perm filename AI.ESS[ESS,JMC] blob
sn#851443 filedate 1988-01-11 generic text, type C, neo UTF8
COMMENT ā VALID 00002 PAGES
C REC PAGE DESCRIPTION
C00001 00001
C00002 00002 THE PRESENT SCIENTIFIC SITUATION IN ARTIFICIAL INTELLIGENCE
C00009 ENDMK
Cā;
THE PRESENT SCIENTIFIC SITUATION IN ARTIFICIAL INTELLIGENCE
The purpose of the present memorandum is to tell how I see the
present situation in AI independent of the problem of support. Any policy
for government (in particular ARPA) support of AI will ultimately be
disappointing if it doesn't take into account the actual scientific
situation.
1. It will ultimately be possible to make artificial intelligences
capable of any intellectual task humans can do, but doing it faster, more
reliably, and taking into account more information. The evidence on this
point has not changed a great deal in the last twenty years. Our
confidence in it is partly based on the general principle that
intelligence is a mechanism and that it is possible to understand and
implement on a computer any mechanism. Besides that there has been
considerable success in implementing a number of mechanisms we have been
able to understand.
2. It will be possible to prevent artificial intelligence from
getting out of hand if people want to do so. There will be no advantage
to anyone in producing an uncontrollable artificial intelligence even
under conditions of human conflict. Therefore, while the potential use of
artificial intelligence in human conflict, like the use of any other
technology has dangers, it is less dangerous than nuclear or biological
technology in this respect. The main danger is its successful use by the
bad guys in some conflict situation.
There is little that can or should be done about this now, because
technology for making sure that artificial intelligence doesn't get out of
hand should be based on increased understanding of the mechanisms of
intelligence.
3. Understanding intelligence remains an extremely difficult
scientific problem. No-one understands it well enough to define a
development program with a substantial probability of producing high
general intelligence in a specific time. The difficulty is probably of
the same magnitude of that of producing an artificial animal using the DNA
genetic mechanisms that will have specified properties and breed true.
(These problems are not especially related; I merely wished to give the
reader a way of calibrating my estimate of the difficulty).
4. Progress in the theory of artificial intelligence has been
disappointingly slow, but the situation is no worse than in other slow
scientific fields like genetics before the DNA breakthrough or embryology
or the present situation in high energy physics.
In my opinion, this is mainly because wrong or inadequate
methodological ideas have prevented people who would like to do theory
from working on the right problems. I still believe that the road to
success lies in separating the epistemological from the heuristic problems
and concentrating initially on the former. This is substantially as
proposed in my papers of 1959, 1963, and 1970. Regrettably, I haven't
convinced many others to work in this direction, but I think the situation
is improving.
Specifically, I think that much present work is misdirected, but I
do not advocate that this meeting or ARPA take a position on these
scientific issues. Rather they should be settled in the traditional way
issues are settled in the scientific community.
5. The problem of vision is mainly separable from the main problem
of artificial intelligence. It is also a major scientific problem, and a
development program to reach human or primate capability cannot promise
success at present.
6. The present state of vision research while not within
development range of human capability, is probably within development
range of capability of useful industrial, consumer (the car), and military
capability. It may be within development range of cost-effective
capability in some of these areas.
7. A number of other applications of AI to special classes of
problems are also within development range of cost-effective capability.