ArticlePDF Available

The dark ages of AI: A panel discussion at AAAI-84

  • Journalist
The Dark Ages of AI:
A Panel Discussion at AAAI-84
Drew McDermott
Yale University, New Haven, Comecticut 06520
M. Mitchell Waldrop
Science Magazine
1515 Massachusetts Avenue NW, Washington, D C. 20005
Roger Schank
Yale University, New Haven, Connecticut 06520
B. Chandrasekaran
Computer and Informatiollal Science Department, Ohio State Ufziversify, Columbus, Ohio 43210
John McDermott
Coqmter Scieme, Carnegie-Melloll Ulziversity, Pittsburgh, Penmylvania 15213
Drew McDermott:
In spite of all the commercial hustle and bustle around AI
these days, there’s a mood that I’m sure many of you are
familiar with of deep unease among AI researchers who
have been around more than the last four years or so.
This unease is due to the worry that perhaps expectations
about AI are too high, and that this will eventually result
in disaster.
To sketch a worst case scenario, suppose that five years
from now the strategic computing initiative collapses mis-
erably as autonomous vehicles fail to roll. The fifth gen-
eration turns out not to go anywhere, and the Japanese
government immediately gets out of computing. Every
startup company fails. Texas Instruments and Schlumber-
ger and all other companies lose interest. And there’s a
big backlash so that you can’t get money for anything con-
nected with AI. Everybody hurriedly changes the names of
their research projects to something else. This condition,
called the “AI Winter” by some, prompted someone to
ask me if “nuclear winter” were the situation where fund-
ing is cut off for nuclear weapons. So that’s the worst case
I don’t think this scenario is very likely to happen,
nor even a milder version of it. But there is nervousness,
and I think it is important that, we take steps to make sure
the “AI Winter” doesn’t happen-by disciplining ourselves
and educating the public.
This panel has been assembled to discuss these issues.
I’ve asked the panelists to discuss the following questions
in particular: Are expectations too high among consumers
of AI, such as business and military? If they are too high,
then why? Is there something we can do to change this
mismatch between expectation and reality? To what ex-
tent is this mismatch our fault? There’s a charge often
leveled against AI people that they claim too much. To
122 THE AI MAGAZINE Fall, 1985
what extent is it due to naivet6 on the part of the public?
What is the role of the press in this mismatch, and how
can we help to make the press a better channel of com-
munication with the public? What is the role of funding
agencies in the future going to be as far as keeping a realis-
tic attitude toward AI? Can we expect DARPA and ICOT
to be stabilizing forces, or is there a danger that they may
cause people in government and business to get a little
bit too excited? Are funding agencies going to continue
to fund pure research, even if AI becomes a commercial
success? Will the perception remain that we need to do
some things that are not of immediate commercial inter-
est? And, finally, what should each of us do to insure his
survival in case of problems?
Here to discuss these issues are Mitch Waldrop, from
Sczence Magazine, representing the press; Ron Ohlander,
from DARPA, representing a funding agency; Roger
Schank, from Yale University; B. Chandrasekaran, from
Ohio State; and John McDermott, from Carnegie-Mellon
University. The first speaker will be Mitch Waldrop.
Mitch Waldrop:
First, I would like to relate an experience I had earlier
this week when I was attending a seminar in New York
state that Isaac Asimov organizes every year. This year
the topic was Artificial Intelligence, and the idea was to
bring in people from all walks of life and, over the course
of several days, work up a human impact statement for
artificial intelligence. Marvin Minsky, as well as myself
and others, were on the resource panel. The result was
what you might expect: A combination of silliness and
seriousness, with not a great deal of informed insight into
AI. But there was a very good cross-section of the general
public, and I gained some very interesting insights while
trying to answer their questions.
AI Magazine Volume 6 Number 3 (1985) (© AAAI)
One, is that most of these people make essentially no
distinction between computers, broadly defined, and arti-
ficial intelligence-probably for very good reason. As far
as they’re concerned, there is no difference; they’re just
worried about the impact of very capable, smart comput-
ers. Enthusiasm and exaggerated expectations were very
much in evidence. The computer seems to be a mythic
emblem for a bright, high-tech future that is going to
make our lives so much easier. But it was interesting to
hear the subjects that people were interested in. Educa-
tion seemed to capture their imagination most-computer-
aided instruction potential. If you want to see some real
passion, start talking about what happens to people’s kids
in their school-room environment.
This was followed by an absolute fascination with cog-
nitive science and what artificial intelligence is telling us
about how we think. As for applications to health-they
were a little vague beyond potential for diagnosis. In fact,
they didn’t make much distinction between artificial intel-
ligence and biotechnology.
There was even some interest in the possibilities of
what could be done with very large databases, searching
it with intelligent database searchers, etc.
I’m not sure what it means, but it’s interesting that
this seems to be in roughly the inverse priority to what AI
people give these subjects.
What really struck me was the flip side of exagger-
ated expectations-exaggerated fears. The computer is
not only a mythic emblem for this bright, high-technology
future, it’s a mythic symbol for much of the anxiety that
people have about their own society. The most obvious,
what you might call the “1984 Big Brother Is Watching
Anxiety,” is that somehow the computer will erode our
freedom and invade our privacy. Who writes the computer-
aided instruction programs for our children? They control
what our children learn, and how they think. There seems
to be an implicit assumption that there’s always going to
be some big, manipulative power structure up there con-
trolling things.
A second anxiety, what you might call the “Franken-
stein Anxiety,” is the fear of being replaced, of becoming
superfluous, of being out of a job, and out on the street.
A third, closely related anxiety might be called the
“Modern Times Anxiety.” People becoming somehow, be-
cause of computers, just a cog in the vast, faceless machine;
the strong sense of helplessness, that we really have no con-
trol over our lives, that computers, being inevitably very
rigid, brittle machines, becoming more and more powerful,
inevitably result in alienation, isolation, enforced confor-
mity, standardization, and all those bad things-leaching
away of humanity. I’m going to come back to this, but I’ll
leave it right now by saying that these fears and expecta-
tions are not groundless.
That brings me to the news and the media. My col-
leagues in the news and media have heard me rant about
imbecile reporters and relentlessly shallow TV reporters,
and I’m not going to give that speech here. It would be
But I bring up the general public’s attitude to point
out that reporters, editors, and TV people, are human
beings; they are reflections of the society in which they live
They write about or film what their readers are interested
in and what they are interested in; what seem to them to
be important issues. That brings us to the key problem in
covering something like AI. The problem is not a matter
of imminent deadlines or lack of space or lack of time or
people straining for “gee whiz” or “Oh, my God” type
headlines. The real problem is that what reporters see
as real issues in the world are very different from what
the AI community sees as real issues, and the trick is to
bring these into consonance. Where does that leave the AI
community? There seems to be broad agreement here that
the coverage of AI is abysmal. So what do you do about
it? Something that is very unhelpful is to take an attitude
that everybody out there is a pack of idiots except us, who
really understand. Cheap shots at reporters get big laughs
at meetings like this but are not very helpful. Some of us
HAVE taken predicate calculus and can understand it very
What would be useful is to ask yourselves: If you don’t
like the coverage that you’re getting now, what would good
coverage consist of? What would be a good story about
AI? What would you like to see? AI is about giving opera-
tional definitions for things. What’s an operational dehni-
tion of a good story? I’m not going to attempt to answer,
but it might be helpful if you think about that.
When a reporter comes to talk to you, what message
are you trying to communicate? It might be helpful to
know ahead of time. If what you want to communicate is
the latest stuff on some nonmonotonic, backward chain-
ing, I don’t think it is going to be too compelling to the
reporter. If it is about the nifty expert system that you
hope to be marketing next month, it’s probably going to
look very funny to the reporter. He would be very suspi-
cious. But just in general, ask yourself what it is you are
trying to communicate. Think of it as an opportunity, not
as an interruption and an irritant.
