weakly related to any improvements that the AI scientists like me and Bruce had made to the reasoning
processes used in DENDRAL’s hypothesis formation.
So in 1968, I called this observation the “Knowledge is
Power Hypothesis.” One data point. Later, as the evidence accumulated from dozens of — or hundreds of
— expert systems, I changed the word “hypothesis” to
“principle.” The title of the 1968 paper was specifically worded to contrast what we called the DENDRAL
case study with the main paradigm of the first generation of AI that focused on the generality of problem-solving. Those of you who are old enough in the audience remember GPS [General Problem Solver]. This
was a major paradigm shift for AI research, but it took
more than five years for the new paradigm to take
Continued modeling of human experts, and in
particular scientists and engineers, led to expert sys-
tems that, for very specific kinds of expertise, could
meet and even exceed the performance of human
experts. This achievement of instrumentality — a
novel capability to do things, namely to exceed some
performances of human experts — eventually led to
a great enthusiasm for expert systems within the arti-
ficial intelligence community and its military
patrons, then quickly drawing in corporations,
investors, entrepreneurs, and the popular press.
Yet the route to these enthusiasms was painstaking
work along the same two developmental lines for
modeling human expertise as computer systems:
changes to the inductive reasoning processes and
also to the representation of expert knowledge and
the means of making that representation. Bruce
Buchanan concentrated his efforts on the latter,
which, he explained eventually became known as
Well, we didn’t use the term “knowledge engineering”
until the 1970s, but we did talk, in a 1969 paper that
Ed and I were coauthors of with Georgia Sutherland,
about knowledge elicitation in AI. It was at a machine
intelligence workshop and people there were somewhat stunned that we were talking about organic
chemistry. John McCarthy rescued me during a talk by
saying to somebody who was giving me a hard time,
“Sit down, be quiet, you might learn something.” I
forever after loved that man.
Well, there were other groups working on knowledge
representation at the same time. Harry Pople and Jack
Myers at [the University of Pittsburgh] were working
with an emphasis on ontologies and mechanisms.
Peter Szolovits was working with Dr. Bill Schwartz, and
that led to a lot of work on the object-oriented frames
paradigm. Cas Kulikowski was working on knowledge
engineering with Dr. Aaron Safir at Rutgers. There was
work in Europe … There was a lot of isolated work in
France replicating some of the early expert systems
work, and several projects in France from commercial
firms, Schlumberger and Elf Aquitaine being two of
the most important. The Japanese Institute for New
Generation [Computer] Technology, ICOT, was working on fifth-generation computing largely from a
point of view of logic. The French were using Prolog
and so did the Japanese.
So I think our lesson there, the important part, was in
coding knowledge. The language you use — Prolog or
LISP or something else — it didn’t matter nearly so
much as the paradigm of starting with an expert’s
knowledge. But we also saw in that time that knowledge engineering could focus on the objects and their
relationships in an ontology: a hierarchy. They could
focus on the inferential mechanisms that were going
on, and in DENDRAL we were very much interested in
what we called the “situation-action rules” at the time.
There was an action in the right-hand side of the rule,
not just a Prolog kind of logical assertion.
For Buchanan, as with Feigenbaum, the motiva-
tion of intelligibility was, at least initially, primary for
the development of expert systems. Buchanan
Well, I was fascinated with the reasoning process....
My dissertation [on the philosophy of science] was on
the process of discovery and trying to orient it into a
framework. In the middle of my dissertation, I got to
know Ed Feigenbaum in 1963 and began reading the
AAAI Archive File Photo.