Cyc 常识知识库项目启动

1984 年，道格拉斯·莱纳特启动了一个野心极大的项目：Cyc。目标很简单也很疯狂——把人类常识全部编码成机器能推理的知识。机器不仅要知道「水往低处流」「人都会死」这类事实，还要能在推理中调用它们。在当时的 AI 界，这被视为通往强人工智能的正途之一。

Cyc 的方法论，是符号主义「知识工程」路线的极致体现。一支团队年复一年地手工录入常识断言，试图覆盖「不言自明」却浩瀚无边的日常知识。项目延续了几十年，录入的知识数以百万条计。这种投入在 AI 史上是罕见的，它更像一场信仰驱动的长期主义实验。

但 Cyc 也遇到了知识工程路线无法回避的困境：常识太多、太碎、太依赖语境，手工编码的速度永远追不上现实世界的复杂度。当机器学习范式崛起，用数据自动学出知识表征时，Cyc 所代表的「手工知识」路线逐渐退居边缘。这成了 1980 年代 AI 寒冬之后，范式更替的一个缩影。

不过，把 Cyc 简单称为「失败」并不公平。它为知识图谱、本体工程提供了直接的方法论遗产，也深刻影响了后来对「知识如何表征」的思考。它提醒行业：常识推理很难靠规则堆出来，却也没有消失——只是换成了从海量数据中学习的形态。

回看 Cyc 项目，它像符号主义路线上的一次长途跋涉：走得最久、投入最大、却未达终点。但它的价值恰恰在过程中——它用数十年的实践，为整个行业标定了「知识工程」的边界与代价，也让后来的学习者用更清醒的目光看待「常识」这个词。

In 1984 Douglas Lenat launched a project of enormous ambition: Cyc. The goal was simple and also mad: encode all human common sense into machine-reasoning knowledge. Machines would not only know facts like "water flows downhill" and "people are mortal," but also call on them during reasoning. At the time, this was considered one of the legitimate roads to strong AI.

Cyc's methodology was the extreme embodiment of the symbolic "knowledge engineering" path. A team hand-entered commonsense assertions year after year, trying to cover the vast, "self-evident" knowledge of daily life. The project ran for decades and encoded millions of assertions. Such commitment is rare in AI history—it reads more like a faith-driven, long-horizon experiment.

But Cyc also met the unavoidable difficulty of knowledge engineering: common sense is too vast, too fragmented, and too context-dependent for hand-coding to ever keep pace with the complexity of the real world. As the machine-learning paradigm rose and models began learning knowledge representations from data automatically, the "hand-built knowledge" route that Cyc represented slipped to the margins. It became a microcosm of the paradigm shift after the 1980s AI winter.

Still, calling Cyc a simple "failure" is unfair. It left a direct methodological legacy for knowledge graphs and ontology engineering, and it deeply shaped thinking about "how knowledge should be represented." It reminded the industry that commonsense reasoning is hard to pile up from rules—yet it never vanished; it simply changed form, learned from vast data instead.

Looking back at the Cyc project, it reads like a long march on the symbolic path: the longest, the most heavily invested, yet never reaching its destination. But its value lies precisely in the journey—with decades of practice, it marked out the boundaries and costs of "knowledge engineering" for the whole industry, and made later learners look at the word "common sense" with clearer eyes.