The idiot reporters and the insensitive editors are al-
ways going to be worthless. I cannot offer you any wis-
dom or magic formula to make them go away. The best
you can hope to do is help the conscientious reporters who
do come by and may just be confused or not know, but
are genuinely willing to try to learn as much as they can
given their time constraints. These people do exist. And
as I said, it helps to have a clear idea of what you want to
A final point is a modest suggestion to the commu-
nity to get out in front of this Frankenstein issue. We’re
going to hear a lot on this panel about overheated expecta-
THE AI MAGAZINE Fall, 1985 123
tions. I think the fears are just as strong, and they’re out
there in the general public. There’s no point is dismiss-
ing them as neurotic or misled. They’re there, no matter
what the source. There’s no point in saying it is not AI’s
problem, that it’s robots who are kicking people out of
the factories. Well, you guys are designing vision systems
for robots. People are going to be thrown out of work by
various forms of expert advisors, at least temporarily, and
people are not going to like it when their careers have been
dashed by some superanimated VisiCalc. Governments can
use large databases to violate people’s privacy and to ha-
rass them. For that matter, credit card companies can do
that, too. One can even envision a natural language sys-
tem that monitors telephone conversations. Computers
and even AI programs can be made to be rigid and con-
straining. There’s no point in saying that you don’t need
computers to do all this. It is true that Hitler managed
to create a totalitarian society with no help from comput-
ers. But computers can also be used to exacerbate these
tendencies and a power drive for human stupidity. If we
expect physicists to be concerned about arms control and
chemists to be concerned about toxic waste, it’s probably
reasonable to expect AI people to be concerned about the
human impact of these technologies.
Just to make sure I am not misunderstood: I’m not
advocating that people run around spouting leftist rhetoric
or crying “chicken little.” I am suggesting some sober and
constructive thought about how one can order, how one
can address these problems over the long run, and that the
community as a whole take some kind of position. After
all, computers, especially with the aid of artificial intelli-
gence, can be made extremely flexible, extremely person-
alized, and can produce a great deal of wealth-we hope.
I’ll conclude by saying that perhaps if people had some-
thing substantive to say on these issues, reporters would
not have to strain for “Gee whiz” or “Oh, my God” head-
Roger Schank:
I’m sorry, I’m not representing the business interests to-
day. I hate to disappoint you. But in line with that, I’ll
tell you that I have a new company that does educational
software. I mention it because I was having a conversation
with Oliver Seldridge, telling him about my new educa-
tional software company. I said, “Well, it really doesn’t
have anything to do with AI at all, except that some of
the software we design has to do with things like teaching
reading and reasoning comes from ideas that we’ve had in
AI, but there’s no AI in the programs in any way.” And
he said, “Oh, sort of like expert systems, huh?”
I came here to relay to you six conversations. That’s
the first. They’re all short.
The second conversation I had was with a real estate
developer, who had a Ph.D in Biology. He wanted to build
an expert systems industrial park; every company in it
would be doing expert systems. I said: “You may have
come to the wrong person; I don’t much believe in ex-
pert systems. “How can you say that?” he said. I asked,
“What do you mean.
7” He explained, “Well, to get comput-
ers to model everything that somebody knows; to put all
the knowledge in and have the thing be just like a person-
that’s terrific.” I replied, “Yes. But we don’t know how to
do that yet.” He said, “No, but that’s what expert systems
are.” That’s the second conversation.
The third conversation was with Bob Wilensky, a for-
mer student of mine, who asked me what I was going to
do on this panel. He asked if I thought doomsday was
coming. I said, “Yes.” And he said, “No, you’re wrong.” I
asked why. He said, ‘It’s already here. There’s no content
in this conference.” Now I think there’s something seri-
ous to be concerned about there. He isn’t the only person
I’ve heard express that view. If that’s true and there’s no
content in this conference, then doomsday is already here.
Conversation four was with a representative of ARPA-
not Mr. Olander-and he said, ‘LYou know, we’ve got a real
problem. We’ve got so much money to spend on scientific
research that we have more money than there are scien-
tific researchers.” And I replied, “And we’re not going to
be able to fix that, are we?” It’s very hard to make more
scientific researchers in an environment where money is be-
ing offered in tremendous amounts to be developers rather
than to be researchers. ARPA can’t raise salaries. It can
only offer money to hire people.
The fifth conversation isn’t a conversation. It’s just a
report. It’s a report of what I heard was the standard for
accepting papers to this conference this year. I’ve been
on the program committee a few times, but missed it this
year. What I heard was that only completed scientific work
was going to be accepted. This is a horrible concept-
no new unformed ideas, no incremental work building on
previous work. I don’t know if that’s actually what hap-
pened. I didn’t attend a lot of sessions. I can tell you
that if that is what happened, that’s frightening. Remem-
ber AI? See, you guys may not be as old as I am AI-wise.
But I remember AI, the first conference and the second
conference. We used to sit and argue about things-not
whether or not we should go public. There were always
people in AI who were interested in software development
tricks. That’s great. There has always been that compo-
nent in AI and there always should be. But if it comes to
dominate AI, then AI becomes applied systems. I don’t
like that.
The last conversation was with Eugene Charniak, and
I had it ten years ago. He kept saying to me, “Roger,
you’re promising too much. You can’t do all the things
you think you want to do. They’re very hard.” And I said,
“Yes, but they’re fun, and I want to work on them. And,
anyway, I think I can do them in ten years.” Here’s why
I mention this. It’s been ten years. I haven’t done them.
Gene Charniak is a wise man. I may not be able to do
124 THE AI MAGAZINE Fall, 1985
the things I thought I could do ten years ago in fifty years.
Yet, at the same time, as I’m beginning to discover more
and more problems about why things are hard, we are get-
ting less and less people working on those hard problems
and more and more people working on applied situations.
That’s frightening.
What do I think the issues are here? First, I think
the press is completely and utterly irrelevant. When I
first got into this field twenty years ago, I used to explain
to people what I did, and they would say, “You mean
computers can’t do that already?” They’ll always believe
that. And it doesn’t matter what the press believes, and it
doesn’t matter what the general public reads in Time and
Newsweek. It really doesn’t make any difference. We have
responsible reporters in Science magazine, but not that
many people read Science magazine in the general public.
I don’t think it’s an issue what the general public believes.
However, it is a very important issue what big business
believes. You see, big business has a very serious role in
this country. Among other things, they get to determine
what’s “in” and what’s “out” in the government.
I got scared and started a company at the same time,
when there were lots of startup companies around. I got
scared when big business started getting into this- Schlum-
berger, Xerox, Hewlett-Packard, Texas Instruments, GTE,
Amico, Exxon,-they were all making investments- they
all have AI group. You start to wonder who could be in the
AI groups. We haven’t got that many people in AI. And
you find out that, those people weren’t trained in AI. They
read an AI book, in many of these cases. They started off
reading all the best AI research. After a while you discover
AI group after AI group whose people were only periph-
erally in AI in the first place. What’s going to happen
is that those companies will find that their groups aren’t
producing as well as they had expected. When they find
that, they will complain; they will say nasty things about
AI. The presidents of those companies will be talking to
the people who are not at ARPA, but at the Secretary of
Defense level. They’ll say things like, “Well, I’ve spent so
many millions of dollars on AI this year, and I’ve had it.
They’re not producing anything.” And it may be that AI
is capable of producing things. It may be that even the
people at those companies are good. But it also may be
that it’ll take them more years than anyone expected.
I’m very concerned about this issue. It’s the reason
I’m on this panel. And I think it’s very important for
people not to worry about the press. Talk to the press;
they’re nice. It doesn’t hurt.
The thing to worry about is when you hear that a
company is starting up an AI effort, you better ask: When
do you expect what? The small companies, the startup
companies, that have been started by AI people, don’t
present the same problem. We’ve had to learn how to
build a product. The more we learn about products, the
more we begin to realize that our products look less and
less like AI. That’s okay-that’s what derivatives are like.
