhiro Fuchi, said that building application systems such as machine translation had been dropped from the goals when the project began in 1982.
What we still don’t know:The (translated) (1981; 7 groups, 26 tasks), which are said to set numerical goals such as the number of words translated and accuracy, could not be found and have not been checked.
Computers of the Fifth Generation (booklet) (Japan Information Processing Development Center (JIPDEC) (incorporated foundation) · 1981 · p.参1 Reference: Functions of fifth generation computers · Original source)
30 March 1993Verified in the originalJapan
The Fifth Generation final evaluation report: inference speed exceeded the target, but a system for talking with people was not built
The committee that had followed the project's progress (the Committee for the Promotion of Basic Computer Technology Development) compiled a final evaluation report at its end. It reported that PIM, a machine that runs many computers in parallel to carry out inference (drawing conclusions from rules), exceeded the target of 100 million inferences per second, reaching more than 150 million. On the other hand, for systems that deal with people through speech and images, it was found that a huge amount of background knowledge is needed. So partway through, the work was narrowed to the basics of natural language processing, and the report records that it did not go as far as building a concrete system. Overall, it sums up that the project (translated).
The popular story:“The Fifth Generation ended without reaching its goals,” people tend to say, lumping everything together The original source says:The government's final evaluation writes goal by goal. The speed of inference (100 million per second) was exceeded. It says itself that a system for talking with people through speech and images was not built.
What we still don’t know:This evaluation was written by the side that ran the project (the government committee). For outside evaluations, see other records (the MITI Research Review 1995, Nilsson 2010, and others).
2 sources and how they were checked
Fifth Generation Computer Project: Final Evaluation Report (Academic and Technical Evaluation Working Group, Committee for the Promotion of Basic Computer Technology Development · 1993 · p.10 Degree of achievement of the original goals for parallel sy… · Original source)
What the figure shows:The “computer that understands what is said” described in a 1981 document stops when phrasing changes slightly and cannot answer without common sense everyone knows. The project was later narrowed to building reasoning machines.
This is the same mechanism as "Try it", written as Python code. Press ▶ Run and the result appears below. Change the numbers and run it again.
u07_knowledge_pipeline.py
031981–1995 · The institute
From a conference to the institute ICOT
Who moved a national project, and how?
On October 19–22, 1981, the (translated title) was held in Tokyo. According to the organizers' booklet, about 300 people from 15 countries came.
In April 1982 the foundation “Institute for New Generation Computer Technology (ICOT)” was established, and in June the institute began its research. Kazuhiro Fuchi came from the Electrotechnical Laboratory to serve as director. The researchers were sent by the Electrotechnical Laboratory, Nippon Telegraph and Telephone Public Corporation and 8 computer companies; the institute started with about 40 people, who were as a rule swapped out every 3 to 4 years. All of the funding was commissioned expenses from the Ministry of International Trade and Industry (MITI).
The figure on the right shows the institute's seats and the people rotating through them. Shun'ichi Uchida (2005) writes that about 200 people were seconded to ICOT over 13 years, extensions included. If 40 seats turn over every 3 to 4 years, you get roughly that number.
People passed through ≈ S × (1 + (T − 1) ÷ L)
S
number of seats
T
number of years (13)
L
years one person stays (3–4)
The conference and the people
The MITI Research Review (1995) says the conference was well received and MITI also began to take the project seriously.
The same account (Uchida 2005) also says that about 80 of the participating researchers later moved to universities. About 200 people also took part in the working groups that linked ICOT and the universities.
We have not confirmed how the number of seats changed over time, so the seat counts in the figure are provisional. The figure of about 40 people is the number Nilsson (2010) quotes from ICOT's 1983 report; we have not confirmed that report itself.
Sources & confidence (2)From a conference to the institute ICOT
19 October 1981Verified in the originalJapan
The in Tokyo: about 300 people from 15 countries gather, and Japan's plan reaches the world
An international conference to present Japan's draft national project (the Fifth Generation Computer) to the world was held in Tokyo. According to the organizers' booklet, about 300 people from 15 countries took part. Research chiefs of major computer companies such as IBM and DEC (United States), Siemens (Germany), and Olivetti (Italy) also came to Japan. A 1995 report from the Ministry of International Trade and Industry's research institute says that the conference was well received, and that the government then moved in in earnest. Later, the director, Kazuhiro Fuchi, looked back and said that the dream told at this conference reached the whole world in a very shocking way.
4 sources and how they were checked
Computers of the Fifth Generation (booklet) (Japan Information Processing Development Center (JIPDEC) (incorporated foundation) · 1981 · p.26 6. Trends in new computer development in other countries (t… · Original source)
ICOT (the Institute for New Generation Computer Technology), the Fifth Generation research institute, is founded, with Kazuhiro Fuchi as director
To carry out the government's (Ministry of International Trade and Industry's) plan, the foundation “Institute for New Generation Computer Technology (ICOT)” was established in April 1982, and the research institute began its activities in June. The director of the institute was Kazuhiro Fuchi, who came from the Electrotechnical Laboratory. The researchers were people sent for a few years each from the Electrotechnical Laboratory (ETL), the Nippon Telegraph and Telephone Public Corporation (NTT), eight computer companies, and others, and at first there were about 40 of them. All of the research funds were commissioned funds from the Ministry of International Trade and Industry.
What the figure shows:With about 40 positions changing hands every 3–4 years, about 200 people passed through the institute over 13 years and then returned to companies and universities.
Cumulative number of people who came (by year)
041972–1984 · Logic programming
Answering with facts and rules
Is writing a program only a matter of writing steps?
In 1972, Colmerauer and Roussel in Marseille created Prolog, a language in which lining up sentences of logic makes them a program as they stand. Kowalski, who worked on it with them, looked back that summer and said .
In Prolog you write not the steps of a computation but facts (Taro is a parent of Hanako) and rules (if X is a parent of Y and Y is a parent of Z, then X is a grandparent of Z). When a question arrives, the computer tries the rules from the top, and if one does not fit it goes back and tries the next (backtracking). Press "Step" in the figure on the right.
The Fifth Generation put logic programming at its center. According to Nilsson (2010), Koichi Furukawa, who later became ICOT's deputy director, learned of Prolog at SRI in the 1970s and brought an implementation back to Japan. In 1984, Kazuhiro Fuchi said he was growing ever more confident in having chosen Prolog.
grandparent(X, Z) ← parent(X, Y), parent(Y, Z)
X, Y, Z
variables
←
if everything on the right holds, the left holds too
The name
Prolog is a shortened form of the French PROgrammation en LOGique (programming in logic). Nilsson writes that Prolog was used mainly outside the United States and became popular in Europe, Canada and Japan.
We could not open Colmerauer and Roussel's own memoir paper for this site.
Sources & confidence (2)Answering with facts and rules
1972Verified in the original
The logic language “Prolog” is born in Marseille: line up logical sentences and you have a program
In Marseille, Alain Colmerauer and the graduate student Philippe Roussel developed Prolog, a language in which writing logical sentences (if ... then ...) one after another makes them directly into a program. The name is short for PROgrammation en LOGique (programming in logic). It grew out of ideas shared with Robert Kowalski, who was visiting from Edinburgh. Looking back on his second visit in the summer of 1972, Kowalski said that .
The words of the Fifth Generation's director, Kazuhiro Fuchi, himself (1984–1988) ── long skeptical about AI, the proposal is a (translated), no direct involvement in the AI business, and (translated) bring a backlash
Kazuhiro Fuchi, director of the Fifth Generation Computer research institute (ICOT), left behind several statements during the project that held back excessive expectations. In an April 1984 interview in an American magazine, he said he had long been extremely skeptical about AI, that amid a lot of optimism he had been pessimistic and critical, but that he had not thought it impossible. He also answered that it is unrealistic to expect applications of expert systems (systems that imitate experts' knowledge), such as raising the catch landed by fishing boats, from the first 10 years of the plan, that it would take far longer, and that if short-term profit were the goal, a joint national project would not be needed. In his keynote at the November 1984 international conference (FGCS'84), he said that their framework was a (translated), and that this project is the proof of it. He also said they were not directly involved in the currently talked-about (translated), and that AI built on top of today's computers will eventually hit limits, so they were building a new foundation. In his keynote at FGCS'88 in 1988, he said that AI had since become a worldwide fad, but that this project is (translated). As one of the temptations of a national research program, he warned that history has examples in which goals were set too high (or were misunderstood as too high), causing a backlash and bringing on a (translated) of research (written exactly so in the printed text; thought to mean “dark ages” — this site's estimate), and that there are also examples of catering to cheerful or dreamy talk.
The popular story:It is often said that the Fifth Generation failed because the government talked too big about the dream of “computers that talk and think” (for how this is told, see the record → The story that the Fifth Generation Com…) The original source says:The director's own words during the project were as follows. In 1984: he had long been skeptical about AI; applications are not possible in the first 10 years and would take far longer; the proposal is a (translated); and they are not directly involved in the (translated). In 1988: this is “not a project about AI itself” (translated); and there are examples in which goals set too high (or misunderstood as such) caused a backlash and brought on a (translated) of research (thought to mean “dark ages”). Many of his words held back excessive expectations (the summary of his words is this site's). However, it is recorded separately that at the same time a large picture of dreams was spreading outside (in the United States and elsewhere) and around him (records → Feigenbaum and McCorduck's book Fifth G… and → The 1981 forecast: it is said that in t…)
What we still don’t know:This site has not confirmed whether the (translated) written in the printed 1988 lecture is a misprint of “dark ages” (the recording and the English version of the lecture were not opened). It is also not written in the lecture which events the examples of backlash refer to
What the figure shows:Prolog lists facts and rules instead of writing procedures. When asked a question, the machine tries rules in order and, if they do not fit, goes back and searches another path to an answer.
This is the same mechanism as "Try it", written as Python code. Press ▶ Run and the result appears below. Change the numbers and run it again.
u07_mini_prolog.py
05Logic programming · Unification
Making shapes match: unification
What does it mean for a question and a rule to “match”?
What Prolog always does before it uses a rule is unification. It compares the shape of the question with the head of the rule (the left side of ←) from the start, and checks whether putting values into the variables (boxes whose contents are not yet decided) can make them the same shape. parent(X, hanako) and parent(taro, Y) become the same if you put taro in X and hanako in Y. parent(X, hanako) and parent(taro, jiro) do not match.
