the same information in several different forms for different purposes.
One advantage of storing genetic information in the modular and data-like
form of DNA is that two individuals' genes can be easily recombined to
form a new genotype. This would be very difficult if the information were
only in proteins. A second reason for storing information in DNA is that it
is easy to transcribe and translate it into proteins. When it is not needed, it
does not take up much room; when it is needed, it serves as a template.
There is no mechanism for copying one protein off of another; their
folded tertiary structures would make copying highly unwieldy. Com-
plementarily, it is almost imperative to be able to get genetic information
into three-dimensional structures such as enzymes, because the recognition
and manipulation of molecules is by its nature a three-dimensional opera-
tion. Thus the argument for purely procedural representations is seen to
be quite fallacious in the context of cells. It suggests that there are advan-
tages to being able to switch back and forth between procedural and
declarative representations. This is probably true also in AI.
This issue was raised by Francis Crick in a conference on communica-
tion with extraterrestrial intelligence:
We see on Earth that there are two molecules, one of which is good for
replication [DNA] and one of which is good for action [proteins]. Is it
possible to devise a system in which one molecule does both jobs, or are there
perhaps strong arguments, from systems analysis, which might suggest (if
they exist) that to divide the job into two gives a great advantage? This is a
question to which I do not know the answer.^14Modularity of Knowledge
Another question which comes up in the representation of knowledge is
modularity. How easy is it to insert new knowledge? How easy is it to revise
old knowledge? How modular are books? It all depends. If from a tightly
structured book with many cross-references a single chapter is removed,
the rest of the book may become virtually incomprehensible. It is like trying
to pull a single strand out of a spider web-you ruin the whole in doing so.
On the other hand, some books are quite modular, having independent
chapters.
Consider a straightforward theorem-generating program which uses
TNT's axioms and rules of inference. The "knowledge" of such a program
has two aspects. It resides implicitly in the axioms and rules, and explicitly
in the body of theorems which have so far been produced. Depending on
which way you look at the knowledge, you will see it either as modular or as
spread all around and completely nonmodular. For instance, suppose you
had written such a program but had forgotten to include TNT's Axiom 1 in
the list of axioms. After the program had done many thousands of deriva-
tions, you realized your oversight, and inserted the new axiom. The fact
that you can do so in a trice shows that the system's implicit knowledge is
modular; but the new axiom's contribution to the explicit knowledge of the
system will only be reflected after a long time-after its effects have "dif-Artificial Intelligence: Retrospects^617