It’s okay to build derivative things. What’s not okay is
to build only derivative things. That’s frightening. So I
am concerned not that expectations are too high, but that
expectations are too low. What expectations am I talking
about? The expectations that we, as an AI community-
I assume that the people left here at this conference are
the actual AI community-have forgotten that we are here
to do science, and that we are nowhere near the solution.
We used to sit and fight about these things in public; now
we all sit and talk about how it’s all solved, and we give
slick talks with slides with pretty pictures on them. I’m
very concerned about the fact that people don’t want to
do science anymore-it is the least appealing job on the
market right now.
It’s easier to go into a startup company and build
products. It’s easier to go into a big company and have a
little respite and do some contract work. It’s easier to do
all those things than to go into a university and try and
organize an AI lab, which is just as hard to do now as it
ever was, and sit there on your own trying to do science.
It’s difficult. But if we don’t do that, we will find that we
are in the “dark ages” of AI.
I take this opportunity to talk to any of you who are
considering the choice between the two and recommend
that you seriously consider that the science of AI is criti-
cally important and also, by the way, a lot of fun.
I leave you with two messages, which will be obvious,
I hope. The first, from one half of my life, is incumbent
upon AI because we have promised so much, to produce.
We must produce working systems. Some of you must
devote yourselves to doing that. And part of me devotes
myself to doing that. It is also the case that some of you
had better commit to doing science. Part of me commits to
doing that. And if it turns out that our AI conference isn’t
the place to discuss science, then we better start finding
a place where we can discuss science, because this show
for all the venture capitalists is very nice. And I hope
all the people back there sell more computers and more
systems, and they should all live and be well. But I am
concerned that people here who are first entering this field
will begin to believe that a Ph.D means building another
expert system. They’re wrong.
Ron Ohlander:
They should always put Roger Schank on last because no-
body can top him.
I’d like to go back to the possible scenarios that Drew
McDermott outlined. I think there are some other possible
scenarios. I’ll mention them, and what the government is
doing concerning them.
One possible scenario is that the current upsurge of
development continues unabated, and we have most re-
markable development going on for the next few years.
Another possible one is that in the near future we’re going
THE AI MAGAZINE Fall, 1985 125
to have some shakeout and realignment. A third one is
we’ll have some disillusionment with the process. There
will be some retrenchment. One way of characterizing it is
we would possibly go back to the state of affairs that were
in existence four or five years ago. Finally there is the last
scenario that McDermott described. And, of course, there
are all the things in between.
I will leave those scenarios for now and discuss some of
the things that are happening in government and how they
portend for the possible downside of the AI technology.
The government has increased its interest in this par-
ticular area. DARPA, has been involved for quite a num-
ber of years in AI research. But the other services are
also starting research efforts. DARPA is continuing its
basic AI program, and it’s still funded at the same level
with some sliding piece over the next few years. In addi-
tion, there’s the strategic computing program, which was
described to you this morning. The Army has started a
support program for an AI laboratory to be placed at a
univeristy. The Navy has established an AI lab at NRL.
The Air Force has started to put together an AI consor-
tium at the universities to support them in research and
development and in education and training. In addition,
the intelligence community is taking a long hard look at
AI and what it offers to their endeavors.
What I conclude from this is that there’s a strong,
healthy interest in AI, that there’s not a lot of wild-eyed
people out there that expect things that are beyond the
state of the art, and that most people are taking a very
orderly approach to the exploitation of AI within the gov-
ernment. The people that are heading these efforts are
pretty level-headed-they know what is going on. I think
that government funding is likely to continue. It’s hard to
get started, but once it gets started, it’s likely to continue
for a period of time.
In addition to these efforts, there are a number of
things going on in terms of system development. This
gets into the actual exploitation of AI for particular sys-
tems. In case anyone thinks that the situation of AI is
unique, let me tell you I have worked for three or so years
at Naval Electronic Systems Command, and I looked at
practically every development that went through there in-
volving computer technology, and every one of them had
trouble. The fact was there were a lot of failures. There
were overruns and systems delivered past schedule. This
is certainly not unique to Naval Electronic System Com-
mand. The most would be systems being acquired for the
government. The government contitiued to acquire such
systems simply because there were also a lot of successes-
key successes. This kind of success encouraged continued
interest and development in the field.
The fact is, there aren’t enough good people to go
around, so the government and everybody else are forced
to pay high prices for successful systems. These high prices
not only come for particular developments, but they come
for multiple efforts to get the same development because
of the failures that occur. I think the government interest
will continue, and that we will see the systems as develop-
ment contractors do more work and get more interested;
they will make representations for the government for in-
corporation of a lot of the technology into these systems.
The implementations will occur and some successes will
also occur. There will also be some failures, but I think
we’ll live through it.
I’d like to make one other comment. What kind of
steps can be taken to circumvent the dark ages or to de-
crease the impact of the downside of the current interest
in AI? I’m not sure that a lot can be done. We’re pretty
much riding the whirlwind.
There are a few conclusions I would like to draw from
my observations of what’s happening in the field; my ob-
servations in funding various efforts. Let me repeat: I
think there will be a lot of failure. On the other hand, I
think there will be enough key successes to override the
failures to keep interest focussed on the area. The shake-
out will come in the not too distant future. There will be
some shakeout in machine areas and some shakeout in all
the expert system technology companies. I think the gov-
ernment will sustain their funding. In fact, it’s likely to
increase as people get more interested in applications. And
in some rebuttal to Schank, I think there’s room for both
scientists and engineers in this field. What we’re seeing
is the rapid growth of an industry that has no underlying
engineering support, so scientists, who would otherwise bc
doing research, are filling that role. It is also my observa-
tion, as it’s been Schank’s, that people currently getting
involved in industry are coming from in-house assets, peo-
ple who are being trained. I see nothing wrong with that
because I think there is a role for people to take this scien-
tific technology and to implement it from an engineering
So revisiting those scenarios that I postulated earlier,
I think that basically the interest will continue, and we’ll
have some shakeup and realignment. In summary, my out-
look for the future is rather positive.
B. Chandrasekaran:
As one of the few academics who is not in modern business,
I have been assigned the role of survivalist. First I want to
ask, “Has AI paid its way?” . . . Or to put it another way,
“Have we earned our keep?” I have three answers to that:
Yes, yes, and yes.
It’s been the most profound paradigmatic change, the
most performed paradigmatic change in human understand-
ing of some of the important issues about ourselves in a
long long time.
The notion of cognition as computation is going to
have extraordinary importance to the philosophy and psy-
chology of the next generation. And for well or ill, this
notion has affected some of the deepest aspects of our self-
126 THE AI MAGAZINE Fall, 1985
image. I think it’s for well, but we’ll have to see. Even at
the technology level, what we have been able to do is min-
imal if you measure it against the capability of the human
mind. But if you measure it against changes in the styles
of programming or styles of building things, significantly
useful accomplishments have been made. The problem is
not that AI is weak in terms of its usefulness or impor-
tance. AI has been the whole expectation of the problem.
When people start building complicated things, there
is a remarkable consensus of what they are looking for.
The reason for that is already there is a software archi-
tecture based on frames; the embedded procedures and
moving around in that kind of space constitute a weak
theory of mind, but nevertheless a theory of mental archi-
tecture. The next is very minimal commitment, which is
for the good, because we don’t know much more at this
point to make strong commitments. In that sense it is al-
ready helpful and will enable us to build entities that we
wouldn’t have been able to build otherwise.
The next question to ask is “Has there been a discon-
tinuity in AI to justify the sudden interest from whatever
viewpoint?” Yes, there has been discontinuity, and the dis-
continuity has been in the idea that knowledge is very im-
portant. We can’t separate knowledge as somethiug that
is pragmatic and go away and do theory that is not con-
nected with knowledge. So in that sense, the discontinuity
that has caused this interest is the importance of knowl-
edge. It’s true in natural language understanding. Roger
Schank and his group’s work and several others have em-
phasized that. Problem solving has also been important.
