The Knowledge-Acquisition Bottleneck

In the [YOU] on AI Field Guide

The cycle that began with [YOU] on AI describes a world in which the bottleneck has been bypassed. The process of getting knowledge into a machine no longer requires expert interviews, knowledge engineers, or years of rule-writing. A system trained on the internet has absorbed more knowledge than any team of knowledge engineers could have extracted in any number of years. The bottleneck is gone. What replaced it is different in kind: not the slowness of human extraction but the opacity of automated learning, not the brittleness of explicit rules but the unpredictability of learned patterns, not the limits of what experts could articulate but the limits of what the training corpus contained.

Feigenbaum's bottleneck thus did double duty. It named the obstacle that defeated his own method, and it predicted the shape of the method that would succeed: a way of getting knowledge into machines that did not require humans to articulate it explicitly. Whether the resulting knowledge—distributed, uninspectable, statistically approximated—is knowledge in the sense he meant is the deepest question the bypassing of the bottleneck leaves open.

Origin

The bottleneck became visible as expert systems scaled. Building DENDRAL and MYCIN was laborious but manageable; the domains were narrow and the key experts were available and cooperative. But as the enterprise expanded into new domains and the knowledge bases grew larger, the rate of knowledge acquisition failed to keep pace with the demand. Each new domain required starting over—finding new experts, conducting new interviews, writing new rules. Each change in the underlying domain required a human to update the rules by hand. The result was systems that were expensive to build, expensive to maintain, and fragile outside the narrow slice of the domain they had been built for.

The bottleneck was structural, not merely practical. It was rooted in the nature of expertise itself: most expert knowledge is not available for conscious inspection, cannot be articulated on demand, and resists translation into explicit propositions. The knowledge engineer's task was always partly archaeological—reconstructing implicit understanding from observable behavior—and archaeology scales badly. Feigenbaum invested heavily in tools and methods to widen the bottleneck, but the bottleneck held.

Key Ideas

Tacit Knowledge as the Root Cause. The bottleneck was ultimately a consequence of tacit knowledge—the vast, inarticulate substrate of expertise that resists explicit articulation. What made MYCIN's physician advisor competent was not a set of rules he could recite but a body of pattern recognition accumulated over thousands of cases and impossible to fully decompose. The knowledge engineer could capture the articulable surface; the tacit depth remained in the expert's head.

The Brittleness Consequence. Systems built from explicitly coded rules were brittle outside their intended domains, because the rules were designed for anticipated situations and the world produces unanticipated ones constantly. An expert system that diagnosed blood infections confidently and correctly could produce confident nonsense about a patient who also had a broken leg—not because the rules were wrong but because the rules had not anticipated the interaction. Common sense, which fills in the gaps that explicit rules leave open, resisted codification in any rule-based form.

The Bypass and Its Costs. Deep learning bypassed the bottleneck by replacing the human extraction of explicit knowledge with the automated extraction of implicit patterns from vast corpora. The result was systems that were not brittle in the same way—systems that generalized more gracefully across domains because they had absorbed more of the implicit background that human common sense relies on. But the bypass had costs: the extracted knowledge is opaque, uninspectable, and unreliable at the edges in ways that are harder to predict than the brittleness of explicit rules.