When you press "Step" in the figure on the right, the place being compared lights up purple. In point(X, X), a value put into X first also applies to the later comparison. If you put into a variable a shape that contains the variable itself, such as X and f(X), the shape goes on forever. The occurs check prevents this.
In the final evaluation (1993), the processors of the Parallel Inference Machine PIM had a dedicated instruction for following the chain of values put into variables (dereferencing).
The logic language “Prolog” is born in Marseille: line up logical sentences and you have a program
In Marseille, Alain Colmerauer and the graduate student Philippe Roussel developed Prolog, a language in which writing logical sentences (if ... then ...) one after another makes them directly into a program. The name is short for PROgrammation en LOGique (programming in logic). It grew out of ideas shared with Robert Kowalski, who was visiting from Edinburgh. Looking back on his second visit in the summer of 1972, Kowalski said that .
1 June 1992Verified in the originalSources disagreeJapan
Fifth Generation results: the Parallel Inference Machine PIM and the language KL1 (the largest machine had 512 computers; the five models together exceeded 1,000)
The Fifth Generation project developed the “Parallel Inference Machine PIM”, which links many computers to carry out inference (drawing conclusions from rules) at the same time, and its dedicated language , and presented them at the international conference FGCS'92 in June 1992. KL1 was designed by Takashi Chikayama, based on the mechanism of the language GHC devised by researcher Kazunori Ueda (from a talk by director Kazuhiro Fuchi). Five models of PIM with different specifications were built; the largest, PIM/p, used 512 computers and could perform 156 million inferences per second. Counting all five models together, there were more than 1,000 computers. The final evaluation states that the target of 100 million inferences per second was exceeded.
The popular story:“The Fifth Generation built an inference machine linking 1,000 computers,” people tend to say The original source says:The largest single machine was the 512-computer PIM/p. Counting all five models together, there were more than 1,000 computers, which is how the final evaluation words it
What we still don’t know:We have not confirmed how PIM was used after the project (how many machines ran where, and for how many years)
4 sources and how they were checked
Where sources disagree:Differences in how the scale is described: the conference page of the ICOT Archive says “a prototype of a parallel inference machine on the scale of 1,000 computers”; the final evaluation says “five models totaling more than 1,000, starting with the 512-computer model p”; Nilsson says “the largest, PIM/p, has 512”
Fifth Generation Computer Project: Final Evaluation Report (Academic and Technical Evaluation Working Group, Committee for the Promotion of Basic Computer Technology Development · 1993 · p.8 Parallel symbolic processing technology (and PDF pages 5, 1… · Original source)
What the figure shows:Unification compares 2 structures from the start and checks whether filling unset boxes, or variables, can make their forms match. Inserting a structure that contains the variable itself would continue forever, so a check prevents this.
This is the same mechanism as "Try it", written as Python code. Press ▶ Run and the result appears below. Change the numbers and run it again.
u07_unify.py
061982–1984 · The world's reaction
“Japan” in the reports
Did the world move to counter Japan's plan?
After the Tokyo conference, large plans began to move in several places. In the United Kingdom, a minister for information technology set up a committee in March 1982, and its report that year recommended a 5-year, £350 million plan of joint research among government, industry and universities. In the United States, companies gathered to found a joint research company, MCC, and DARPA, the Department of Defense research agency, published a plan for its in October 1983. The European Community proposed ESPRIT in 1983.
But how Japan is written about in these documents is completely different. The thickness of the lines in the figure on the right shows how strongly each document refers to Japan. The British report says at the start of its main text that the committee began because of . The European proposal mentions it just once, as an example in a note, and in the DARPA plan neither Japan nor the Fifth Generation appears even once.
Switch to "Japan's view (1995)" and you get how the 1995 MITI Research Review writes about it. It records that all of these plans were inspired by the Fifth Generation.
What you think you knowThe popular story:The world began AI plans in response to Japan The original source says:The British report states plainly that Japan's announcement was the trigger. In the US Department of Defense research agency's plan (1983), however, neither Japan nor the Fifth Generation appears even once (a history book records that it was a large factor in the mood for persuading Congress). In the European proposal (1983), it appears only once as an example in a note.
Europe's circumstances
The ESPRIT proposal gave as its reasons Europe's trade deficit in information technology and fierce competition from the United States and Japan. It also records that it was Europe's big companies who strongly asked for a large plan. In one of its fields, , it called knowledge processing .
MCC was founded at the call of William Norris, founder of Control Data, with 12 companies including DEC providing money and people, and it decided to dissolve in June 2000.
Sources & confidence (4)“Japan” in the reports
1982–1983Verified in the original
The British : the report states that Japan's Fifth Generation announcement was the trigger
A committee set up by the British government's minister for information technology (chaired by John Alvey, a senior executive of British Telecommunications) issued a report recommending a 5-year program of joint research among the government, companies, and universities (a total of £350 million, with the government paying two-thirds). At the start, the report states plainly that the committee came about because of and Japan's invitation to take part. It also says that the meeting, after hearing reports from the British people who went to the Tokyo conference, saw the Japanese plan as a major competitive threat. AI that uses knowledge (IKBS) was one of the four pillars.
1 January 1984–1988Verified in the originalSources disagree
Europe's joint research program begins ── in the proposal document, Japan's Fifth Generation is one example in a footnote
The European Community (the forerunner of the EU) started the first phase (1984–1988) of ESPRIT, a program for carrying out information technology research jointly across national borders. One of its five areas, , put the shift from data processing to knowledge processing as . The Community's contribution was 750 million euros (normally half of the cost). The European Commission's 1983 proposal document gave as reasons Europe's trade deficit in information technology and the fierce competition from the United States and Japan. In a note it wrote, , as an example. It also records that it was Europe's large companies that strongly called for this big program.
The popular story:It is often said that Europe began ESPRIT to counter Japan's Fifth Generation The original source says:The 1983 proposal document mentions Japan's Fifth Generation only once, in a note, as an example of how the . The central reasons were Europe's information technology trade deficit and strong demands from companies. The emphasis differs from Britain's Alvey Report, which stated in its main text that Japan's announcement was the trigger.
What we still don’t know:The text of the decision itself (Council decision of February 1984, Official Journal L 67) could not be opened, because EUR-Lex does not accept our requests.
3 sources and how they were checked
Where sources disagree:Starting year: Nilsson says it was started in 1983, while CORDIS gives the start of the first phase as January 1, 1984 (this can be read as 1983 being the proposal and 1984 the decision and start). Budget: CORDIS gives the Community's contribution as 750 million, and Nilsson gives the first phase's budget as 1.5 billion ECU (split in half); read together, they do not conflict.
American computer companies form the joint research company MCC (a private-sector answer to Japan's Fifth Generation)
American computer companies, alarmed by the news of Japan's Fifth Generation, set up the joint research company MCC (Microelectronics and Computer Technology Corporation) at the call of William Norris, who founded Control Data. Twelve companies, including DEC, supplied money and people, and it studied software, parallel processing, artificial intelligence and knowledge systems, and more. Austin, Texas, was chosen as the site (July 1983). About 400 people worked there around 1985, but it began to lose momentum around 1987, when its first president stepped down, and in June 2000 it was decided to dissolve it. Nilsson writes that by then the Fifth Generation was no longer thought to be such a threat.
The popular story:A US historical record (the Handbook of Texas) says that Japan's Ministry of International Trade and Industry (MITI) announced The original source says:In Japan's own national tables, the money the government paid was about 54.1 billion yen over 11 years (about $380 million by Nilsson's conversion). Abroad, the size of Japan's plan was estimated in many ways (a 1983 European document said $500 million, and Britain's 1982 report gave a projected amount and then noted, as the view of many people, that ).
What we still don’t know:MCC's own documents from the time could not be opened. It is not known which announcement or article of the time the $850 million figure came from.
2 sources and how they were checked
Where sources disagree:The Handbook of Texas describes Japan's Fifth Generation as a plan of $850 million over ten years announced by MITI in April 1982. In Japanese records, the plan was announced at the international conference of October 1981 (another record), and April 1982 is the month ICOT was created (another record). The money paid by the government was about 54.1 billion yen over 11 years (another record; Nilsson writes about $380 million at 1990 exchange rates).
The Quest for Artificial Intelligence (Nils J. Nilsson · 2010 · p.354 22.2.1 The Microelectronics and Computer Technology Corpora… · Secondary summary)
28 October 1983–1993Verified in the originalSources disagree
The US Department of Defense's DARPA announces the ── Japan's name never appears in the plan
DARPA, part of the US Department of Defense, released a large-scale plan to develop . It set three goals: autonomous systems such as vehicles that move by themselves, a pilot's associate (pilot support), and battle management (support for commanding operations). To get there, it would grow AI that has experts' knowledge written into it, research on images, speech, and language, and the computers to run them. The budget was about $600 million for the first 5 years and $50 million for fiscal 1984. The money DARPA put in over 1983–1993 was a little over $1 billion. It is often said to be “a counter to Japan's Fifth Generation,” but the plan itself never mentions Japan or the Fifth Generation.
The popular story:It is said that the United States began the Strategic Computing Initiative to counter Japan's Fifth Generation The original source says:The 1983 plan never mentions Japan or the Fifth Generation (it notes that its list of foreign programs is classified). According to history books, alarm about Japan played a large part in getting the plan approved in Congress. The “counter” was not a stated reason in the plan; it belongs to the mood of the time and to the work of persuading Congress.
What we still don’t know:How Japan was talked about in explanations to Congress (such as the 1983 hearings) has not been confirmed, because the hearing records could not be opened (this relies on Roland's work as cited by Nilsson).
3 sources and how they were checked
Where sources disagree:The Japanese side (MITI Research Review 1995) counts it as one of the plans begun “inspired by the Fifth Generation.” The American plan itself, on the other hand, does not mention Japan and gives military need and “national security and economic strength” as its reasons. According to history books (Nilsson citing Roland), Japan's plan was one of the reasons and was especially effective in persuading Congress.
What the figure shows:Major projects began elsewhere after Japan's project, but their documents mention Japan to very different degrees. Britain described Japan as the trigger, while a US defense proposal mentioned it 0 times.
071983 · Prediction
The book and the prediction clock
Who painted Japan's plan as a big threat?