The problem has been, however, that we have very weak
theories of knowledge and even weaker theories of how to
use them. People don’t understand that very well. So
they’re being mistaken and misunderstood with respect to
their power. Most importantly, there have been no char-
acterizations of what is possible. So the real problem is
the very, very strong belief in omnipotence of simple ar-
There is also a real confusion because the computer
science community, the AI community, has refugees from
so many areas. Symbol-level theories, which may even
be right, are being mistaken for knowledge-level theories.
This is one of the conceptual problems that has been be-
deviling us.
So, basically, the problem isn’t underestimation of the
problems of multiplicity of generic knowledge structures.
Knowledge use invalidities has been the cause of misun-
derstanding. We’re seeing some systems and extrapolating
that all it takes is more of the same. I also believe that
faster architecture could do the trick. Let’s run 30,000
rules. Let’s run 50,000 frames. The idea being that all it
takes is more architecture, faster systems.
We need to characterize the things that we can do-
that’s do-able. Then spin off a list of things that we know
how to do and let people go ahead and do them. We could
also more clearly understand the research issues that need
Regarding the commercial prospects; that may no long-
er be a very interesting problem from an AI viewpoint, but
it would not have been possible without AI having been
there. This class of problems can be categorized as knowl-
edge-rich and problem solving-poor. All that is needed is
getting some form of knowledge, organizing it, and making
it available to people. It’s going to have very little prob-
lem solving capability, but without the recent history of
AI, such a thing would not have been possible. But it’s
not going to solve all the problems. A reasonable num
ber of problems can be handled with this appoach. The
formula I normally use, 10% AI and 90% other, is what
is going to make it useful. In fact, this might even spin
off, and in some ways it may even be better off for AI if it
dots. Then AI can get back to concentrating on research
The problem in applying even current technology is
that they still require epistemic analysis. This is very
hard. To think that all it takes is engineering is a mis-
taken uotion.
Epistemic analysis is hard to teach. Some people can
use the same tools and build extremely interesting systems;
other people cannot. So AI is going to be blamed for
the failures of people who do not have the capabilities of
sufficiently powerful epistemic analysis. Also, AI is going
to be blamed for what I call dilution of AI. We started from
AI, which then became expert systems, which then became
rules, which then became LISP. So people think they are
doing AI when they do LISP programs, for example. It is
important to keep reemphasizing that AI is not all those
things. AI is something else. We must keep emphasizing
the importance of all those “something elses.”
Will there be an AI Winter? I think there may be
an AI dusk, which may be even better, rather than the
hard sunshine that we have been having in Austin recently.
Instead of the bubble bursting, the bubble may become
somewhat smaller and less fragile. So it may be an in-
teresting place to be. I think AI has already contributed
enough, and I believe that AI will contribute enough to
justify itself at that level.
There are all sorts of historical analogies about what’s
going on in AI. One relates to the automatic high-quality
natural language translation. I don’t think that analogy
is valid. I think that it was based on too small a number
of ideas. It didn’t have enough robustness and solidity
to it. AI today has more robustness and solidity to it iu
so many levels that that’s not the real analogy. A truer
analogy is probably closer to biotech, where even five years
ago people thought incredible claims were being made.
But biotech companies have not gone bust. Many
them have gone back to solid research. Even five years ago,
AI people used to go around talking about how they’re go-
ing to clone the human miud. That used to get me really
THE AI MAGAZINE Fall, 1985 127
scared because that showed that big theories were being
mistaken for strong theories.
My hope is that AI will evolve more like biotech in
the sense that certain technologies wil get spun off, and
researchers will remain and extremely interesting progress
will be made.
With respect to projects such as the Japanese fifth
generation, there is a nightmare of some bureaucrat at
the DARPA finally taking all those things and stamping
“Failure, “Failure, “Failure,” on each one of them. Our
position ought to be to plan for success and to realize that
in many ways we cannot fail. If we have some personal
and professional integrity in what we do, we cannot fail
to come out ahead. Even at a technological level, enough
things will be happening for DARPA to get its money back.
That’s the sort of attitude to encourage. That requires not
pushing weak theories too far.
What am I doing as a surviror? I believe the next four
or five years are going to be some of the most exciting years
in research. In spite of all the over-promising and over-
expectations, the last five years have identified extremely
important and interesting problems. I hope to be involved
in that kind of research. Also, unlike some other areas, we
don’t have to decouple technology from research.. . . You
can build things and still do research. In that respect,
AI is in a better situation than other so-called theoretical
sciences. As long as we characterize each of the advances in
knowledge-level terms, identify what it’s capable of doing,
and identify what it’s not capable of doing so that users
know what kind of problems can be solved, I think we
can come out in reasonably good shape at the end of this
John McDermott:
I want to revisit Drew McDermott’s original thinking on
why this panel is a good idea. His position is that there
are these things called expectations. The people who have
false expectations are going to become angry, upset, or
unhappy with the violation of their expectations, and walk
away. That applies to our situation, if the people of the
world find out that AI isn’t what it expected, then AI
isn’t going to be supported anymore. Presumably, one of
the kinds of support that we need most is funding for our
research efforts. So that’s going to go away.
Schank offered a slight variant of that, which is that
we don’t have to worry about what the people think, but
we do have to worry about what big business thinks. If
big business gets angry at us, then the funding source will
be cut off, and we won’t be able to do science.
I think that we all agree that we would like to have
the science of AI continue to be supported. I don’t think
anybody, either up here or out there, is at all unclear on
the fact that we haven’t yet made much progress in the
field, and that we’ve got a long way to go, and most of the
exciting discoveries are still ahead of us. So we somehow
want to insure that the funding base for the science of AI
continues to be there so we can do good research.
What bothers me about Drew McDermott’s premise
is that it’s not completely clear to me that people who
have violated expectations are going to end up withdraw-
ing support. If we look at the kinds of expectations that
could be incorrect, there are presumably many. I have
four that I would like to go through. If you focus on dif-
ferent types of expectations and the ways those could be
violated, it’s hard for me to see a clear connection between
violating those expectations and not having research get
The first kind of expectation that could be off would
be an expectation about the kinds of tasks that current AI
systems or AI systems five years from now, would perform.
I have encountered people who have a science fiction view
of the world and think that computers now can do just
about anything. But that view never seems to manifest
itself in a personal way. These people have a feeling that
computers can do wonderful things, but if you ask them
how exactly could an AI program help in work, they don’t
have the sense that within a week or two they could be
replaced or that computers can come in and do a much
better job than they do in their work. So I don’t think
that at a concrete level, people have a naive view that
super-intelligent computer programs are right around the
corner. In fact, I find the opposite. When people talk
with me about systems that could be developed to solve
particular problems, I’m often the one that describes a
more positive or grandiose role for the systems than they
do. The people who might fund these efforts end up having
what seems to be an extremely healthy caution in what
they expect. So I don’t see a lot of wildness in people’s
ideas of what systems can do. Even if there are some
people who do have overblown expectations along those
lines, I think that the AI technology has developed to an
extent now that it is possible to produce software that can
do some extremely helpful things. Because of that, people
are going to be happy. They might not be as happy as
they would be if their wildest dreams were satisfied, but
the slope is upward, and I think they’ll like it.
So if someone’s expectations are violated because he
or she gets some good, helpful thing, but it’s not as good
and helpful as hoped for, I don’t think that’s the kind of
expectation violation that’s going to result in the funding
rug being pulled out from under us.
