The Fifth Generation Computer Project — MITI and ICOT's Parallel Inference Machines

The Fifth Generation Computer Systems project was a national research programme that Japan's Ministry of International Trade and Industry (MITI, today's Ministry of Economy, Trade and Industry) started in fiscal 1982. The Institute for New Generation Computer Technology (ICOT) was established in April 1982 to run it, with the technical goal of research and development on new computer technology oriented toward knowledge information processing. At its centre were parallel inference machines, which ran logic-based reasoning across many processors at once, and a parallel logic programming language to drive them. The project ended in March 1993, and the government's final evaluation concluded that it had met its initial goals. Yet almost none of its results became products, and Japan's computer industry went on to fall behind in the era of PCs and networks — something the project's own leaders later wrote themselves.

The size of the budget, and the rule-based AI that spread through companies in the same years, are covered on the expert-systems boom page. This page deals with why the project came about, what it built, how the world reacted and what it left behind.

Why a "fifth generation"?

Shunichi Uchida, who later headed ICOT, set out the background in an account written in 2005. From around 1965, he explains, MITI supported domestic manufacturers through subsidies and large-scale industrial technology programmes, and the strategy of catching up with America and Europe paid off spectacularly in consumer electronics, semiconductors and cars. But it also brought semiconductor trade friction and growing accusations that Japan contributed nothing to basic research and only copied. MITI needed to show that Japan, too, could do original research and contribute to the world.

A committee chaired by Tohru Moto-oka of the University of Tokyo spent about two years working out the content, and a MITI official, Nakano, coined the name "fifth generation computer". The plan adopted was the one proposed by Kazuhiro Fuchi's group at the Electrotechnical Laboratory: a machine whose main application would be artificial intelligence, built on parallel hardware of around 1,000 processors. Uchida recalls it as a unique, non-von Neumann computer that nobody in the world had conceived of, and writes that the parallel inference language, the parallel operating system and nearly everything else had to start from zero. The introduction to the final evaluation report says the name was chosen to signal a technology that would be creative and pioneering by international standards.

What it built

According to the final evaluation report, the prototype system was built consistently around a single idea, the parallel logic paradigm, from the hardware up to the application software.

The report also listed figures among its "notable results". PIM reached an inference speed above 150 MLIPS (150 million logical inferences per second), which the report called about 100 times a general-purpose mainframe and the fastest in the world at the time. The parallel theorem prover MGTP showed a speed-up of more than 200 times on 256 processors and automatically proved part of a previously unsolved problem in group theory. All of these are the self-assessment of the committee that promoted the project.

How the world reacted — America, Britain, Europe

Abroad, the project was first read as a threat. In their 1983 book The Fifth Generation, Edward Feigenbaum and Pamela McCorduck argued that whichever country first mastered reasoning computers would gain great economic and strategic advantages. According to a review in The Christian Science Monitor, the two warned that America was rapidly losing its technological lead and urged a forceful response, including a national laboratory. The review quotes them: "We are writing this book because we are worried."

Other countries launched their own programmes.

Uchida writes that the United States judged Japan incapable of building the fifth generation computer on its own, but expected the AI and massively parallel research inside it to produce important technologies, and that DARPA started an AI-centred programme with a larger budget than Japan's. The National Research Council's history, however, does not mention Japan when it describes the announcement of Strategic Computing. How far the Japanese project drove each country's decisions depends on which source one reads.

In August 1983 The Christian Science Monitor wrote that Japan's drive had spurred similar collaborative ventures in America and Europe, and also that the West's first response to Japan's invitation to collaborate had been mainly scorn: researchers feared Japan would take Western results without giving much back. Brian Oakley, who headed Britain's Alvey directorate, told the paper that the right thing was to get organised first and only then talk about collaboration.

A legacy of open research

At home the project also took an unusual decision. According to Uchida, MITI declared from the start that the results would be made public — a break with the standard at the time, under which results paid for by taxpayers belonged to the state and were released only for a fee. That decision opened the door to exchanges with researchers abroad. About 200 researchers and engineers were seconded to ICOT, and about 200 more took part in working groups linking ICOT with universities.

The final evaluation report presented the hosting of foreign researchers and the free release of software as a model for future national research, and quoted Gilles Kahn of the French national research institute INRIA to the effect that before the Fifth Generation one could see Japan's technology but not the faces of its researchers. The software was published as "ICOT Free Software": 71 programs at the FGCS'92 international conference in June 1992, reaching 100 by the end of March 1995. KLIC, which ran KL1 on UNIX machines, continued to be improved by a volunteer group after the project ended.

Results that never became products

Verdicts on the outcome divide sharply by vantage point. The final evaluation report concluded that the project had fully met its initial expectations and offered a model for Japanese national projects. Uchida's same account acknowledges that while it trained people — about 80 participants moved to universities — its direct outputs were too far ahead of their time, and few could be turned into products. Meanwhile the market moved from mainframes to workstations and PCs linked by fast networks. Japanese manufacturers, he writes, expected the mainframe era to last another ten years, misjudged, and fell behind on the spread of the internet and PCs; missing the shift to distributed software led by UNIX was, in his words, close to fatal for the industry. The Wikipedia article lists the rising performance of general-purpose processors, which overtook the specialised parallel hardware, among the reasons the project failed commercially.

Uchida also writes that this was the last large programme on the "Japan Inc." model, with the state leading and industry moving as one. Expectations for AI cooled at the end of the 1980s, and in the 2010s the field moved from hand-written rules to deep learning that learns from data (the turning point is AlexNet). Japan would not have a law dealing directly with AI until the AI Promotion Act of 2025, more than forty years later. The Wikipedia article adds that as processor clock speeds stopped climbing and parallel computing spread through multi-core chips and GPUs, many of the project's themes came back in a different form; the final evaluation report itself had noted in 1993 that parallel computers were coming to be seen as the mainstream of the next era.