When Cyc was first envisioned, the idea was to start by encoding the contents of a one volume desk encyclopedia, hence the name Cyc. Issues such as the representation of time, substances, agents, and causality were largely ignored in favor of highly specific issues. However, soon after, the team realized that it wasn't the contents of the encyclopedia that was urgent, it was the complement of it, i.e. the underlying common sense knowledge that the writers of encyclopedias and newspaper articles assumed that their readers already possessed. This seems like a step in the right direction. However, as I will argue below, it wasn't a sufficient step. Other reviewers [McDermott, 1993,Neches, 1993] are also skeptical about the premise that what is holding AI back is programs' lack of explicitly represented knowledge. McDermott comments in his usual bold style that ``you can't make the problems go away by burying them under a truckload of frames.''
Throughout the articles and books of the Cyc team, there is an implicit dream that shows itself in random places. For example in the middle of a section about how many entries Cyc will need, they say
``... probably another ten to fifty million entries would suffice for general intelligence (for example, the intelligence required for acquiring knowledge in school and in extra-curricular conversations). Quite a bit more may be needed for qualitatively super-human intelligence.''[Lenat and Guha, 1990]
This is not a proposition that they officially defend, that Cyc is going to encompass general intelligence when loaded with enough knowledge. But every once in a while, you can see glimpses of the idea that makes you feel that this is their background hope.
A related hope is their anticipation of a kind of crossover [Guha and Lenat, 1990] from manual knowledge entry to primarily automatic entry via natural language understanding. It is observed that the role of human knowledge enterers today, manually entering assertion after assertion, is akin to teachers who must instruct by surgically manipulating brains. Reaching the crossover would mean that the role of humans would become much more like tutors, with a great deal of question answering going on in both directions. Cyc would start reading books and newspapers itself, initiating a dialogue with humans to clarify ambiguous points.
When I heard about the crossover point, and the quest for knowledge supporting general intelligence, I thought what they would focus on is the knowledge of a five year-old, or the knowledge and inference capabilities that seem to be common to human species throughout all world cultures. I thought they were after constructing an infrastructure, not the building itself. Apparently, I was wrong. Here is a typical example of a Cyc axiom [Guha and Lenat, 1993]
(#%ist #%LargeCorpInternalsMtTranslated into English, it says that the human resources department of a company plays the primary role of mediating the hiring of employees.
(#%ForAll x (#%HumanResourcesDepartment #%allInstances)
(#%actsInCapacity x #%mediatorInProcesses
#%EmployeeHiring #%MainFunction)))
Certainly most of the Australian aboriginals, for example, probably haven't heard about large corporations. A five year-old might not know what a vice-president or a house speaker is. I certainly did not have much knowledge about what kind of a game baseball was before I came to this country, and I still don't understand how it works. But we are all considered intelligent. We have the capacity to learn about corporations, politics or sports. In addition, we all have some kind of common-sense (the kind that is common to all of us), i.e. we can all tell that a flying bird does not touch the ground, that after I enter a room I am inside the room, if I was able to enter, then probably the room was bigger than me etc. etc. We will all be able to answer a question such as ``Can a lone king reach every square of an empty chess board?'' after given a brief description of the rules. We will all have a hard time answering the same question for a knight [McAllester, 1991].
An argument can be made for this kind of common-sense reflecting the infrastructure for intelligence. An argument cannot be made for knowledge about large corporation internals being a necessary prerequisite for intelligence. That is static encyclopedic knowledge. If Cyc ever reaches the ``crossover point'', it can read about these things later. I presume that the pressure from the large corporations which were supporting the Cyc project forced the team to focus on such issues prematurely. It is unfortunate that this was at the expense of handling the infrastructure in an adequate way.