The second kind of expectation has to do with the
background people have to have in order to be able to pro-
duce or to build AI systems. One thing that we all have a
tendency to say about AI systems is that they are easier
to build than more traditional programs. Often we end
up forgetting to say “easier” and say “easy” to build. It
is conceivable, and I suspect that some people have trans-
lated “easy to build” into “with a one-day tutorial in AI, I
can go out and build an AI system.” I think that’s mostly
128 THE AI MAGAZINE Fall, 1985
Let’s say that a lot of businesses have some middle
managers who are convinced that all they need in order
to build an AI lab is to hire five or ten people, send them
each out for a one-day tutorial, give them each a book,
and at the end of that one day, they’ll go off and’11 start
producing a great system. I’m hard pressed to believe that
management is somehow going to get so caught up in that
myth that they are going to be able to then react vio-
lently when they find out that those people don’t produce
as much or as quickly as they had hoped. That kind of
judgment doesn’t seem to me to be the kind of judgment
about how long one might have to spend in order to ac-
quire aptitude to do some task. The notion that all you
need is a day or week of training isn’t the kind of judg-
ment that people make about most tasks. There might
be some wishful thinking, and people might try to build a
group because they can’t hire trained AI people. I would
expects that management has a notion of the risk and that
they understand that they are taking a path less likely to
lead to success, at least quickly, than a path that was built
on a stronger experience base.
The third kind of expectation is that every attempt
to build a knowledge base system will succeed gloriously.
If the builders of the system are people with absolutely
no experience in AI, I don’t have any sense at all as to
what might be a reasonable success rate to bet on, but I
suspect it’s low. Even if the people had significant exper-
ience in AI, we still know that there are many attempts
to build AI systems that go on for a while, and then the
people discover that they can’t build the system that they
thought they could. I don’t think that’s the kind of ex-
pectation violation that’s going to pull away support for
AI research, because as long as there’s some reasonable
amount of success, as long as some fraction of the systems
that are attempted turn out to be truly helpful, there’s a
positive, forward-moving attitude. And people say, “Gosh,
I bet I’ll be more fortunate next time.”
The fourth kind of expectation is the expectation about
what the level of performance of a successful AI system is
going to be. There may be people who believe that the
successful AI system will never make a mistake. We all
know that the most we could expect from an expert sys-
tem is a performance level that’s as high as the experts,
and experts, of course, make mistakes. Presumably, we
would be delighted with a performance level that was sub-
stantially less than that. So there are going to be systems
being used that make mistakes. If the people who are us-
ing the systems or asked for the systems to be built started
out with the expectation that those systems would never
make mistakes and then find, much to their surprise, that
they do make mistakes, that’s an expectation violation.
But again I don’t see that that implies, that that suggests
in any way, that these people are somehow going to turn
away from AI. They are going to become better informed.
They are going to understand that the technology that
they are dealing with has certain limits, and they will pre-
sumably be in a position to understand why those limits
are necessary.
If you ask yourself what are the kinds of expectations
that people might have and how might those expectations
be violated, it seems to me that the nature of the viola-
tion is going to be such that the people who had the false
expectation will simply become better informed about the
nature of AI technology. The nature of the violation is not
going to result in rejecting the technology, feeling cheated,
or feeling that somehow AI doesn’t have any promise, with
a result of these people turning away and putting their re-
search dollars into something else.
Questions from the Audience
Following up on Dr. Schank’s suggestion that big busi-
ness is the driver in this expectations failure model or the
AI Winter, I was wondering if people on the panel could
identify any operationally successful expert systems that
have been implemented over the last five years. I know
John McDermott can tell us about R-l, but that is fast
becoming ancient history. Where are some systems, as
John suggested, being implemented that are being viewed
as being successful? I don’t see any.
John McDermott:
Let me try to restate the question. There is, at most, one
successful AI system out there, and if AI has any promise,
why is that? What are the names of some of the other
successful AI systems?
When I give talks on successful AI systems, I put up
slides that have systems with names. Some of the slides
that I put up refer to systems like ACE, CATSl, the
Drilling Advisor-there are a number. If somebody says,
“Just how successful are those systems? To what extent
are they being used on a regular basis? How much money
are they saving? And so on...” I’m not directly involved
with any of those systems. I don’t know how the compa-
nies that participated in the development of those systems
are using them and whatever. I’m told that the systems
are getting better and that the people who work with them
think they are valuable and see promise in them. I do have
occasion to go to Digital Equipment Corp. from time to
time, and in addition to the R-l system, Digital has four or
five other systems that, though they have been used for a
shorter period of time and aren’t used as widely. Neverthe-
less they are being used. There’s a system called XCEL,
which a large number of sales people are now beginning
to use. There’s a system called IMAX, a fairly simple sys-
tem, a part of which is now being used in a manufacturing
There are AI systems out there, and the people who
use them say, “Gosh, I’m glad these systems are here.”
THE AI MAGAZINE Fall, 1985 129
And I suspect that this phenomenon is going to continue.
I believe that the technology is able to provide a lot of
assistance, but it’s taking us some time to develop those
systems. The transfer of AI systems into real working en-
vironments is going to take a few more years. In a few
more years when somebody is asked to name systems, the
list will be much longer, and it will no longer be an issue
of trying to find enough systems so that we’re not embar-
As has been pointed out several times, it’s clearly true that
the explosion in applications work has siphoned off a lot
of our precious Ph.D level talent for AI research. It’s also
true, I think, that it has created an entirely new resource,
and that is what you might rcfcr to as journeyman level
or master’s level AI people. Not all those who have very
little or no formal AI background are not ex-Ph.Ds. Not
all these people are of the sort who read one book, go
to a one-day tutorial, or spent three days working on AI.
In fact, in some of the industrial labs, you have people
who do have, say, a Master’s in Computer Science, who
started out perhaps by reading one book and going to a
tutorial, but over a year or two or three of work in the field,
have actually developed some capability in the field. Now
these are people who never could be and never would be
researchers. They’re not going to get a Ph.D. But they are
capable of being the equivalent of our lab technicians. The
question is: Do you see any way to use this new resource
in our Ph.D style research to alleviate some of the lack of
Roger Schank:
I think that it’s wonderful that those people are being cre-
ated, if they are. I’m a little anxious about how well they
are being created, but I think it’s important. So don’t mis-
understand me on this statement. They’re not researchers.
The worry I have is that they will begin to think they are
researchers. I don’t think we should make that demand
of them. My concern about AI, by the way, and about
the complexity of it and why it’ll take so long, is not so
much that there is a tremendous lacking of ideas that we
need, but also that there is so much engineering involved.
We have tremendous amounts of knowledge to put into
these systems or else get them to learn them on their own.
Either way, there is a tremendous amount of engineering
involved. So we do need to have this engineering class built
in. I don’t think AI research is very easy though, and I
think some Ph.Ds aren’t very good at it. So I’m concerned
about the Masters people doing it.
I’m addressing this question to Ron Ohlander. The ques-
tion deals specifically with the history of speech under-
standing research and the funding of DARPA for that,
but I think it reflects the general problems.with funding
for AI. From 1970 to 1975 there was a project aimed toward
continuous speech understanding undertaken by DARPA,
and it funded several centers. It was perceived as hav-
ing failed. There was a severe retrenchment in the latter
1970s for speech understanding research, and now I see
that under the strategic computing projects, continuous
speech understanding once again is being funded with, I
gather, fairly high expectations of success. My question is,
what has been learned from the previous round of funding
speech understanding research and also of A& so that this
time things will work out differently than they did fifteen
years ago?
Ron Ohlander:
I assume you are addressing the question from a political
standpoint rather than a technical standpoint. And that
is, are we going to have that kind of experience again? I
wasn’t there at the time, so I can’t tell you from my own
personal knowledge, but I’ve talked to people. I think the
problem was the casuality of the political process and that
there were new things that people wanted to do. I think
that program was sacrificed to some of the new things
they wanted. I will also point out that the bulk of the
AI program stayed intact. I think the chances of that
same kind of thing happening again are nil. There is a lot
more consciousness of the world of AI, and the programs
are supported very strongly. In fact, we wouldn’t have
gotten the strategic computing program started without
the enthusiastic endorsement of Dr. Cooper, who is head
of the agency.
AI appears to be a fairly young field. I’m sure that most
people in the room over thirty or so are fully aware of
this, but as a word of kindness for those who are young
and starting in the area, most of our high technology areas
go through boom and bust cycles. Think back through
aerospace industry, back when Sputnik kicked things off.