In 1983, Edward Feigenbaum, a professor of computer science at Stanford University, and Pamela McCorduck, a writer who had also written a history of AI, published a book introducing Japan's plan, . In a magazine article that printed part of the book, the two wrote that if even part of the Fifth Generation succeeded, Japan would pull far ahead of the world in the information industry, and that the United States has no national plan to match it. They also said that many Americans were skeptical, but that the chance of this country's gamble succeeding was higher than people thought. A history book (Nilsson 2010) says the book heightened the sense of crisis in Congress, which was debating research budgets.
The two were writing to appeal to Americans that . When you move "Current year" on the prediction clock on the right, the 1981 booklet, this book and the director's 1992 outlook each reach the year when their answer comes due, and what came after is shown.
On MCC, the group of American companies, Nilsson writes that by the time its first president stepped down in 1987, the Fifth Generation was no longer thought of as such a threat.
What you think you knowThe popular story:Japan itself exaggerated the threat of the Fifth Generation The original source says:Among those who painted Japan's plan as a great threat and spread that picture were American AI researchers themselves. The book was written to make an issue of the fact that the United States had no national plan.
Sources we checked
We have not confirmed the book itself (1983). What we confirmed is the magazine Creative Computing (August 1984), which reprinted part of the book with the publisher's permission. The book estimates the budget the Japanese government would spend over 10 years at about $450 million (section 08 of this chapter).
The director's outlook：In a 1992 talk, Kazuhiro Fuchi said that for the results to truly take root in society, a further maturing period of about 5 years would probably be needed. This site has not checked in sources what happened after that.
Sources & confidence (2)The book and the prediction clock
1983–1984Verified in the originalSources disagreeJapan
Feigenbaum and McCorduck's book (1983): if even part of the Fifth Generation succeeds, they wrote, Japan will take a commanding lead in the world's information industry
Edward Feigenbaum, an AI researcher at Stanford University, and the writer Pamela McCorduck wrote a book introducing Japan's Fifth Generation plan. In a magazine article that printed part of the book, the two wrote that Japan is aiming, by the beginning of the next decade, for a standout position in world trade using humankind's knowledge as its weapon; that if even part of the Fifth Generation succeeds, Japan will run far ahead in the world's information business; and that the United States has no plan to counter it. They also wrote that many Americans are doubtful but that the chances of Japan's challenge succeeding are higher than expected. A history book (Nilsson) says this book heightened the sense of alarm in the US Congress (which handles DARPA's budget).
The popular story:It is often thought that the Fifth Generation “threat” was loudly hyped by Japan itself The original source says:Among those who drew Japan's plan large and spread it were AI researchers themselves, writing for Americans. This book was written to make an issue of the fact that the United States had no national plan, and to argue to Americans that (the article says so itself).
What happened next:The Fifth Generation finished its main stage in fiscal 1992. Nilsson writes that no notable uses were found for the software it built, and that workstations and personal computers overtook the parallel inference machines in speed. At the US MCC, by around 1987 the Fifth Generation was no longer thought to be much of a threat (Nilsson). A 1995 report from the Ministry of International Trade and Industry's research institute also describes the Fifth Generation as a project (translated) (while also writing that calling it a failure on that basis is an opinion that does not fully understand its ideals), and said that its architecture is not thought likely to be used in future technology. In the materials I looked at, there is no record of Fifth Generation machines or software taking the lead in the world's information industry.
What we still don’t know:The book itself (1983) has not been opened. What is said in parts not in the excerpt (the preface and the main text, pp.19-20) was confirmed only through Nilsson's citation.
2 sources and how they were checked
Where sources disagree:Size of Japan's plan: this book puts the government's 10-year budget at “about $450 million” (and expects companies to put in the same amount or twice as much). In Japan's national table it is about 54.1 billion yen over 11 years (another record; about $380 million at 1990 exchange rates by Nilsson's conversion).
American computer companies form the joint research company MCC (a private-sector answer to Japan's Fifth Generation)
American computer companies, alarmed by the news of Japan's Fifth Generation, set up the joint research company MCC (Microelectronics and Computer Technology Corporation) at the call of William Norris, who founded Control Data. Twelve companies, including DEC, supplied money and people, and it studied software, parallel processing, artificial intelligence and knowledge systems, and more. Austin, Texas, was chosen as the site (July 1983). About 400 people worked there around 1985, but it began to lose momentum around 1987, when its first president stepped down, and in June 2000 it was decided to dissolve it. Nilsson writes that by then the Fifth Generation was no longer thought to be such a threat.
The popular story:A US historical record (the Handbook of Texas) says that Japan's Ministry of International Trade and Industry (MITI) announced The original source says:In Japan's own national tables, the money the government paid was about 54.1 billion yen over 11 years (about $380 million by Nilsson's conversion). Abroad, the size of Japan's plan was estimated in many ways (a 1983 European document said $500 million, and Britain's 1982 report gave a projected amount and then noted, as the view of many people, that ).
What we still don’t know:MCC's own documents from the time could not be opened. It is not known which announcement or article of the time the $850 million figure came from.
2 sources and how they were checked
Where sources disagree:The Handbook of Texas describes Japan's Fifth Generation as a plan of $850 million over ten years announced by MITI in April 1982. In Japanese records, the plan was announced at the international conference of October 1981 (another record), and April 1982 is the month ICOT was created (another record). The money paid by the government was about 54.1 billion yen over 11 years (another record; Nilsson writes about $380 million at 1990 exchange rates).
The Quest for Artificial Intelligence (Nils J. Nilsson · 2010 · p.354 22.2.1 The Microelectronics and Computer Technology Corpora… · Secondary summary)
What the figure shows:US AI researchers themselves were among those who presented Japan's project as a major threat. Predictions have target years, and their results become visible when those years arrive.
081982–1995 · Money
54.1 billion yen, or 56.9 billion?
How much did the state spend on the Fifth Generation Computer?
According to a table in the 1995 MITI Research Review (sourced to Ministry of International Trade and Industry materials), the budget the state spent was about 54.1 billion yen over the 11 fiscal years 1982–1992 (8.3 billion in the first phase, 21.6 billion in the middle phase and 24.2 billion in the final phase). After that, about 2.8 billion yen was spent over the 2 fiscal years 1993–1994 on a follow-on plan, which makes about 56.9 billion yen in total. From fiscal 1986, money also came from a separate special account, in addition to the general account (the ordinary budget).
Try moving "How far to add" in the bar graph on the right. Stop at fiscal 1992 and you get 54.1 billion yen; add up to fiscal 1994 and you get 56.9 billion yen. The difference between the often-seen “about 54 billion yen” (1993 final evaluation) and “57 billion yen” (from remarks in the Diet in 2017 and elsewhere) is the difference in the period that was totaled. You can also display the foreign estimates by pressing the button.
Total = Σ (general account + special account)
Σ
add up fiscal year by fiscal year
Fiscal year
April to the following March
What you think you knowThe popular story:Overseas, Japan was seen as pouring huge sums into the Fifth Generation and hiding the real amount The original source says:The 1982 British report noted a view that , but the 1993 final evaluation reported (translated), and the 1995 report carries a table broken down by fiscal year. The foreign estimates vary: $450 million, $500 million and $850 million, and the largest figure came from an encyclopedia written after the project ended.
What we don't know
The table lists only state spending; the costs and staff time borne by the participating companies are not included.
From a slide by Shun'ichi Uchida, Nilsson (2010) gives the 10-year budget as 54.2 billion yen (about $380 million at the 1990 exchange rate). This site has not converted between dollars and yen to compare them.
Sources & confidence (2)54.1 billion yen, or 56.9 billion?
April 1982–1995Verified in the originalSources disagreeJapan
Government spending on the Fifth Generation: about 54.1 billion yen over the main 11 years, about 56.9 billion yen with the later 2 years added
According to a table compiled in 1995 by the Ministry of International Trade and Industry's research institute (the MITI Research Institute), government spending on the Fifth Generation Computer project was about 54.1 billion yen over the 11 fiscal years 1982–1992 (initial stage 8.3 billion yen, middle stage 21.6 billion yen, final stage 24.2 billion yen). After that, about 2.8 billion yen was spent over the two fiscal years 1993–1994 on a finishing plan (a plan to make what had been built usable on ordinary computers), bringing the total to about 56.9 billion yen. The commonly seen figures of “about 54 billion yen” and “57 billion yen” differ only in whether these last two years are counted.
The popular story:People give different numbers: “the Fifth Generation cost 57 billion yen,” or “no, it was 54 billion yen” The original source says:In the government's table, the main 11 years come to about 54.1 billion yen, and adding about 2.8 billion yen for the later 2 years makes about 56.9 billion yen. Both numbers are in the table; the difference is how far the counting goes
What we still don’t know:The table covers only the money paid by the government (MITI). Funds and staff that participating companies bore on their own are not in this table (as far as can be read from the table's notes)
4 sources and how they were checked
Where sources disagree:“About 54 billion yen” comes from the final evaluation report (1993) and the main text of the MITI Research Review (1995), “total 54.1 billion yen”; both cover the 11 fiscal years 1982–1992. ／ “57 billion yen” comes from a 2017 Diet statement and the like; it matches about 56.9 billion yen, the total with the later two fiscal years 1993–1994 (about 2.8 billion yen) added. ／Nilsson 2010 gives “54.2 billion yen” from Uchida's slides. All of them are explained by the same table, and the disagreement is only in the range of years counted
1982–2020Verified in the originalSources disagreeJapan
Foreign estimates of the Fifth Generation's cost vary widely: $450 million, $500 million, $850 million... In Japan's own national table it is about 54.1 billion yen over 11 years
The cost of the Fifth Generation was estimated in many ways abroad. Britain's Alvey Report (1982) wrote that the Japanese government had committed £24 million for 1982–84 and was expected to spend £60 million to £100 million in 1985–88. It also wrote that many people think the real amount may be much higher and that the true figure will not be disclosed. A footnote in the European ESPRIT proposal (1983) says ; the book by Feigenbaum and McCorduck (1983) says the government is expected to put in (the government's part over 10 years, with companies expected to match it, and perhaps double); and the Texas encyclopedia (published online in 2007) writes . In a table in a 1995 report from the Ministry of International Trade and Industry's research institute (source: Ministry documents), the government paid about 54.1 billion yen over the 11 fiscal years 1982–1992.
