From the very beginning of this article I knew I was in for a treat. Most things were explained very clearly and with humor that not only made the reading enjoyable (not that it usually isn't) but that also helped to understand the particular point or issue at hand. One complaint I have, although minor and specific, is the example of finding the genetic component of IQ. I attribute this mostly to my opinion that all measures of IQ that I know of seem inadequate and often give misleading measures of intelligence. This is easily made up for by the inclusion of actually meaningful pictures, diagrams and the like. Cheese cutting is an often neglected domain in AI in Design research. I could have done with less examples of rules and run traces of AM though.

      I was surprised when Lenat presents the paradigm of AI research and uses the term "information processing" rather than "symbol processing" given his previous summary of different scientific views of people. Does he mean to use these terms interchangeably or is he saying there a fundamental difference between representations machines and people can work with? It would seem that information is explicit in representation while a symbol is not. Symbols seem more open to interpretation.

      It was interesting to see statements about automatic language translation and other such things that didn't seem very possible or to work well back in 1978 but seem to be common and generally accepted these days. Despite all the advances made over the last 23 years, we still have a ways to go though. Lenat says that it seems that it should be possible for a computer to provide driving directions for example, maybe even in Boston, but a quick look at the services available on the web shows there is still some work to be done.

      I was very happy when Lenat addresses the limitations imposed upon AM by not taking advantage of the knowledge it discovers to control its search. By treating the control knowledge (heuristics) and the concepts it seems very possible to improve the system. From the examples of rules and AM execution it doesn't seem like a trivial change to make though. Also, it seems limiting to think of all control knowledge as heuristics.

      "AI programs make explicit, in a very usable form, the knowledge necessary to perform some complicated activity, thereby de-mystifying it." What about neural networks?