I’m a mathematician. We had a cycle in the early 1960s
when we could not produce enough Ph.Ds. Many kids
started training, and then in the late 1960s the market was
glutted. Many Ph.D mathematicians couldn’t get jobs in
mathematics at all, much less in a university at the Ph.D
level. This happens in all fields. The universities try to
control it as best they can. So do the industries. But these
things happen. You also appear to be on a slight boom
period with a potential bust coming. And there will be
survival. That will happen. I agree with that. The field
will exist. It will grow. It will be robust. But there will be
many of you in this room or your students who may get
shaken out in the long run. As a touch of kindness to you,
there are several things that can be considered. Check with
the other disciplines. We have been through this before.
In the MAA Journals and AMS Journals for Mathematics
130 THE AI MAGAZINE Fall, 1985
in the mid-1960s, around the 1965-1967 time-frame, have
very serious articles discussing this and how to get out of it.
IEEE transactions for engineers did the same thing. So did
the aerospace industry. Someoue jokingly said: Find what
else to call yourself. For those in AI, a very easy bailout
will be as a computer scientist. If that fails, for those of
you that are appropriately trained, about five years from
now, or seven, or ten, if I read the statistics correctly, call
yourselves mathematicians. We’ll need you again.
I’d like to ask Roger Schank a question concerning the fla-
vor of this convention. I’m a student, and I came to this
convention with certain hopes and expectations, and the
flavor is a little bit different than I expected, to say the
least. What would we do to change the flavor of IJCAI-85
or AAAI-86? What could we do to change this to gear
it not so much toward business and more toward the re-
search? Or more toward getting back to arguing and less
about selling?
Roger Schank:
Well, you’re not going to change this. AI conferences have
an evolution, and if you’ve been to all of them, you begin
to see the process. I can see that this one is going to get
more business, not less. The issue is to be able to start
other forums in other places where people can start to have
those kinds of dialogues. We’ve tried to do that from time
to time. It requires a certain amount of energy for some-
body who actually wants to run a conference. But I think
that that’s probably what’s needed, that AAAI ought to
concern itself with running more than one conference, and
it ought to have one where there aren’t any booths and
there aren’t any tutorials, and see how many people show
up. Maybe the people who showed up would want to talk
about science.
I hate to pick on Professor Schank, but I have another
question directed at him. If corporate centers can’t re-
ally produce researchers, isn’t it the responsibility of an
organization like AAAI to provide more tutorials during
the year where people from corporate research centers can
and get more information and become closer to Ph.D level
researchers or at least not lead corporations down the gar-
den path, thinking that they are actually doing AI work?
Roger Schank:
There’s a presupposition in your statement that is wrong,
that through lots of tutorials you’ll learn to be an AI re-
searcher. I think that’s blatently false. AAAI should go
on the way it is. Tutorials and shows are wonderful. But
if that’s the only conference we have, that’s the problem.
We just don’t have the other ones. With respect to train-
ing people, Ph.D research, at least in my laboratory, used
to take three to four years. Now it seems to be taking five
to six. It’s a long process. You have learn a lot of stuff and
then try to create something on your own. That’s what
a Ph.D means. All I can say is that that isn’t the same
thing as being trained through tutorials. There’s a mas-
ter’s degree program at Stanford, but I don’t know how
many other AI master’s programs really exist. Probably
what we need are more of those, where people in industry
can take off for a year and learn about AI techniques and
then return to their company. I don’t think tutorials will
ever do it because if you don’t have hands-on experience,
it doesn’t work in AI.
John McCarthy (from the floor):
I’d like to comment on Schank’s point in the question to
him. The AAAI does sponsor workshops, and anyone who
wishes to organize a workshop in some special topic should
get in touch with me, because I have undertaken to con-
tinue that.
I have a question about historical perspective on this en-
tire issue you’re discussing. There have been a number of
technologies that have run into dead ends, like dirigibles
and external combustion engines. And there have been
other ones, like television and, in fact, the telephone sys-
tem itself, which took between twenty and forty years to
go from being laboratory possibilities to actual commercial
successes. Do you really think that AI is going to become
a commercial success in the next ten to fifteen years, or
will it have a longer gestation period? That’s addressed
to the entire panel.
Roger Schank:
There are differences between AI and derivatives of AI.
AI means that you are going to make machines intelligent.
Remember that? We were interested in the mind and what
it meant to be intelligent and how thinking took place.
Remember all those issues? That is what AI is about.
Now that doesn’t mean that the companies discuss those
issues on a daily basis. Companies are great-they should
exist. And they should do derivative work. That means
providing software which is better than the software that
we had before, which has all kinds of AI flavor, all kinds
of AI derivatives. It just shouldn’t be confused with AI.
That’s all.
John McDermott:
Let me add to that. As we learn some things in AI, it
simply makes it possible to develop systems whose be-
havior is better in certain kinds of circumstances. And
that’s going to continue. So I don’t see that there are any
sharp boundaries. I don’t know what it would mean to say
that we’re going to try to go back and have increasingly
THE AI MAGAZINE Fall, 1985 131
dumber programs. We’re going to understand better how
to solve certain classes of tasks, and that’s going to con-
tinue. People are going to like that, and so they’re going
to encourage these derivative companies to produce that
kind of software.
What I was trying to point at was that you can have a
technology develop and then continue at a very low level.
We still have steam turbines. We still have lighter-than-air
aircraft. But they’re not a major commercial force.
Mitch Waldrop:
Picking up on your comment about historical analogies
and also picking up on Chandrasekaran’s talk about work
on hard problems, historically you can look at develop-
ment of the steam engine back toward the end of the 18th
century, beginning of the 19th century. This was done by
people just trying to solve the immediate problems. In
the long run, however, because they had to understand
the nature of heat and work in steam engines, they dis-
covered thermodynamics. Carnot was an engineer trying
to understand the efficiency of steam engines. In the long
run, forcing yourself to work on real-world hard problems
can help.
This is just a comment on what John McDermott had to
say about companies not really having high expectations.
A lot of the real tight purse strings, the big money for
AI, is being directed just toward places that are gearing
their people up from reading the book, going to a tuto-
rial, and, believe it or not, those companies do have high
expectations of the software that they are hoping to come
up with. And I’d just like to say that dismissing that as
not a
big issue, as far as money drying up, an AI Winter,
or an AI dusk, is not something we should forget about, if
we’re worried about funds for research.
I’d like to say also that I’ve visited a number of labs of
large companies that have very serious AI efforts going on
with extremely large research programs. Those companies
are succeeding in turning out good products. I think that
will counterbalance the effect of the companies who do not
seriously address the technology.
B. Chandrasekaran:
I would like to make one remark about the whole trend of
using expert systems. At least it has done one thing well.
It has concentrated a lot of analytic effort in some areas
of the company’s operation or in some body of knowledge.
But at the end of it you might conclude, ‘LGee, I don’t
need any expert system. I understood the whole thing
very well. I can summarize the whole thing in eighteen
rules. Or I can just write up a simple system to do that.”
But that still says it has a good effect on the operations of
the company. I think if those things take place, AI should
also take some credit for it too, because it has emphasized
looking at knowledge and analyzing it properly.
I’d like to address this to anybody on the panel who wants
to answer. What feelings do you have about the sentiment
that AI is rather fragmented and is a bunch of solutions
looking for problems and that there is very little methodol-
ogy being developed to allow people to vigorously analyze
problems and pick and choose the methodology to use in
different parts of the problem? Right now it’s a very ar-
cane art.
I don’t think anybody would disagree with that. It’s what
you would expect in the early stages of an engineering
B. Chandrasekaran:
I would like to comment that most people have a false no-
tion of rigor as proving mathematical theorems. I think
a lot of people are pretty rigorous. They get up in the
morning and tear their hair out over issues, and they don’t
accept the first solution to come to mind. But they don’t
express their solutions as mathematical theorems. So with
that particular proviso, I think people are aware of it and
people are doing rigorous work. But they may not recog-
nize it as rigorous because people might have wrong no-
tions about the definition of rigor.