The popular story:Abroad, Japan was seen as having poured huge sums into the Fifth Generation (and hiding the true amount) The original source says:The 1982 British report recorded the view of many people that , but the 1993 final evaluation report says (translated), and the 1995 report from the Ministry's research institute carries a year-by-year table. The estimates differ from writer to writer, and the largest, $850 million, comes from an encyclopedia article written after the project.
What we still don’t know:None of the sources says which announcement or article of the time the foreign estimates were based on. The total cost paid by the participating companies themselves is not known.
6 sources and how they were checked
Where sources disagree:The estimates disagree: $500 million (EC 1983), $450 million plus the companies' share (Feigenbaum and McCorduck 1983), $850 million (Handbook of Texas). The British report gives projected amounts by period in pounds, and the Japanese table is in yen (about 54.1 billion yen, 11 years). Because the currencies, exchange rates, and scope of the totals (government only, or including private money) differ, this site does not make a converted comparison.
What the figure shows:The difference between about 54 billion yen and 57 billion yen comes from which fiscal years were included, rather than from lying or hiding money.
091982–1991 · Japan's boom
The AI boom that was imported back
Where did Japan's second AI boom come from?
The (supervised by a division of the Ministry of International Trade and Industry) says that (translated). It adds that Japan's boom began with the Fifth Generation, and that (translated). It also says there seems to have been an idea like a myth, (translated), with company heads ordering AI development from the top down. In the Diet in 1985, too, AI was described as being in (translated).
So since when had the term itself been used? When we line up digitized books and magazines in order of age in the National Diet Library's full-text search, the oldest example in the sense of AI was a magazine from October 1982. The magazine says (translated). If you search for in the figure on the right, results from 1978 and 1981 come up first. They are characters read from the author name A.I. Bloomfield on a book and from the MANZAI boom in comedy. Even when the character strings match, they are not necessarily the same term.
What you think you knowThe popular story:Japan's second AI boom was Japan-born excitement over the Fifth Generation The original source says:According to the white paper supervised by a Ministry of International Trade and Industry division in 1991, Japan's boom “was generated in a form where the response from overseas was imported back” (translated) from the Fifth Generation.
Cool voices in the middle of the boom
A 1985 issue of the magazine says (translated). A supplement to the same year also pointed out that the AI boom was repeating a history of rising many times and ending in disappointment.
The extracted passages are machine-read text, and we have not checked page images. The confidence grade is “weakly confirmed”.
Sources & confidence (3)The AI boom that was imported back
1985–1991Verified in the originalJapan
Japan's AI boom was a “reverse import” (translated) — a 1991 white paper supervised by the government said the overseas reaction to the Fifth Generation project bounced back to Japan and became a boom
The , supervised by a division of the Ministry of International Trade and Industry (MITI), wrote that an AI boom arose around the world, including the United States and Europe, starting from Japan's Fifth Generation project. In Japan, from around 1985 the term was embraced by the media, and new AI magazines and reports kept being launched. The white paper says that Japan's boom arose as a “reverse import” (translated) of the Fifth Generation and the reaction to it overseas. It adds that company heads, holding a myth-like belief that (translated), seemed to order development from the top down. At first, money was poured without hesitation into expensive AI-dedicated machines (such as those of Symbolics and Lisp Machines, Inc.) and tools.
The popular story:Many people think Japan's second AI boom was Japan's own, with Japan getting excited about its own Fifth Generation project The original source says:The 1991 white paper supervised by the government says that Japan's boom “arose as the overseas reaction was reverse-imported” (translated) from the Fifth Generation project (big talk overseas moved Japan's company heads)
What we still don’t know:This white paper does not check when the media first used the term “AI boom,” or in what year each magazine was launched (for example, ).
2 sources and how they were checked
AI White Paper 1991: AI Technology and Its Use (Supervised by the Electronics Policy Division, Machinery and Information Industries Bureau, Ministry of International Trade and Industry (MITI); edited by the ICOT-JIPDEC AI Center; published by the Japan Information Processing Development Association (JIPDEC) · 1991 · p.11 Part 2, The Present State and Outlook of AI, Chapter 1, sec… · Original source)
October 1982–1985Weakly confirmedSources disagreeJapan
Magazines in October 1982 already use “artificial intelligence boom” and ── (translated). By 1985 there were cooler views too
In a full-text search of books and magazines digitized by the National Diet Library, I listed the unnumbered “artificial intelligence boom” from the oldest publication date, and the oldest were two magazines of October 1982 (within what was searched on September 23, 2026). (vol. 20, no. 10) has a sentence saying that the race to develop computers that understand natural language and think is intensifying, followed by “The artificial intelligence boom of a while back has come around again” (translated), and it also has the words (translated). (vol. 19, no. 10) has a sentence saying it will reveal the true nature of the recent artificial intelligence boom. The same issue also has a sentence saying that unless a way is made to overcome a fatal flaw, the on both sides of the Pacific may end as just another boom again (the oldest example of the form being used to mean AI). Of the 20 oldest hits, 15 are from 1985, and the writers of that year named different things as the trigger of the boom (the Fifth Generation plan, the success of the expert system MYCIN, Feigenbaum's proposal of knowledge engineering [1977], and “AI tools”, meaning software tools for building AI). In the same year, 1985, there were also words of a cooler view from the middle of the boom: (translated) (an article in question-and-answer form in the March issue of ), today's AI boom has hardly any substance behind it (July issue of ), and AI booms have repeated a history of disappointment, rising many times and fading (the special issue of , , November 1985).
The popular story:A 1991 government-supervised white paper says (translated) (record → Japan's AI boom was a reverse import (t…). This can be read as meaning that people in Japan began to speak of an around 1985 The original source says:In the full-text search, the term itself already appears in two magazines of October 1982. And it was written as the return of a boom of the past: “The artificial intelligence boom of a while back has come around again” (translated). However, 15 of the 20 oldest hits are from 1985, which agrees with the white paper's statement that it was widely embraced from around 1985 (the amount of digitized material differs by year, so be careful about simple comparisons of counts).
What we still don’t know:I have not seen the articles' titles, authors, or pages (the page images are limited to use inside the library and so were not opened). The excerpts come from OCR (optical character recognition) and are cut off at both ends. Materials that were not digitized or were misread cannot be searched, so it is not known whether there are examples older than October 1982. It is also unclear from the excerpt when in of October 1982 refers to (a month later the same magazine writes of (translated) (translated), record → A 1982 magazine already has the first a…).
11 sources and how they were checked
Where sources disagree:Trigger of the boom: “Jimu to Keiei” (Office and Management), April 1985, points to the Fifth Generation plan (“is thought to have originated from”). The “Preprints of Academic Lectures” (Society of Automotive Engineers of Japan, October 1985) points to the success of MYCIN (“one of the triggers”). “Automation,” May 1985, points to Feigenbaum's proposal of knowledge engineering (1977) (the excerpt is cut off at “and, that,” so other triggers may be listed as well). (November 1985) calls “AI tools” “one trigger.” The 1991 government-supervised white paper says the overseas response to the Fifth Generation was (translated) into Japan (another record). Momentum of the boom: “Jimu to Keiei,” April 1985, says it “has finally begun to show calm movement,” while “Eizo Joho” (Image Information), July 1985, writes of “a bit of an artificial intelligence boom lately.”
President 20(10) (1982 · Approximate frame numbers in the National Diet Library Digi… · Original source · Could not be opened)
4 April 1985Verified in the originalSources disagreeJapan
In the 1985 Diet, AI is explained as being said to be (translated)
Takemochi Ishii of the University of Tokyo was called to the Diet (a committee of the House of Representatives) and explained, (translated). He said the first was an effort centered on the United States from the mid to late 1960s, and the second was Japan's Fifth Generation Computer project. The idea of counting booms was already around in the middle of the second boom.
What we still don’t know:We do not know who started the phrase (translated) before 1985 (this is the oldest instance found within the Diet records).
2 sources and how they were checked
Where sources disagree:Ishii gives the period of the first boom as “the mid to late 1960s.” Many Japanese sources of the 2010s say “the late 1950s to the 1960s” (such as the 2016 White Paper of the Ministry of Internal Affairs and Communications; see another row)
What the figure shows:Japan's second AI boom arose when overseas reactions to the Fifth Generation project were “reverse-imported.” In word searches, matching text does not guarantee matching meaning.
101984–1988 · The director's words
invite a backlash
In the middle of the project, what was the director himself saying?
The director, Kazuhiro Fuchi, left behind words more than once that held back talk of big dreams. In an April 1984 interview with an American magazine he said he had long been skeptical of AI, and answered that it was not realistic to expect applications in the first 10 years. In a keynote speech in November the same year, too, he said that their framework was a (translated) and that this project was there to test it, and that it was not directly involved in the (translated).
In a 1988 talk he said this project was (translated). Naming one of the temptations national research plans easily fall into, (translated), he warned that history has examples where a goal that was too high (or was misunderstood as too high) invited a backlash.
On the right is a figure of this way of thinking (all numbers are provisional values of this site). Even if you keep the goal modest, if it is conveyed as something big outside, expectations fall sharply in the year of the promise.
Expectation in the year of the promise = ability − (the amount by which goal × multiplier exceeds ability). Expectation cannot fall below 0.
ability
results achieved by the deadline
goal × multiplier
the promise the world received
What you think you knowThe popular story:The Fifth Generation failed because the state talked up the dream of a computer that could speak and think too grandly The original source says:Many of the director's own words in the middle of the project held back excessive expectations. But another source records that an exaggerated image was spreading around him at about the same time (sections 02 and 07 of this chapter).
The
The record of the 1988 talk says the backlash brought on a (translated) of research. The Japanese word was probably meant as “dark age” (this site's inference), but it has not been confirmed whether it is a misprint. The talk also does not say which events the examples that invited a backlash refer to.
