专家系统市场退潮

项目演示通过。规则库能在限定案例里给出像专家一样的建议。接下来才是昂贵的部分。

新情况出现时，谁去采访专家、澄清判断，再把它写成规则？两条规则冲突时，以哪一条为准？组织流程变化、产品更换、法规更新后，旧知识怎样识别和删除？专家的经验常带条件、例外与默认背景，写进系统时却必须变成明确的符号。规则越多，修改一处影响另一处的机会也越多。所谓“知识获取瓶颈”并不发生在系统启动前一次，而是贯穿它的整个寿命。斯坦福 AI100 报告与计算机历史博物馆对专家系统先驱的口述史，都反复回到同一主题：把默会知识变成可维护的代码，劳动量被系统性低估。

1980 年代前半，企业为专家系统购买 Lisp 机器、开发工具和知识工程服务。Symbolics、Lisp Machines Inc. 等厂商出售面向 AI 开发的专用硬件；DEC 的 XCON（也称 R1）等配置系统被当作商业成功案例引用——它能根据客户订单约束生成 VAX 等机型的配置，减少人工配置错误。窄领域的成功被外推成广泛的知识自动化前景，专用软硬件由此形成市场。可当系统进入真实组织，购买价格只是第一笔费用。维护规则、验证输出、追踪环境变化，以及让稀缺专家持续参与，才决定它能否留下。

与此同时，通用工作站变得更快、更便宜。Sun 等工作站与不断改进的通用 CPU，使越来越多 AI 开发可以在非专用机器上完成；Common Lisp 与各类编译器也在通用硬件上成熟，削弱了“必须买 Lisp 机器才能做 AI”的销售叙事。Lisp 机器原本以适合人工智能开发的软硬件环境建立优势；当普通设备能够承担越来越多同类工作，专用机器的价格和封闭性便从卖点变成负担。需求下降，供应商收缩，围绕专用硬件、工具与咨询构成的市场在 1987 年后迅速退潮——年份是后设的标记，而不是某日开盘铃响起的崩盘。

这次收缩常被称为第二次 AI 寒冬，但“寒冬”容易掩盖它的商业性质。研究并未同时停止，规则系统也没有从软件中消失；编译器、诊断程序、业务规则引擎仍在工业软件里延续。变化的是企业对成本与收益的判断：一个能展示专业推理的系统，未必是一个能在预算内长期更新的生产系统。“人工智能”再次成为难以向投资者和采购部门解释的标签，一些项目转而使用“机器学习”“智能系统”等名称，以避开被灼伤的词。

专家系统的教训也不能被简化为“手写规则错了，数据学习才对”。后来的机器学习系统不再主要要求专家逐条编码知识，却仍要处理数据漂移、评估、监控、异常与责任归属。表示知识的方式改变了，生产环境不断变化这一事实没有改变。规则库变成了数据集与模型版本；维护账单换了科目，没有消失。

1987 年之后，市场重新计算的是一项被低估的长期负债。一次演示只需在选定案例里答对；生产系统必须跟着规则、人员和业务一起变化。知识工程最昂贵的部分，原来并不在把第一版规则写进去，而在此后没有终点的维护。账单上的科目后来改叫数据标注、模型监控与漂移处理，但“部署之后谁付钱让系统继续正确”这个问题没有退场。

The demo passed. The rule base could give expert-like advice on a bounded set of cases. Then came the expensive part.

When a new situation appears, who interviews the expert, clarifies the judgment, and writes it as a rule? When two rules conflict, which wins? After processes change, products turn over, or regulations update, how is old knowledge found and removed? Expert experience is full of conditions, exceptions, and default background that must become explicit symbols in a system. The more rules, the more chance a change in one place disturbs another. The “knowledge-acquisition bottleneck” does not happen once before launch; it runs through the system’s entire life. The Stanford AI100 report and Computer History Museum oral histories of expert-system pioneers return to the same theme: converting tacit knowledge into maintainable code systematically underestimates labor.

In the first half of the 1980s, enterprises bought Lisp machines, development tools, and knowledge-engineering services for expert systems. Vendors such as Symbolics and Lisp Machines Inc. sold specialized hardware aimed at AI development; DEC’s XCON (also called R1) and similar configuration systems were cited as commercial successes—generating VAX and related configurations from customer-order constraints and reducing human configuration errors. Narrow-domain success was extrapolated into a broad prospect of knowledge automation, and a market for specialized software and hardware formed. Once systems entered real organizations, purchase price was only the first invoice. Maintaining rules, validating outputs, tracking environmental change, and keeping scarce experts engaged decided whether a system stayed.

At the same time, general-purpose workstations grew faster and cheaper. Suns and improving general CPUs let more AI development run on non-specialized machines; Common Lisp and compilers matured on general hardware, weakening the sales story that AI required a Lisp machine. Lisp machines had built an advantage as software–hardware environments suited to AI work; when ordinary equipment could carry more of the same load, price and closedness flipped from selling points to burdens. Demand fell, suppliers contracted, and the market of specialized hardware, tools, and consulting retreated rapidly after 1987—the year is a retrospective marker, not a crash that rang the open-market bell on a single morning.

The contraction is often called the second AI winter, but “winter” can hide its commercial character. Research did not stop all at once, and rule systems did not vanish from software; compilers, diagnostic programs, and business-rule engines continued inside industrial systems. What changed was enterprise judgment of cost and return: a system that can display professional reasoning is not necessarily a production system that can be updated within budget. “Artificial intelligence” again became a label hard to defend to investors and purchasing departments; some projects switched to names such as “machine learning” or “intelligent systems” to avoid a burned word.

The lesson of expert systems also cannot be simplified to “handwritten rules were wrong; learning from data is right.” Later machine-learning systems no longer mainly require experts to encode knowledge line by line, yet still must handle data drift, evaluation, monitoring, anomalies, and assignment of responsibility. The representation of knowledge changed; the fact that production environments keep changing did not. Rule bases became datasets and model versions; the maintenance bill changed line items without disappearing.

After 1987 the market re-priced an underestimated long-term liability. A demo need only answer correctly on chosen cases; a production system must change with rules, people, and business. The most expensive part of knowledge engineering was never writing the first version of the rules, but the maintenance without an endpoint afterward. Line items later renamed data labeling, model monitoring, and drift handling. The question—who pays after deployment to keep the system correct—did not leave the room.