日本第五代计算机计划

1981 年，日本通产省发布了一份研究报告，宣布要研制第五代计算机。1982 年项目正式启动，目标定得极其宏大：十年内造出以逻辑推理为核心的智能计算机，能处理知识、理解自然语言、并行推理。当时的背景是日本在半导体和工业机器人领域势不可挡，欧美普遍担心——如果日本人把「智能」也做成了标准件，下一个计算时代就是日本的。

FGCS 的技术路线选得很鲜明：以 PROLOG 语言和逻辑程序设计为核心，围绕推理机、知识库、智能接口三大件展开。政府牵头，联合富士通、NEC、日立等大企业，投入的预算以百亿日元计。项目建了 ICOT（新一代计算机技术研究所），集中了当时日本最优秀的年轻研究人员，摆出了一副举国攻关的架势。

这个计划立即刺痛了西方。美国推出了战略计算计划（Strategic Computing Initiative），欧洲启动了 ESPRIT 项目。三个大玩家同时把 AI 写进国家科技战略——在此之前，AI 主要是大学里自娱自乐的研究课题。FGCS 在这一点上影响深远：它让「AI 竞争」第一次成为国家议题，这个框架一直延续到今天的算力竞争。

项目的实际进展则另一番景象。ICOT 确实造出了 PSI、PIM 等原型推理机，PROLOG 的研究也取得了扎实的学术成果。但「智能计算机」的核心承诺——处理自然语言、通用知识推理——远远没有实现。1990 年代初，随着泡沫经济破裂和人工智能寒冬加深，项目在 1995 年前后悄悄收场。事后复盘，批评者说它押错了技术路线：逻辑推理没有成为计算的基础，而连接主义在几年后带着深度学习的浪潮卷土重来。

FGCS 的遗产是双重的。对西方而言，它像一记警钟，促使美欧重新投入 AI 研究，客观上为后来的 AI 复兴储备了人才与项目。对日本而言，它是一场昂贵的试错，提醒后来的国家计划：把整个产业押在一种尚未被验证的技术愿景上，风险远超预期。今天回看，FGCS 既不是笑话，也不是丰碑，而是一次真实的、代价高昂的国家级赌局——它的成与败，都被后来的 AI 史反复引用。

In 1981 Japan's Ministry of International Trade and Industry (MITI) issued a report announcing a fifth-generation computer. The project formally began in 1982 with an enormous goal: build knowledge-processing machines centered on logic reasoning within ten years—computers that could handle knowledge, understand natural language, and reason in parallel. The backdrop was Japan's dominance in semiconductors and industrial robots, and broad Western anxiety that if Japan industrialized "intelligence" as a commodity, the next computing era would be Japan's.

The technical bet was distinctive: PROLOG and logic programming at the core, organized around three pillars—inference machines, knowledge bases, and intelligent interfaces. Government led, joined by Fujitsu, NEC, Hitachi and others, with a budget in the hundreds of billions of yen. MITI founded ICOT (the Institute for New Generation Computer Technology), gathered many of Japan's best young researchers, and presented a national project in full siege mode.

The plan stung the West immediately. The United States launched the Strategic Computing Initiative; Europe launched ESPRIT. Three major powers wrote AI into national science strategy at once—before FGCS, AI had been mostly a university research topic. In that respect FGCS changed the game permanently: it made "AI competition" a national-policy issue, a frame that persists into today's compute race.

Actual progress was a different story. ICOT did build prototype inference machines such as PSI and PIM, and PROLOG research yielded solid academic results. But the core promise—intelligent computers that process natural language and reason about general knowledge—was nowhere near realized. As the bubble economy deflated and the AI winter deepened, the project quietly wound down around 1995. Critics later said it bet on the wrong route: logic programming did not become the foundation of computing, and connectionism returned with the deep-learning wave a few years later.

FGCS's legacy is double-edged. For the West it was an alarm that pushed the US and Europe back into AI research, objectively helping stockpile talent and projects for the AI revival. For Japan it was an expensive experiment, reminding later national programs that staking an entire industry on an unproven technological vision carries risk far beyond the budget. Today FGCS reads neither as a joke nor a monument, but as a real, costly national bet—whose wins and losses AI history keeps citing.