Sources & confidence (1) invite a backlash
April 1984–1988Verified in the originalJapan
The words of the Fifth Generation's director, Kazuhiro Fuchi, himself (1984–1988) ── long skeptical about AI, the proposal is a (translated), no direct involvement in the AI business, and (translated) bring a backlash
Kazuhiro Fuchi, director of the Fifth Generation Computer research institute (ICOT), left behind several statements during the project that held back excessive expectations. In an April 1984 interview in an American magazine, he said he had long been extremely skeptical about AI, that amid a lot of optimism he had been pessimistic and critical, but that he had not thought it impossible. He also answered that it is unrealistic to expect applications of expert systems (systems that imitate experts' knowledge), such as raising the catch landed by fishing boats, from the first 10 years of the plan, that it would take far longer, and that if short-term profit were the goal, a joint national project would not be needed. In his keynote at the November 1984 international conference (FGCS'84), he said that their framework was a (translated), and that this project is the proof of it. He also said they were not directly involved in the currently talked-about (translated), and that AI built on top of today's computers will eventually hit limits, so they were building a new foundation. In his keynote at FGCS'88 in 1988, he said that AI had since become a worldwide fad, but that this project is (translated). As one of the temptations of a national research program, he warned that history has examples in which goals were set too high (or were misunderstood as too high), causing a backlash and bringing on a (translated) of research (written exactly so in the printed text; thought to mean “dark ages” — this site's estimate), and that there are also examples of catering to cheerful or dreamy talk.
The popular story:It is often said that the Fifth Generation failed because the government talked too big about the dream of “computers that talk and think” (for how this is told, see the record → The story that the Fifth Generation Com…) The original source says:The director's own words during the project were as follows. In 1984: he had long been skeptical about AI; applications are not possible in the first 10 years and would take far longer; the proposal is a (translated); and they are not directly involved in the (translated). In 1988: this is “not a project about AI itself” (translated); and there are examples in which goals set too high (or misunderstood as such) caused a backlash and brought on a (translated) of research (thought to mean “dark ages”). Many of his words held back excessive expectations (the summary of his words is this site's). However, it is recorded separately that at the same time a large picture of dreams was spreading outside (in the United States and elsewhere) and around him (records → Feigenbaum and McCorduck's book Fifth G… and → The 1981 forecast: it is said that in t…)
What we still don’t know:This site has not confirmed whether the (translated) written in the printed 1988 lecture is a misprint of “dark ages” (the recording and the English version of the lecture were not opened). It is also not written in the lecture which events the examples of backlash refer to
What the figure shows:During the project, the director repeatedly cautioned against speaking too grandly about its dreams. Even modest goals can produce a sharp drop in expectations at the promised deadline if outsiders present them in a much bigger way.
111985–1992 · Parallel processing
Splitting work, and the cost of linking
If you use 512 computers, does it get 512 times faster?
The heart of the Fifth Generation was parallel processing, running inference on many computers at once. The tree of case splits you search for an answer is cut into work for each branch and handed out to each machine. But the branches vary in size, and every handoff of work also costs effort to connect.
Try increasing the number of machines in the figure on the right. If you just deal out the work in order at the start, some machines finish early and wait. If idle machines go and take the remainder, you get close to a speedup equal to the number of machines. Even so, you cannot finish faster than the single largest piece of work (the tree is an example made by this site).
The final evaluation (1993) writes that in a logic simulation the speedup was 166 times on 256 machines, and for some problems nearly 500 times on 512 machines, and it packaged load balancing (how work is distributed) as reusable components. Nilsson (2010) writes that outside special problems suited to it, it was hard to make full use of the strengths of parallelism.
S(P) = T(1) ÷ T(P)
P
number of machines
T(P)
time for everything to finish on P machines (including the cost of handing out work)
S(P)
speedup factor
Sources & confidence (2)Splitting work, and the cost of linking
1 June 1992Verified in the originalSources disagreeJapan
Fifth Generation results: the Parallel Inference Machine PIM and the language KL1 (the largest machine had 512 computers; the five models together exceeded 1,000)
The Fifth Generation project developed the “Parallel Inference Machine PIM”, which links many computers to carry out inference (drawing conclusions from rules) at the same time, and its dedicated language , and presented them at the international conference FGCS'92 in June 1992. KL1 was designed by Takashi Chikayama, based on the mechanism of the language GHC devised by researcher Kazunori Ueda (from a talk by director Kazuhiro Fuchi). Five models of PIM with different specifications were built; the largest, PIM/p, used 512 computers and could perform 156 million inferences per second. Counting all five models together, there were more than 1,000 computers. The final evaluation states that the target of 100 million inferences per second was exceeded.
The popular story:“The Fifth Generation built an inference machine linking 1,000 computers,” people tend to say The original source says:The largest single machine was the 512-computer PIM/p. Counting all five models together, there were more than 1,000 computers, which is how the final evaluation words it
What we still don’t know:We have not confirmed how PIM was used after the project (how many machines ran where, and for how many years)
4 sources and how they were checked
Where sources disagree:Differences in how the scale is described: the conference page of the ICOT Archive says “a prototype of a parallel inference machine on the scale of 1,000 computers”; the final evaluation says “five models totaling more than 1,000, starting with the 512-computer model p”; Nilsson says “the largest, PIM/p, has 512”
Fifth Generation Computer Project: Final Evaluation Report (Academic and Technical Evaluation Working Group, Committee for the Promotion of Basic Computer Technology Development · 1993 · p.8 Parallel symbolic processing technology (and PDF pages 5, 1… · Original source)
The Fifth Generation final evaluation report: inference speed exceeded the target, but a system for talking with people was not built
The committee that had followed the project's progress (the Committee for the Promotion of Basic Computer Technology Development) compiled a final evaluation report at its end. It reported that PIM, a machine that runs many computers in parallel to carry out inference (drawing conclusions from rules), exceeded the target of 100 million inferences per second, reaching more than 150 million. On the other hand, for systems that deal with people through speech and images, it was found that a huge amount of background knowledge is needed. So partway through, the work was narrowed to the basics of natural language processing, and the report records that it did not go as far as building a concrete system. Overall, it sums up that the project (translated).
The popular story:“The Fifth Generation ended without reaching its goals,” people tend to say, lumping everything together The original source says:The government's final evaluation writes goal by goal. The speed of inference (100 million per second) was exceeded. It says itself that a system for talking with people through speech and images was not built.
What we still don’t know:This evaluation was written by the side that ran the project (the government committee). For outside evaluations, see other records (the MITI Research Review 1995, Nilsson 2010, and others).
2 sources and how they were checked
Fifth Generation Computer Project: Final Evaluation Report (Academic and Technical Evaluation Working Group, Committee for the Promotion of Basic Computer Technology Development · 1993 · p.10 Degree of achievement of the original goals for parallel sy… · Original source)
What the figure shows:Using 512 computers does not make a job 512 times faster. Tasks have unequal sizes and take effort to pass around, and the whole job cannot finish sooner than its largest task.
Number of machines and speedup factor (both axes in steps of 2 times)
121985–1992 · GHC and KL1
Waiting until the data arrives
When several computers use the same data, how is the order kept?
Running Prolog on several computers at once was hard (Nilsson 2010). According to Kazuhiro Fuchi's 1992 talk, at the start of the middle phase the researcher Kazunori Ueda proposed GHC (Guarded Horn Clauses). Building on it, the section chief Takashi Chikayama designed KL1, and all research in the final phase was unified on KL1.
GHC rules have a guard. When a rule needs to look at a variable that has no value yet, the task sleeps there (stops and waits) and wakes when a value goes in. Once a rule that passed its guard has been chosen, it does not go back and redo. In the figure on the right, machine 1 sends numbers and machine 2 adds them up. Machine 2 sleeps when the next number does not arrive. If you press "Remove the producer", everything stops asleep (deadlock).
According to the final evaluation (1993), the first version of PIMOS, an operating system written in KL1, was about 44,000 lines, and its debugger could find deadlocks automatically (the figure is a small example of the idea).
head :- guard | body
guard
the condition for whether this rule may be chosen (wait if a value is missing)
body
the work after it is chosen
Sources & confidence (1)Waiting until the data arrives
1 June 1992Verified in the originalSources disagreeJapan
Fifth Generation results: the Parallel Inference Machine PIM and the language KL1 (the largest machine had 512 computers; the five models together exceeded 1,000)
The Fifth Generation project developed the “Parallel Inference Machine PIM”, which links many computers to carry out inference (drawing conclusions from rules) at the same time, and its dedicated language , and presented them at the international conference FGCS'92 in June 1992. KL1 was designed by Takashi Chikayama, based on the mechanism of the language GHC devised by researcher Kazunori Ueda (from a talk by director Kazuhiro Fuchi). Five models of PIM with different specifications were built; the largest, PIM/p, used 512 computers and could perform 156 million inferences per second. Counting all five models together, there were more than 1,000 computers. The final evaluation states that the target of 100 million inferences per second was exceeded.
The popular story:“The Fifth Generation built an inference machine linking 1,000 computers,” people tend to say The original source says:The largest single machine was the 512-computer PIM/p. Counting all five models together, there were more than 1,000 computers, which is how the final evaluation words it
What we still don’t know:We have not confirmed how PIM was used after the project (how many machines ran where, and for how many years)
4 sources and how they were checked
Where sources disagree:Differences in how the scale is described: the conference page of the ICOT Archive says “a prototype of a parallel inference machine on the scale of 1,000 computers”; the final evaluation says “five models totaling more than 1,000, starting with the 512-computer model p”; Nilsson says “the largest, PIM/p, has 512”
Fifth Generation Computer Project: Final Evaluation Report (Academic and Technical Evaluation Working Group, Committee for the Promotion of Basic Computer Technology Development · 1993 · p.8 Parallel symbolic processing technology (and PDF pages 5, 1… · Original source)
What the figure shows:When several machines share data, a task sleeps while its value is missing and wakes when the value arrives. If no task produces it, all remain asleep and progress stops.
This is the same mechanism as "Try it", written as Python code. Press ▶ Run and the result appears below. Change the numbers and run it again.
u07_ghc_stream.py
131982–1992 · PIM
156 million inferences per second
How is the speed of inference measured, and how fast did it get?
The Fifth Generation measured inference speed in LIPS (logical inferences per second). According to Nilsson (2010), computers in the early 1980s managed about 100,000 per second, and Japan tried to raise that by more than 1,000 times. The target was 100 million per second.
In the first phase it developed ESP, an extended Prolog, and the PSI machine that infers on its own, and used about 500 machines, improved versions included, in research. At the end of the middle phase it built the 64-machine Multi-PSI, and in the final phase the Parallel Inference Machine PIM. According to the final evaluation (1993), the 512-machine PIM/p reached a peak of 156 million per second and the 256-machine PIM/m 153 million, about 100 times a general-purpose mainframe.
The figure on the right has this page's JavaScript solve a small problem, reversing a list (30 items, 496 inferences), and measures the speed. A single processing element of PIM handled 300,000–600,000 per second. With the slider on the right, you can compare an example calculation adding up the speeds of the elements with PIM/p's actual performance.
