OPINIONS


1. The Japanese plan requires the solution on schedule of some
fundamental conceptual problems of artificial intelligence.
Therefore, it seems unlikely that the system they develop
will attain the form the project on the time scale they
project.  Of course, the 5th generation project may still produce
worthwhile results.

2. Artificial intelligence is not ready for a Manhattan Project.
The equivalent of nuclear fission has not yet been discovered.

3. The Middle East war shows that technological advantage is
militarily more important than it has been in the past.
Better electronics and better ways of using it allowed the
Israelis to destroy planes and missiles much more effectively
than in 1973.  I apologize for mentioning an opinion in an
area in which I am not expert.

4. Artificial intelligence has the potential for large technological
surprises.  While IBM can gamble that it can readily hire the people
to take advantage of any commercial AI breakthroughs that develop,
the Defense Department must be prepared for events that take place
on a much shorter time scale.

5. The DARPA Speech Project is not a model that should be followed
for AI in general.  While it had some success, for close to seven
years, almost all research in computer speech recognition in the
U.S. was oriented towards selling concepts to the committee in charge
of the Speech Project.  While it was probably the most competent
committee that could be formed, it seems likely to me that the amount
of money spent would have achieved greater results if a more
diverse collection of approaches had been followed.

6. Basic research in AI has been somewhat neglected.  This is partly
the fault of the funding agencies who have responded best to proposals
that promise (but usually don't deliver) a short term applied payoff.
However, this also corresponds to the ideology of the majority of
AI researchers, which I believe to be mistaken.  In fact a greater
diversity of ideas were being pursued at the end of the 1950s than
are being pursued today.  DARPA and other agencies need to encourage
basic research in AI - in heuristic search, in representation, in
epistemology and in reasoning.

7. Perhaps the most obvious potential artificial intelligence
military application is to much smarter weapons.  It is potentially
possible for a weapon that fires many shots per second to aim each
one of them at a different target instead of merely bombarding an
area.

8. Another issue is the rapid controllability and modifiability of
software.  Surely much software will benefit from modification
on the basis of information received in the first few days of
conflict.  The NASA Apollo experience is instructive.  Their original
intention was that all the Lunar Module software would be in
read-only memory for greater reliability.  Fortunately, as I understand
it, the read-only memory never worked and they had to use ordinary
core memory, because on the first landing they had to modify the
software to ignore a faulty switch.  Because of their lack of
preparation for modifying software, they had to radio the
modifications in octal to the astronauts, who had to put them in
by hand.  Fortunately, the required modifications were small.

	One of the main characteristics of AI languages, including
LISP and the higher level languages built on top of it, is that
software is readily modifiable and that programs to modify software
are easily written, because the language has access to its own
syntax.  Unfortunately, this is not true of ADA and its relatives
such as Pascal and Algol.  It will be important to adapt the AI
languages for use in "imbedded software".