I’m probably one of the few business people still here. I
wanted to say a couple of words of reassurance about busi-
ness expectations. I come from a Fortune 50 company.
We’ve got a small AI group of four, but we’re not claiming
to do AI. We’re attempting to find things that we can pick
now and apply. I think you’ll see that creating a constant
demand-a constant demand for perhaps what you’re cal-
ling engineering. We’re not seeking to be AI researchers;
we never claimed to be. What we will become will be the
moral equivalents of systems analysts. That kind of think-
ing we can learn from doing. And I think that that is a
point that has not yet been said. There’s a different kind
of demand for a constant kind of application that you’re
going to see, and that will flatten your boom and bust
Well, I concern myself with the following issues. I don’t
know who you are or what company you come from, but
132 THE AI MAGAZINE Fall, 1985
I concern myself with the training of the people who are
at that company, who you now claim are an AI group. I
worry how much AI they know.
Very little. And we’ll be up front about that.
Roger Schank:
But soon they will be doing applications in your company,
which if it’s a Fortune 50 company, is an important com-
pany and will be claiming or knowing about the successes
or failures of this AI group and its ability to do applica-
tions. I’m concerned whether those people know anything
at all about AI.
And what we’ll say in response to that is we have found one
more information-handling technology that is applicable,
possibly in conjunction with others, that helps solve some
problems. We’re not fools. We will pick applications where
we can have some incremental value. We’re not going to
try for a home run and strike out swinging.
Roger Schank:
It would make me a lot more comfortable if you didn’t say
we have a small AI group-if you said we have some smart
programmers who are working on some hard problems.
Then perhaps that’s the term you would feel more com-
fortable with. We’re already moving away to engage in our
own expectation management. We’re already moving away
from terms such as “expert systems” toward “knowledge
systems.” We’re trying to drop the term “AI.” We’re doing
all the expectation management you’re talking about.
Roger Schank:
I wish I could get a follow up on this two years frorn now
and see how it worked out.
We’ll still be here.
Roger Schank:
I guess so.
I just wanted to ask what you think people can do about
avoiding the sensationalist journalism about artificial in-
telligence. You see a lot of articles now about machines
that can think. People get a lot of expectations, especially
those who don’t have a technical background or expertise.
They arc starting to look for a lot to come out of artificial
intelligence. I have a feeling that this boom-bust cycle, at
least in the popular sense, could easily be fueled by current,
sensationalist journalism. I’m wondering what you think
can be done to cahn people down about the possibilities
of AI and make thern more realistic.
Mitch Waldrop:
In part, I have to say I can’t help you. The Nutzonnl
Enquirer is always with us. I can’t make it go away. In
part, it’s inevitable. There’s the phenomenon of the three-
day wonder. This is more than three days, but it’s the
sarne phenomenon, that something new appears on the
horizon, people get excited about it, get enthusiastic about
it, and they gush. After awhile, it goes away. I will say that
in my talks with lay people about artificial intelligence, I
have exactly the same impression that John McDermott,
talked about in that they seem to have a vague idea that
great things can happen. have sublime confidence in you
people, that you can do anything. But when it gets down
to the nitty-gritty, they tend to be pretty unimaginative
and have pretty low expectations as to what can be done
Referring back to this last weekend that I spent, I had to
keep boosting people’s ideas about possibilities, such as
what a teaching machine might be able to do. I think it,
might be less of a problem than you really imagine. Yes,
the sensationalism will go along for a while, and then it’ll
blow itself over when they find something else to be excited
about. I’m surprised Michael Jackson hasn’t stopped it
already. But I don’t know how tremendous an impact
that will have on serious decision makers. Let me add
something else, too, that I keep hearing. I’m again going
to paraphrase John McDermott in terms that might be
much blunter than he would prefer. There seems to be
this unconscious assumption around here that everybody
in the world except artificial intelligence researchers is kind
of dumb, mindless, unreasonable, inflexible, and incapable
of learning, and in short, behave not unlike computers.
Most people I know out there are rather reasonable and
sensible-it may not be as much of a problem as you think.
It seems that academic AI people tend to blame everyone
but themselves when it comes to problems of AI in terms
of relationship to the general society. Charges of arro-
gancc are traditional, but I’m concerned with a different
one. It seems that there’s a need for master’s level en-
gineering education in a more organized fashion than the
zero start situation of so many AI groups in fairly large
companies that are starting up efforts. It seems that, Ph.D
level people who can teach or master’s level people with
background can usually do much better things in terms of
personal reward than teach It seems that there is a need
for some kind of mechanism, either within the universities
or possibly in doing an endrun, that creates an alterna-
tive institution for a master’s level engineering education
THE AI MAGAZINE Fall, 1985 133
in AI. I’d like to raise that as an issue, and ask the panel
to comment on that.
Roger &hank:
The problem with the master’s training is not that we
don’t want to do it. I think the problem is that everyone
wants someone else to do it. In fact, representing most of
the AI faculty at Yale sitting at this table, I don’t want to
do it. You don’t want to do it. We’re happy to sit there
and say we should have master’s students, but the dearth
of good Ph.Ds are going to the universities to do research
and none of them want to do these kinds of training pro-
grams. I hope they’ll come, but I don’t know who is going
to start it. You have to get someone who is very dedicated
to that proposition as being something important. When
that happens, they occur. But it’s very hard to get anyone
to really care, even though you’re right.
Do you think the initiative is then more appropriate in
coming from the industrial community to try and stim-
ulate special programs in the universities? I know that
Stanford has had a master’s program oriented toward AI,
and it was established partly because of very considerable
financial support and promises over at least medium term
rather than short term from industry. Notably, Bell Labs
helped sponsor the masters in computer science, which
was the reason why there was enough masters’ programs
to support masters in AI at Stanford. So is it appropriate
for industry to push on the universities and come up with
a lot of dollars to support these masters’ programs?
Roger Schank:
That might help, but again you would have to get some-
body who cared about it. There’d have to be enough in
it. Suppose you took a guy who was starting out at a new
university, and he was concerned about how was he going
to get good computing facilities and attract graduate stu-
dents. If they came to a person like that and said we’ll
give you extra money so you can do that, by the way, if
you’ll do the masters program, he might find himself at
the center of a very nice situation. I think you’d have to
find the right person, but it could be done.
AAAI Membership Benefits:
Subscription to the
AI Mayazzne
AAAI Membership Directory
Reduced subscription rate to the AI Journal
. Reduced registration fee at IJCAI Conference
Early announcement of AAAI-sponsored activities
Affiliation with the principal AI association
by Robert Wilensky,
University of California, Berkeley
The perfect introduction to LISP for courses in
artificial intelligence or programming languages.
@ Norton
Solutions Manual
by Robert Wilensky
A Taste of Smalltalk ,0o+
by Ted Kaehler, $P ‘?-
Apple Computer formerly Xerox PARC) Go
and Dave Patterson,
University of California, Berkeley
A brief tour of the Smalltalk-80TM language and
environment, with program examples that gently
demonstrate the power of object-oriented
Pascal! Second Edition
by Doug Cooper and Michael Clancy
both of the University of California, Berkeley
Still stylish and witty, now with
a brief introduction to systems and software
early section on one-dimensional arrays
procedures and parameters combined
greater emphasis on recursion
new chapter on software engineering
arrays treated before records
new material on analysis of algorithms, program
& correctness, sorting, searching, and matching
Instructor’s Manual by Doug Cooper
W W Norton&Company, Inc.
500 Fifth Avenue, New York, NY 10010
Standard Pascal User
Reference Manual
by Doug Cooper
The most readable and complete reference for
Standard Pascal.