LIPS = number of inferences ÷ seconds
one inference
using one rule once (this site's way of counting)
What you think you knowThe popular story:The Fifth Generation built an inference machine connecting 1,000 computers The original source says:The largest single machine is the 512-machine PIM/p. Counting all 5 types (512, 256, 256, 16 and 16 machines) together comes to more than 1,000, which is how the final evaluation puts it.
Sources & confidence (2)156 million inferences per second
1 June 1992Verified in the originalSources disagreeJapan
Fifth Generation results: the Parallel Inference Machine PIM and the language KL1 (the largest machine had 512 computers; the five models together exceeded 1,000)
The Fifth Generation project developed the “Parallel Inference Machine PIM”, which links many computers to carry out inference (drawing conclusions from rules) at the same time, and its dedicated language , and presented them at the international conference FGCS'92 in June 1992. KL1 was designed by Takashi Chikayama, based on the mechanism of the language GHC devised by researcher Kazunori Ueda (from a talk by director Kazuhiro Fuchi). Five models of PIM with different specifications were built; the largest, PIM/p, used 512 computers and could perform 156 million inferences per second. Counting all five models together, there were more than 1,000 computers. The final evaluation states that the target of 100 million inferences per second was exceeded.
The popular story:“The Fifth Generation built an inference machine linking 1,000 computers,” people tend to say The original source says:The largest single machine was the 512-computer PIM/p. Counting all five models together, there were more than 1,000 computers, which is how the final evaluation words it
What we still don’t know:We have not confirmed how PIM was used after the project (how many machines ran where, and for how many years)
4 sources and how they were checked
Where sources disagree:Differences in how the scale is described: the conference page of the ICOT Archive says “a prototype of a parallel inference machine on the scale of 1,000 computers”; the final evaluation says “five models totaling more than 1,000, starting with the 512-computer model p”; Nilsson says “the largest, PIM/p, has 512”
Fifth Generation Computer Project: Final Evaluation Report (Academic and Technical Evaluation Working Group, Committee for the Promotion of Basic Computer Technology Development · 1993 · p.8 Parallel symbolic processing technology (and PDF pages 5, 1… · Original source)
The Fifth Generation final evaluation report: inference speed exceeded the target, but a system for talking with people was not built
The committee that had followed the project's progress (the Committee for the Promotion of Basic Computer Technology Development) compiled a final evaluation report at its end. It reported that PIM, a machine that runs many computers in parallel to carry out inference (drawing conclusions from rules), exceeded the target of 100 million inferences per second, reaching more than 150 million. On the other hand, for systems that deal with people through speech and images, it was found that a huge amount of background knowledge is needed. So partway through, the work was narrowed to the basics of natural language processing, and the report records that it did not go as far as building a concrete system. Overall, it sums up that the project (translated).
The popular story:“The Fifth Generation ended without reaching its goals,” people tend to say, lumping everything together The original source says:The government's final evaluation writes goal by goal. The speed of inference (100 million per second) was exceeded. It says itself that a system for talking with people through speech and images was not built.
What we still don’t know:This evaluation was written by the side that ran the project (the government committee). For outside evaluations, see other records (the MITI Research Review 1995, Nilsson 2010, and others).
2 sources and how they were checked
Fifth Generation Computer Project: Final Evaluation Report (Academic and Technical Evaluation Working Group, Committee for the Promotion of Basic Computer Technology Development · 1993 · p.10 Degree of achievement of the original goals for parallel sy… · Original source)
What the figure shows:Reasoning speed is measured in rule applications per second, called LIPS. PIM reached about 156 million per second, exceeding the goal of 100 million.
Number of machines in the 5 types of PIM
This is the same mechanism as "Try it", written as Python code. Press ▶ Run and the result appears below. Change the numbers and run it again.
u07_lips_meter.py
141992 · Theorem proving
A machine proved part of an open problem
How does a computer prove that “there is no such thing”?
According to the final evaluation (1993), the theorem-proving program MGTP on PIM automatically proved part of an unsolved problem about quasigroups posed by the Canadian mathematician Bennett. It reached a speedup of more than 200 times on 256 machines, and added tricks to hold down the explosion of combinations from case splits. In 1992 Kazuhiro Fuchi said this prover was also used as the inference engine of a system that does legal reasoning.
The multiplication table of a quasigroup is a table in which every symbol appears exactly once in each row and each column (a Latin square). In the figure on the right, the cells are filled one at a time to search for a table, and when it hits a dead end it goes back. If every case split hits a dead end, that is a proof that no such table exists. Tables that satisfy x*x = x and x*y = y*x are found for sizes 3 and 5, but do not exist for 4 and 6 (the conditions are an example chosen by this site).
x*x = x, x*y = y*x
x*y
the symbol in row x, column y
lookahead
spotting early any cell that has no symbol left to go in it, and cutting off the search
Sources & confidence (2)A machine proved part of an open problem
30 March 1993Verified in the originalJapan
The Fifth Generation final evaluation report: inference speed exceeded the target, but a system for talking with people was not built
The committee that had followed the project's progress (the Committee for the Promotion of Basic Computer Technology Development) compiled a final evaluation report at its end. It reported that PIM, a machine that runs many computers in parallel to carry out inference (drawing conclusions from rules), exceeded the target of 100 million inferences per second, reaching more than 150 million. On the other hand, for systems that deal with people through speech and images, it was found that a huge amount of background knowledge is needed. So partway through, the work was narrowed to the basics of natural language processing, and the report records that it did not go as far as building a concrete system. Overall, it sums up that the project (translated).
The popular story:“The Fifth Generation ended without reaching its goals,” people tend to say, lumping everything together The original source says:The government's final evaluation writes goal by goal. The speed of inference (100 million per second) was exceeded. It says itself that a system for talking with people through speech and images was not built.
What we still don’t know:This evaluation was written by the side that ran the project (the government committee). For outside evaluations, see other records (the MITI Research Review 1995, Nilsson 2010, and others).
2 sources and how they were checked
Fifth Generation Computer Project: Final Evaluation Report (Academic and Technical Evaluation Working Group, Committee for the Promotion of Basic Computer Technology Development · 1993 · p.10 Degree of achievement of the original goals for parallel sy… · Original source)
1 June 1992Verified in the originalSources disagreeJapan
Fifth Generation results: the Parallel Inference Machine PIM and the language KL1 (the largest machine had 512 computers; the five models together exceeded 1,000)
The Fifth Generation project developed the “Parallel Inference Machine PIM”, which links many computers to carry out inference (drawing conclusions from rules) at the same time, and its dedicated language , and presented them at the international conference FGCS'92 in June 1992. KL1 was designed by Takashi Chikayama, based on the mechanism of the language GHC devised by researcher Kazunori Ueda (from a talk by director Kazuhiro Fuchi). Five models of PIM with different specifications were built; the largest, PIM/p, used 512 computers and could perform 156 million inferences per second. Counting all five models together, there were more than 1,000 computers. The final evaluation states that the target of 100 million inferences per second was exceeded.
The popular story:“The Fifth Generation built an inference machine linking 1,000 computers,” people tend to say The original source says:The largest single machine was the 512-computer PIM/p. Counting all five models together, there were more than 1,000 computers, which is how the final evaluation words it
What we still don’t know:We have not confirmed how PIM was used after the project (how many machines ran where, and for how many years)
4 sources and how they were checked
Where sources disagree:Differences in how the scale is described: the conference page of the ICOT Archive says “a prototype of a parallel inference machine on the scale of 1,000 computers”; the final evaluation says “five models totaling more than 1,000, starting with the 512-computer model p”; Nilsson says “the largest, PIM/p, has 512”
Fifth Generation Computer Project: Final Evaluation Report (Academic and Technical Evaluation Working Group, Committee for the Promotion of Basic Computer Technology Development · 1993 · p.8 Parallel symbolic processing technology (and PDF pages 5, 1… · Original source)
What the figure shows:Searching every possibility by filling squares one at a time and going back when stuck proves that no such table exists if every possible case ends in a dead end.
This is the same mechanism as "Try it", written as Python code. Press ▶ Run and the result appears below. Change the numbers and run it again.
u07_latin_square.py
151986–1995 · Electronic dictionary
Building a dictionary of meaning by hand
What is needed for a computer to handle the meaning of words?
In April 1986, a 9-year plan began at the “Japan Electronic Dictionary Research Institute (EDR)” to build a large “electronic dictionary” (translated) as a foundation for letting computers handle everyday language. It consists of a “word dictionary” (translated) linking words and meanings (concepts), a “concept dictionary” (translated) showing the up-and-down relations between concepts, a “co-occurrence dictionary” (translated) of how words connect, a Japanese–English “bilingual dictionary” (translated), and more.
The figure on the right is a small dictionary that borrows this idea. The word “atsui” (あつい; thick, hot to the touch, or hot weather: 厚い・熱い・暑い) picks its meaning by the pair of words it appears with. Even when there is no pair, you can go up through the concepts (coffee → drink) and choose if a pair exists there.
According to the current information page, it has about 270,000 Japanese words, about 410,000 concepts and about 900,000 co-occurrence phrases. The final evaluation also wrote that collecting knowledge from outside language is beyond what ICOT alone can do, and that it pins its hopes on EDR and on CYC at MCC in the United States.
Score = 2 if there is direct co-occurrence, 1 if there is no direct co-occurrence but there is co-occurrence at the concept above, and 0 if there is neither
co-occurrence
a pair of words that appear together
how the score is set
scoring unique to this site
Money and size
The money came from the Key Technology Research Promotion Center and 8 computer companies. The total amount is not given in the sources we saw.
At the time of the 1991 white paper it was just over 200,000 words each in Japanese and English and just under 500,000 concept headings; the current information page gives 270,000 Japanese words, 190,000 English words and 410,000 concepts. This is the difference between the midway numbers and the numbers after revision.
