XCON 商业专家系统

1980 年代初，DEC 公司面临一个甜蜜的烦恼：VAX 小型机卖得很好，但每一台机器的配置都极其复杂——型号、内存、外设、电缆的组合数以万计，人工配置错误频出，报价和布线经常返工。这份烦人的工作，成了 AI 第一次大规模商业化的舞台。

卡内基梅隆的约翰·麦克德莫特接下这个挑战，写了一个叫 R1 的程序，后来改名为 XCON。它的原理很「规则化」：把 DEC 工程师配置机器的经验，一条条写成条件规则——如果订单里有某个组件，那么就配上对应的电源、线缆和插槽。输入一份订单规格，XCON 就能自动产出完整的配置清单，准确率甚至超过熟练员工。

DEC 把 XCON 投入生产后，效果立竿见影：配置错误大幅减少，每年节省数百万美元。这件事在业界引起轰动——它证明了 AI 不是只能解谜题、下棋的玩具，而能在真实企业的流水线上创造真金白银的回报。一大批公司受此鼓舞，纷纷效仿部署专家系统，1980 年代的专家系统产业热潮由此点燃。

热潮之下也有隐忧。XCON 的规则库需要不断维护——VAX 产品线一更新，几千条规则就要跟着改。这种「知识工程」的维护成本，是后来产业退潮的重要原因之一。但至少在 1980 年代中叶，XCON 是 AI 商业价值最响亮的宣言。

回看 XCON，它像是 AI 产业化的第一次演练：证明了价值，也暴露了局限。它告诉我们，把人类知识编码成规则可以赚钱，但规则系统很难自己成长。当 1990 年代机器学习重新抬头时，人们开始寻找另一种路径——让机器从数据里自己学，而不是靠人一条条喂规则。这条从 XCON 到深度学习的转折，正是 AI 商业史上一道深刻的弧线。

In the early 1980s, DEC faced a sweet problem: its VAX minicomputers sold extremely well, but configuring each machine was brutally complex. The combinations of models, memory, peripherals, and cables numbered in the tens of thousands; manual configuration was error-prone, and quotes and wiring plans often had to be redone. That tiresome job became the stage for AI's first large-scale commercialization.

John McDermott at Carnegie Mellon took on the challenge and wrote a program called R1, later renamed XCON. Its logic was thoroughly rule-based: DEC engineers' configuration experience was encoded as conditional rules—if an order contains a certain component, then add the matching power supply, cable, and slot. Feed in an order specification and XCON would automatically produce a complete configuration, with accuracy that matched or beat experienced staff.

The results were immediate. DEC put XCON into production, configuration errors fell sharply, and the company saved millions of dollars a year. The industry took notice: AI was not just a toy for puzzles and chess—it could generate real returns on a real factory floor. A wave of companies followed, deploying expert systems of their own, and the 1980s expert-system boom was lit.

Yet beneath the boom lay a warning. XCON's rule base had to be maintained constantly—every time the VAX product line changed, thousands of rules had to be updated. That "knowledge engineering" burden was one reason the industry later cooled. But in the mid-1980s, XCON was the loudest statement of AI's business value.

Looking back, XCON reads like a first rehearsal for the AI industry: it proved value and exposed limits. It showed that encoding human knowledge into rules could make money, but rule systems could not grow on their own. When machine learning resurged in the 1990s, people began seeking another path—letting machines learn from data rather than hand-feeding rules. That arc, from XCON to deep learning, is one of the deepest curves in AI business history.