In 1984 Douglas Lenat began Cyc, a project to write down, by hand, the common-sense knowledge that humans take for granted - that water flows downhill, that you cannot be in two places at once, that a dead person stays dead. The bet was that intelligence needed an explicit base of millions of such facts and rules, and that once enough were encoded, reasoning would follow.
By the time Lenat described the effort in a 1995 Communications of the ACM article, the scale was already enormous. He reports that “since 1984, a person-century of effort has gone into building CYC,” a schema of roughly 10^5 general concepts with on the order of 10^6 common-sense axioms hand-crafted and entered, and millions more inferred. The article reviews the assumptions behind doing a project this large and the lessons learned along the way.
The remarkable thing is that the project never stopped. Cyc has run continuously for about four decades. Its company, Cycorp, still markets the technology - its website describes Cyc as “Machine Reasoning AI that uses codified human common sense and knowledge (not patterns and statistics) for human-like cognitive processing,” and points to deployments in industry. What it never delivered was the general intelligence that the manual-encoding thesis pointed toward.
The contrast with what came later is the point. The dominant approach to common sense turned out not to be hand-coded axioms at all but statistical learning from vast text - “patterns and statistics,” the very thing Cyc’s own marketing positions itself against. Cyc is not a failure in the sense of a project that crashed; it is a dead end in the sense of a road that was followed faithfully for forty years while the traffic went somewhere else.