# What was Japan's Fifth Generation computer project, and why did it worry the rest of the world?

Date: 2020-01-06 · Author: Robert Scales · https://guidetoartificialintelligence.com/en/learn/what-was-japans-fifth-generation-computer-project/

**Answer:** The Fifth Generation Computer Systems project was a ten-year national program launched by Japan's Ministry of International Trade and Industry in 1982 to build parallel computers that reason with logic, as a platform for artificial intelligence. It alarmed the United States and Europe into starting rival programs, produced impressive prototype machines, and ended in 1992 having spent about 50 billion yen (roughly $400 million) without the commercial breakthrough it promised.

In the early 1980s, Japan seemed to win every industry it entered: steel, cars, consumer electronics, memory chips. So when its powerful trade ministry announced a plan to leapfrog the world in computing, and to do it with artificial intelligence, people in the United States and Europe took it very seriously. The plan was called the Fifth Generation Computer Systems project, usually shortened to FGCS.

The goal was a new kind of computer. Instead of mainly calculating, it would reason: draw conclusions from large stores of knowledge, fast enough to support programs that could understand language, diagnose diseases or analyze legal cases. To get there, Japan bet on two ideas at once, **logic programming** and **massively parallel** hardware.

## What happened, and when

The project had a long run-up. In the mid-1970s, the Ministry of International Trade and Industry (MITI) asked JIPDEC, the Japan Information Processing Development Center, for a study of where computing was heading. In 1979 it funded a deeper three-year study with industry and universities, and the phrase "fifth-generation computer" came into use during that work. MITI also held an international conference to present its ideas and invite foreign researchers.

In 1982 the government gave the go-ahead and set up ICOT, the Institute for New Generation Computer Technology, in Tokyo, funded jointly with Japan's computer companies. Kazuhiro Fuchi led the project as its director. The plan ran for ten years in three stages: three years of initial research, four years building subsystems, and a final three years to complete a working prototype.

The name came from the usual way of counting computer generations by hardware: vacuum tubes, then transistors, then integrated circuits, then microprocessors. A "fifth" generation signaled a leap beyond all of them.

## How it was supposed to work

At the heart of FGCS was logic programming, the approach behind the language Prolog, created in Marseille in 1972. In logic programming, a program is a set of facts and rules, and running the program means proving that a statement follows from them. One of the field's founders summed it up as two equations: a program is a set of axioms, and a computation is a proof.

The planners believed logic was the "missing link" between knowledge, stored as facts and rules, and parallel hardware. Many proof steps could, in principle, run at the same time on many processors. So they measured speed in a new unit, **LIPS**, logical inferences per second, where one inference is roughly one step of reasoning.

Here is the scale of the ambition:

| Machine | Speed in LIPS | Compared with a typical workstation |
|---|---|---|
| Typical workstation, early 1980s | about 100,000 | 1× |
| FGCS target, low end | 100 million | 1,000× |
| FGCS target, high end | 1 billion | 10,000× |

To make the idea concrete, imagine a medical knowledge base with a rule such as "a patient with symptoms A and B may have condition C." A query checks thousands of rules like that. A fifth-generation machine was meant to try many of them at once, on hundreds of processors, rather than one after another.

> **Try it:** Take a short Prolog-style family program (a few `parent` facts and a `grandparent` rule) and count how many fact checks are needed to answer one question. Then imagine 1,000 workers each taking one check at the same time, and think about which steps must still wait for others to finish.

In 1982, during a visit to ICOT, the computer scientist Ehud Shapiro designed Concurrent Prolog, a logic language built for running many processes at once. His work pushed the project toward "concurrent logic programming." That line led to Ueda's Guarded Horn Clauses (GHC) and finally to KL1, the project's core language.

## Why it worried the world

FGCS landed at a time of anxiety about Japanese competition. In response, a group of U.S. firms set up the Microelectronics and Computer Technology Corporation (MCC), a research consortium in Austin, Texas, and the U.S. Defense Department launched its Strategic Computing Initiative, which included work on intelligent systems such as self-navigating tanks. Britain started the Alvey programme, and Europe launched ESPRIT and a joint research center in Munich backed by ICL, Bull and Siemens. Some American scientists later admitted that colleagues may have overstated the threat in order to win more research funding at home.

## What it achieved

By 1992 ICOT had built five Parallel Inference Machines, with up to 1,000 processors working together, plus the software to program them and applications such as a parallel database, a legal reasoning system and an automated theorem prover. Experts described some of this as technically impressive.

Commercially, though, it missed. General-purpose computers, like the workstations made by Sun Microsystems, became fast and cheap enough that few buyers wanted specialized AI machines. The concurrent logic languages also never fully joined logic-based knowledge with parallel hardware as cleanly as hoped. After spending about 50 billion yen, over $400 million at 1992 exchange rates, the project ended, and MITI offered its software free to anyone, including foreigners. Stanford's Edward Feigenbaum, who had co-written a 1983 book about the project, observed that no one, even in Japan, was using the technology.

Its defenders pointed to a different payoff: hundreds or perhaps thousands of Japanese engineers trained in advanced computer science. As one Japanese research director put it, "I think the side effect is the main effect."

## What lessons does the Fifth Generation project hold for national AI programs today?

The first lesson is about betting on one technical path. FGCS chose logic programming and specialized hardware early and stuck with them for ten years while the computer industry moved toward cheap, general-purpose machines. In 2020, many governments are announcing national AI strategies. A useful question to ask of each one is whether it funds broad skills and research, or a single machine or method.

The second lesson concerns expectations. The project's director argued that outsiders exaggerated its aims, which made the outcome look like a failure. Grand public promises, whether from planners or alarmed rivals, set a bar that real research rarely clears on schedule.

The third lesson is that the value of a program can lie in its people and its spillovers. FGCS trained a generation of engineers and pushed the United States and Europe to invest in computing research. The ideas also live on: parallel processing is now everywhere, from multi-core laptops to the graphics processors that train today's neural networks, although it is unclear how much the Japanese project itself contributed to that.

## Sources

- 'Fifth Generation' Became Japan's Lost Generation: https://web.archive.org/web/20191218221357/https://www.nytimes.com/1992/06/05/business/fifth-generation-became-japan-s-lost-generation.html
- Fifth generation computer: https://web.archive.org/web/20191127093726/https://en.wikipedia.org/wiki/Fifth_generation_computer