Sources & confidence (1)Building a dictionary of meaning by hand
April 1986–1995Verified in the originalSources disagreeJapan
Japan Electronic Dictionary Research Institute (EDR): building by hand, over 9 years, a huge dictionary that lets computers handle the meaning of Japanese and English (1986–1995)
As a foundation for getting computers to understand natural language, a company to develop a large-scale, high-quality “electronic dictionary”, the “Japan Electronic Dictionary Research Institute (EDR)”, was established, and the project began in April 1986. The Japan Key Technology Center invested in it (AI White Paper 1991), and eight computer companies also put in funds (EDR's 1996 paper). The institute made a “word dictionary” describing the grammar and meaning of words, a “concept dictionary” organizing the relations among concepts, a “co-occurrence dictionary” recording how words go together, and a “bilingual dictionary” showing the correspondences between Japanese and English, among others. The dictionaries now introduced on the page of the National Institute of Information and Communications Technology (NICT) are of this size: about 270,000 Japanese words, about 190,000 English words, about 410,000 concepts, about 200,000 Japanese example sentences and so on. The final evaluation of the Fifth Generation project wrote that collecting the vast knowledge that lies outside of words has limits for ICOT alone, and expected much from EDR's dictionaries.
What we still don’t know:The total cost of the project is not stated in the sources we looked at.
4 sources and how they were checked
Where sources disagree:Period: the AI White Paper 1991 says “started in April 1986 (Showa 61), planned to run until March 1995 (Heisei 7),” and EDR's 1996 paper says “nine years, fiscal 1986–1994” (the same period written in months and in fiscal years). Dictionary size: at the time of the 1991 white paper, a little over 200,000 words each for Japanese and English and just under 500,000 concept headings; on the current NICT page, 270,000 Japanese words, 190,000 English words and 410,000 concepts (the difference between the numbers during development and after repeated revisions).
What the figure shows:For a computer to choose a word's meaning, it needs a dictionary with large numbers of word combinations and broader–narrower word relationships collected by people.
This is the same mechanism as "Try it", written as Python code. Press ▶ Run and the result appears below. Change the numbers and run it again.
u07_cooccurrence.py
161992–1995 · Evaluation
“Failed” by which yardstick?
Was the Fifth Generation a success or a failure?
In June 1992, after the international conference FGCS'92 where the results were presented, an editorial in the newspaper The Baltimore Sun wrote of the plan's . But in that conference's keynote speech, the director, Kazuhiro Fuchi, mentioned that (translated). Fuchi explained that the source of such reviews was the big picture described at the 1981 conference, that the plan that began in 1982 had a more modest content, and said he believed it was a success.
The 1993 final evaluation summarized the results goal by goal. It exceeded the goal for inference speed, but did not get as far as building mechanisms for interacting with people by speech and images. Overall it said the project (translated). The 1995 MITI Research Review said the view that it failed because it did not become a product does not fully understand the project's philosophy, but also wrote that the hardware mechanisms would not be used in future technology. Nilsson (2010) writes that many people think most of the results now have only historical significance. In the figure on the right, switch the yardstick and the same plan's report card changes.
What you think you knowThe popular story:The Fifth Generation failed (and Japan admitted it) The original source says:At the 1992 conference the director did not admit failure; he explained that the source of the reviews calling it a failure was the big picture of 1981. The final evaluation too is written goal by goal. That the computers and software that were built did not spread in the world is pointed out both by the MITI Research Review and by books from outside researchers. The answer changes depending on what you count as the goal.
What we don't know
We have not been able to confirm where the reputation of failure first spread from in the United States. We have not been able to open the June 1992 New York Times article . The editorial also does not say what it based “acknowledged” on.
Sources & confidence (3)“Failed” by which yardstick?
June 1992–1995Verified in the originalSources disagreeJapan
The story that “the Fifth Generation Computer failed”: what the people involved, the evaluation committee and outside researchers each wrote
In June 1992, right after the international conference that wrapped up the project (FGCS'92), an editorial in the newspaper The Baltimore Sun (written for American companies) wrote of the project's . But in the keynote speech at that conference, the director, Kazuhiro Fuchi, introduced the situation by saying (translated). He then explained that big images such as (translated) and (translated) had spread from the 1981 conference before the project. He said that the project begun in 1982 had a more modest content (parallel inference), and that he believed it was a success. The promotion committee's final evaluation (1993) also said it had “fully met the initial expectations” (translated). A report of the Ministry of International Trade and Industry's research institute (1995) wrote that calling it a failure because it did not become a product is an opinion that does not fully understand the project's aims (basic research and international contribution). At the same time, the same report also writes that the structure of the computers made is not expected to be used in future technology, because of the progress of small computers. A history book by an outside researcher (Nilsson 2010) writes that many people think most of the results now have only historical significance.
The popular story:It is said that “the Fifth Generation Computer failed (and Japan admitted it)” The original source says:At the 1992 conference, the director did not admit failure. He said the source of the (translated) was the big dream image of 1981, and that the project goals from 1982 (parallel inference) were achieved. The promotion committee's final evaluation too said it had “fully met the initial expectations” (translated). However, the fact that the hardware and software developed did not spread is written in both the report of the Ministry of International Trade and Industry's research institute and the book by an outside researcher. The answer changes depending on what you take as the goal.
What we still don’t know:Where the reputation of first spread from in the United States (for example the June 1992 New York Times article ) has not been confirmed, because we could not open that article.
5 sources and how they were checked
Where sources disagree:About the same June 1992 conference, the newspaper editorial wrote that the failure was acknowledged. The director's keynote introduced that there are commentaries saying it is a failure, and then said he believes it was a success. What statement (whose words) the editorial based its “acknowledged” on is not written in the editorial.
The Fifth Generation final evaluation report: inference speed exceeded the target, but a system for talking with people was not built
The committee that had followed the project's progress (the Committee for the Promotion of Basic Computer Technology Development) compiled a final evaluation report at its end. It reported that PIM, a machine that runs many computers in parallel to carry out inference (drawing conclusions from rules), exceeded the target of 100 million inferences per second, reaching more than 150 million. On the other hand, for systems that deal with people through speech and images, it was found that a huge amount of background knowledge is needed. So partway through, the work was narrowed to the basics of natural language processing, and the report records that it did not go as far as building a concrete system. Overall, it sums up that the project (translated).
The popular story:“The Fifth Generation ended without reaching its goals,” people tend to say, lumping everything together The original source says:The government's final evaluation writes goal by goal. The speed of inference (100 million per second) was exceeded. It says itself that a system for talking with people through speech and images was not built.
What we still don’t know:This evaluation was written by the side that ran the project (the government committee). For outside evaluations, see other records (the MITI Research Review 1995, Nilsson 2010, and others).
2 sources and how they were checked
Fifth Generation Computer Project: Final Evaluation Report (Academic and Technical Evaluation Working Group, Committee for the Promotion of Basic Computer Technology Development · 1993 · p.10 Degree of achievement of the original goals for parallel sy… · Original source)
The 1981 forecast: it is said that in the early 1990s computers with eyes, ears, and a mouth, that understand speech and can do automatic translation and interpreting, will appear
The report of the survey before the plan (1979–1981; chaired by Tohru Moto-oka of the University of Tokyo) stated that the goal was to complete a prototype in the year 1990. A booklet explaining it for the general public wrote that fifth-generation computers are (translated) and that research must start early to (translated) (to build a trial machine). As the pictures of what they aimed for, it listed eight functions, including (translated) that can handle sound, figures, images, and characters as they are; (translated) that can be spoken to in everyday language (and so can also do automatic translation and interpreting); computers that have common sense and solve unfamiliar problems by reasoning; and computers that can be used without writing programs.
The popular story:It is said that the Fifth Generation promised to build, in 10 years, a computer that can talk like a human The original source says:The 1981 survey documents do say “computers that understand speech” (translated), (translated), and (translated) (a trial machine). However, in 1992 the director, Kazuhiro Fuchi, said that in the project that actually began in 1982, building applications such as machine translation had been dropped from the goals and the project was narrowed to machines for reasoning. The story changes depending on which year's, and which document's, goals you mean.
What happened next:The project ran in fiscal years 1982–1992. According to the 1993 final evaluation report, the goal for inference speed (100 million per second) was met and exceeded, at over 150 million per second. On the other hand, the report says that systems that interact with people by voice or images turned out to need a huge amount of background knowledge, so partway through the project narrowed its focus to a base for language research and did not get as far as building concrete systems. In 1992 the director, Kazuhiro Fuchi, said that building application systems such as machine translation had been dropped from the goals when the project began in 1982.
What we still don’t know:The (translated) (1981; 7 groups, 26 tasks), which are said to set numerical goals such as the number of words translated and accuracy, could not be found and have not been checked.
Computers of the Fifth Generation (booklet) (Japan Information Processing Development Center (JIPDEC) (incorporated foundation) · 1981 · p.参1 Reference: Functions of fifth generation computers · Original source)
What the figure shows:Whether the Fifth Generation project succeeded or failed depends on which goals are used to judge it.
171987–1995 · Downsizing
Special machines and ordinary computers
Why didn't the world's fastest computer spread?
PIM's inference speed was about 100 times that of a general-purpose mainframe. But according to Nilsson (2010), even ordinary workstations and personal computers became high-performance and surpassed PIM. The MITI Research Review (1995) also wrote that the hardware mechanisms would not be used in future technology because of downsizing (the move to smaller computers) and other trends.
Move how fast the ordinary machine grows in the figure on the right. The special machine stays at the speed it had when it was built. Even with a 100-times gap, if the ordinary machine becomes 1.5 times faster each year it catches up in about 11 years (the growth rate is a provisional number).
The same thing happened to Lisp machines, computers made specially for AI. In the recollection of one of their developers, Lisp came to run well on workstations such as Sun's, and the advantage of the special machines suddenly shrank. The 1991 white paper also writes that the computers used for AI in Japan had become mainly ordinary models such as personal computers.
n = log(gap) ÷ log(G)
n
years until it catches up
gap
the gap in speed at the start (100)
G
how many times faster the ordinary machine gets in a year
Was it only the machines that went unused?
Nilsson also writes that in the late 1980s and the 1990s, GUIs, operated on a screen, spread, and for human interaction AI language understanding and speech recognition were no longer needed much, at least for a while.
The 1991 white paper also writes that Japan did not go through an (translated) like the one seen earlier in the United States (next chapter).
Sources & confidence (3)Special machines and ordinary computers
June 1992–1995Verified in the originalSources disagreeJapan
The story that “the Fifth Generation Computer failed”: what the people involved, the evaluation committee and outside researchers each wrote
In June 1992, right after the international conference that wrapped up the project (FGCS'92), an editorial in the newspaper The Baltimore Sun (written for American companies) wrote of the project's . But in the keynote speech at that conference, the director, Kazuhiro Fuchi, introduced the situation by saying (translated). He then explained that big images such as (translated) and (translated) had spread from the 1981 conference before the project. He said that the project begun in 1982 had a more modest content (parallel inference), and that he believed it was a success. The promotion committee's final evaluation (1993) also said it had “fully met the initial expectations” (translated). A report of the Ministry of International Trade and Industry's research institute (1995) wrote that calling it a failure because it did not become a product is an opinion that does not fully understand the project's aims (basic research and international contribution). At the same time, the same report also writes that the structure of the computers made is not expected to be used in future technology, because of the progress of small computers. A history book by an outside researcher (Nilsson 2010) writes that many people think most of the results now have only historical significance.
The popular story:It is said that “the Fifth Generation Computer failed (and Japan admitted it)” The original source says:At the 1992 conference, the director did not admit failure. He said the source of the (translated) was the big dream image of 1981, and that the project goals from 1982 (parallel inference) were achieved. The promotion committee's final evaluation too said it had “fully met the initial expectations” (translated). However, the fact that the hardware and software developed did not spread is written in both the report of the Ministry of International Trade and Industry's research institute and the book by an outside researcher. The answer changes depending on what you take as the goal.
What we still don’t know:Where the reputation of first spread from in the United States (for example the June 1992 New York Times article ) has not been confirmed, because we could not open that article.
5 sources and how they were checked
Where sources disagree:About the same June 1992 conference, the newspaper editorial wrote that the failure was acknowledged. The director's keynote introduced that there are commentaries saying it is a failure, and then said he believes it was a success. What statement (whose words) the editorial based its “acknowledged” on is not written in the editorial.
Companies making AI-specific “Lisp machines” lose to ordinary workstations (late 1980s)
From around 1980, companies such as Symbolics built and sold “Lisp machines,” special computers that run Lisp, a programming language often used in AI research, fast. According to the recollection of one of Symbolics' founders, new from Sun, Apollo and others appeared, and even ordinary computers could run Lisp quite well. So the advantage of the dedicated machines suddenly became small. He also writes that most of the sales depended on expectations for expert systems (AI into which experts' knowledge is written), and that the backlash from those expectations swelling too far came around 1988 and became “The AI Winter.” The company repeatedly cut staff. According to a student paper (2001) written for an MIT course, Symbolics' product sales fell to $101.6 million, $82.1 million and $55.6 million in fiscal 1986, 1987 and 1988. In Japan too, the 1991 white paper wrote that the machines used for AI had come to center not on dedicated machines but on ordinary computers, especially personal computers.
The popular story:It is often said that in 1987 the AI winter suddenly collapsed the Lisp machine market The original source says:In the founder's recollection, two reasons overlapped: the technical change of ordinary workstations (such as Sun) and Lisp running on ordinary machines coming out, and the backlash from expectations for AI into which experts' knowledge is written swelling too far (around 1988). In the student paper's figures too, sales fall over several years starting from fiscal 1986.
What we still don’t know:We have not opened primary company documents such as the annual reports of Symbolics or LMI. We have not obtained MIT's 1999 bachelor's thesis because a human-verification screen appeared on MIT's site.
5 sources and how they were checked
Where sources disagree:When it collapsed: the student paper says “crisis by 1987,” and the recollection says “a big backlash around 1988.” The sales figures were confirmed only in the student paper (we have not opened the company's annual reports themselves).
AI White Paper 1991: AI Technology and Its Use (Supervised by the Electronics Policy Division, Machinery and Information Industries Bureau, Ministry of International Trade and Industry (MITI); edited by the ICOT-JIPDEC AI Center · 1991 · p.11 Part 2, Chapter 1 (survey feature 7) · Original source)
20 March 1991Verified in the originalSources disagreeJapan
1991, a government-supervised white paper: Japan (translated)
The wrote that Japan did not enter an AI “winter” (translated) like the one seen earlier in the United States. As the reason, it suggests that users may have found for themselves the uses in which AI is effective (it also writes that the United States had already been through its “winter”). In the white paper's survey (December 1989 to January 1990, sent to 7,084 establishments, with 1,844 replying), 37.8% of the responding establishments (26.2% of the establishments that use computers, 379 sites) had introduced AI. The white paper reported this as 17.2 points more than the 20.6% of the 1988 survey. Of the adopters, 46.2% used expert systems (mechanisms in which experts' knowledge is written) and 20.6% used machine translation.
The popular story:It is often said that after the second boom, Japan too went into winter along with the world The original source says:The government-supervised white paper of 1991 wrote that Japan had not entered a winter like the one seen in the United States. The survey too showed that the share of establishments that had introduced AI grew from 1988 to 1989 (though that is the share among the establishments that replied).
What we still don’t know:This white paper cannot tell how AI use in Japan changed after 1991 (when winter came, or whether it came at all). The survey results are also shares among only the establishments that replied (26% of those sent), not shares for Japan as a whole.
1 sources and how they were checked
Where sources disagree:Japan's “second winter”: this 1991 white paper writes that Japan “did not enter a winter like the one seen in the United States.” On the other hand, the 2016 white paper of the Ministry of Internal Affairs and Communications writes “a winter again from around 1995” (another record). They were written in different years, so it can also be read that in 1991 winter had not yet come.
AI White Paper 1991: AI Technology and Its Use (Supervised by the Electronics Policy Division, Machinery and Information Industries Bureau, Ministry of International Trade and Industry (MITI); edited by the ICOT-JIPDEC AI Center; published by the Japan Information Processing Development Association · 1991 · p.11 Part 2, Chapter 1, 1.1 (and the survey in Part 1, PDF pages… · Original source)
What the figure shows:Even a dedicated machine that is 100 times faster can be overtaken within a few years as ordinary machines become faster each year. The dedicated machine stays at the speed it had when built.
181992–1995 · Afterward
100 software packages fly off
When a national project ends, what happens to what it built?
The main body of the Fifth Generation was extended by 1 year from the planned 10 and ended in March 1993. From April the (translated title) continued for 2 years. Because the software that had been developed ran only on the dedicated Parallel Inference Machines, they developed KLIC, a language processor that converts KL1 into the C language, so that it also runs on UNIX parallel machines, which were then starting to spread.
The software was released free of charge as “ICOT Free Software”. At FGCS'92 in June 1992, 71 packages were released, then 6 in June 1993, 7 including KLIC in November 1994, and 16 added at the end of March 1995, for 100 in all. For software whose copyright belonged to ICOT rather than to the state, they released it after removing restrictions that would hinder research. State-funded research was all completed in March 1995.
When you press "Translate one line" in the figure on the right, a logic rule is turned into the loop (for) and the conditional branch (if) of an ordinary programming language. In 1992 the director, Kazuhiro Fuchi, said he hoped that (translated) and sprout in many places around the world.
What you think you knowThe popular story:The Fifth Generation was a 1982–1992 plan The original source says:The main body was extended by 1 year and continued until March 1993 (the end of fiscal 1992). After that, for 2 years, a plan was carried out to make the results run on UNIX computers too and release them free of charge. All state-funded research ended in March 1995.
Afterward
This site cannot confirm how much the 100 released packages were used afterward. It is also not known how the Parallel Inference Machines were used after the project ended.
On the period of the follow-on plan, the ICOT archive's conference page and the MITI Research Review give the 2 fiscal years 1993–1994, while the same archive's free software page gives fiscal 1993–1995.
Sources & confidence (1)100 software packages fly off
April 1993–1995Verified in the originalSources disagreeJapan
Two years after the Fifth Generation: 100 programs released free of charge, with UNIX support (1993–1995)
The main Fifth Generation project ended in March 1993, and in April a successor, the , began. Software that ran only on the parallel inference machines was ported, using a mechanism that translates KL1 into the C language (KLIC) and other means, so that it would also run on the UNIX parallel computers then appearing on the market. The software developed was released free of charge as “ICOT Free Software”, growing from 71 programs in June 1992 to a total of 100 by the end of March 1995. The software, whose copyright belongs to ICOT rather than the government, was released with restrictions that would hinder research removed. About 2.8 billion yen was spent on this successor plan, and all government-funded research and development was completed in March 1995.
The popular story:“The Fifth Generation was a 1982–1992 project,” is often all that is written (even the table in the notes from when this site was first being made says 1982–1992) The original source says:The main project ran one year past its planned 10 years and ended in March 1993 (the end of fiscal 1992). A plan then continued for two more years, making what had been built usable on UNIX computers and releasing it free. All research funded by the government ended in March 1995.
What we still don’t know:This record does not confirm how much the 100 released programs were used afterward (for example, download counts); Figure 5 on p.126 of the MITI Research Review shows access counts up to February 1995.
4 sources and how they were checked
Where sources disagree:Duration of the later plan: the conference page of the ICOT Archive and the table of contents of the committee's materials say “Heisei 5–6 (2 years)” (fiscal 1993–1994), and the MITI Research Review says “fiscal 1993–1994 (2 years)”. Only the IFS page of the same archive says “from Heisei 5 to Heisei 7” (fiscal 1993 to 1995) (it does not say whether this means that the release was finished by the end of March 1995).
What the figure shows:Software that once ran only on dedicated machines was translated into an ordinary programming language so it could run elsewhere, and 100 programs were released for free.
Closing
What this chapter still doesn't know
There is still a great deal we do not know about the Fifth Generation. The costs and staff time borne by the cooperating companies are not in the state's table. How the Parallel Inference Machines were used after the project ended, and how much the released software was used, have not been confirmed on this site either. We have not found the 1981 (translated title) that set out numerical targets such as numbers of translated words. Where the reputation of failure began in the United States (for example the 1992 New York Times article) we also cannot say, because we have not been able to check the article.
Whether it was a success or a failure changes with what you count as the goal. The dream picture of 1981, the plan that began in 1982, or whether it was actually used in society. Even for the same plan, change the yardstick and the report card changes. It also remains on record, from the middle of the boom, that in 1988 the director himself warned that goals set too high invite a backlash.
In the next chapter we look at the sprouts that were quietly growing as the second boom cooled. The 1991 white paper had written that Japan did not go through a winter like the one seen earlier in the United States. (For teachers: this chapter connects to the history in (1) of Information II, an advanced elective high-school informatics course in Japan, and to the development of information systems in (4). See the "For teachers" page for details.)