Visit us at booth #602
134 THE AI MAGAZINE Fall, 1985
... This succinct review of artificial neural networks promptly highlights that the symbiosis between the prevailing science of the brain, as composed of neurons and their connections, and logic/mathematics, both in the application of functions and the (early!) application of network theory, was made apparent from the start. 1 Subsequent synergies with computer science have developed along these lines such that improved storage and processing power now make feasible achievements foregone in the periods of reduced funding and interest in AI during the 1970s, '80s, and '90s, known as the "AI winter" (McDermott et al., 1985;Russell & Norvig, 2003, p. 22;Howe, 1994; see also Lighthill, 1973). Current deep neural networks now have many more nodes, and, crucially, additional hidden layers, but the very foundations of the system remain pretty much the same since its beginnings. ...
Cognitive scientists deal with technology in a very particular way: they use technology to understand perception, action, and cognition. This particular form of human-machine interaction (HMI) is very well illustrated by the use cognitive scientists make of artificial neural networks as models of cognitive systems and, more concretely, of the brain. However, the activity of cognitive scientists in this context suffers from the shortcoming of epistemic opacity: artificial neural networks are too difficult to interpret and understand, so in many cases they remain black boxes for researchers. In this paper, we provide a diagnostic for such epistemic opacity based on dominant cognitive science’s lack of theoretical resources to account for the activity of artificial neural networks when taken as models of the brain. Then, we offer the guidelines of a solution founded on the notion of information developed in ecological psychology.
... 等,同时它们也 严重影响了地表荒漠化程度 [6] 。 面对风沙危害问题, 学者、工程人员和政府部门积极研究风沙防护治理 方案 [7] 。根据工程原理,可以通过增加地表粗糙度、 降低贴地表风速、减少被风吹起的沙颗粒数,以及 增加可供植物生长的土壤养分或者水分等方法实 现对风沙的有效防护治理。这些方法主要包括物理学 方法 [8][9] 、生物学方法 [10] 以及化学方法 [11] 。根据防 护功能,防沙体系依据防沙专家常年积累的经验知 识,主要从阻沙作用、固沙作用、输沙作用和导沙作 用进行防护技术的遴选以及时空布局的动态调整。由 于这些经验并没有被系统地总结与分析,当防沙专家 不在场时,风沙防护治理无法得到有效开展。 防沙体系方案的构建、 优化以及评估需要依据防 沙专家常年积累的经验知识。应用弱的知识提取方 法,可能无法提取隐性知识,或者造成知识库的不精 确性。刻意的知识表达方案会导致知识的提取更趋 于生成表达驱动的知识库,而不是知识驱动的知识 库 [12] 。专家知识提取方法涉及多个研究领域,例如生 态学 [13] 、认知心理学 [14] 和人工智能 [15] 等。如何将大 量的知识转化为一套精准的事实和规则,是专家知 识提取应用的主要"瓶颈" [16] 。人工智能寒冬 [17] 的打破也为知识提取带来了新的方法。模糊模型可 将测量的信息通过隶属度函数(MFs)语言解释联系 起来,成为知识提取的一种有效方法 [18] ,此种能力 使其既可以处理定量信息,也可以处理定性信息 [19] 。 风沙防护治理系统是典型的复杂系统 [20] ,具有 自适应性、自组织性和涌现行为的特点。随着系统复 杂性的增加,如果缺乏有效的管理与控制,模拟系统 往往会逐渐偏离实际系统,直接影响风沙防护治理的 效果。首先,以往的系统分析方法往往难以刻画系统 各部分之间的相互关系,而基于虚拟系统或者人工 系统的研究方法不再以逼近实际系统为研究目标, 而是构建人工对象,通过"观察"对象的行为模式 生成模型;其次,面对复杂系统不能进行有效的实验 的问题,利用人工系统的思想和方法,通过建立计算 实验来分析复杂系统 [21] 。在计算实验中,计算机不只 是仿真的工具,同时也作为一个"实验室"进行与 系统行为和决策支持等相关的"实验" [22][23] [25] 。Dong Z B 等人 [2] 在阻沙栅 栏的基础上加入草方格防沙体系来增加地表空气 动力学粗糙度、固定流沙 [26] 和降低地表风速 [27] ,在 国内外引起广泛关注。Bo T L 等人 [28] 应用数值模型 计算出在给定草方格长度消退速率和年度风况风 速情况下的最佳草方格间距和宽度,计算过程遵循了 物理过程计算规则和语义。风沙防护治理人员为优化 防沙体系进行了一系列改进,例如采用耐久材质 [29][30][31] , 改进对防沙体系的管理和控制 [32] 等。 1999 年王雪芹 和雷加强 [33] ...
Full-text available
Using Taklimakan desert highway and its sand-breaking system as the research object for complex system management and control, the ACP-based parallel intelligence theory to deal with the problems of the difficulty in model-ing, analyzing, and predicting of sand-breaking system was applied, to realize the intelligent decision support for sand-breaking system management and control, support the sustainable development of aeolian environment. The expert experience knowledge to construct the artificial desert highway sand-breaking system was extracted by simulating physical process. The sand-control efficiency index of the artificial system was calculated using equations, then evaluated and modified by comparing and learning with the actual system. Using parallel intelligent theory, the artificial system and the actual system can learn from each other to provide decision support for sand control.
... This work is in fact regarded as the origin of today's deep CNNs. R1, developed by McDermott in 1982, was the first successful commercial expert system employed in the digital equipment industry, for the configuration of new computer systems' orders [56]. In nearly 4 years, the firm added 40 million dollars of revenue using R1. ...
Full-text available
Artificial intelligence has witnessed exponential growth in the past decade. Advances in computing power and the design of sophisticated artificial intelligence algorithms have enabled computers to outperform humans in a variety of tasks. Yet, artificial intelligence's path has never been smooth, having essentially fallen apart twice in its lifetime after periods of popular success. We provide a brief rundown of artificial intelligence's evolution, highlighting its crucial moments and major turning points from inception to the present. In doing so, we attempt to learn, anticipate the future, and discuss what steps may be taken to prevent another winter.
... This work is in fact regarded as the origin of today's deep CNNs. R1, developed by McDermott in 1982, was the first successful commercial expert system employed in the digital equipment industry, for the configuration of new computer systems' orders [56]. In nearly 4 years, the firm added 40 million dollars of revenue using R1. ...
The field of artificial intelligence (AI), regarded as one of the most enigmatic areas of science, has witnessed exponential growth in the past decade including a remarkably wide array of applications, having already impacted our everyday lives. Advances in computing power and the design of sophisticated AI algorithms have enabled computers to outperform humans in a variety of tasks, especially in the areas of computer vision and speech recognition. Yet, AI's path has never been smooth, having essentially fallen apart twice in its lifetime ('winters' of AI), both after periods of popular success ('summers' of AI). We provide a brief rundown of AI's evolution over the course of decades, highlighting its crucial moments and major turning points from inception to the present. In doing so, we attempt to learn, anticipate the future, and discuss what steps may be taken to prevent another 'winter'.
... In fact, the early 1950's and 1960's as well as the early 1980's saw similar AI bubbles (see this interesting summary by Luke Muehlhauser for Open Philanthropy (Muehlhauser, 2016)). These AI bubbles have been followed by what have been dubbed "AI Winters" (McDermott et al., 1985). ...
Full-text available
Machine Learning (ML) is one of the most exciting and dynamic areas of modern research and application. The purpose of this review is to provide an introduction to the core concepts and tools of machine learning in a manner easily understood and intuitive to physicists. The review begins by covering fundamental concepts in ML and modern statistics such as the bias-variance tradeoff, overfitting, regularization, and generalization before moving on to more advanced topics in both supervised and unsupervised learning. Topics covered in the review include ensemble models, deep learning and neural networks, clustering and data visualization, energy-based models (including MaxEnt models and Restricted Boltzmann Machines), and variational methods. Throughout, we emphasize the many natural connections between ML and statistical physics. A notable aspect of the review is the use of Python notebooks to introduce modern ML/statistical packages to readers using physics-inspired datasets (the Ising Model and Monte-Carlo simulations of supersymmetric decays of proton-proton collisions). We conclude with an extended outlook discussing possible uses of machine learning for furthering our understanding of the physical world as well as open problems in ML where physicists maybe able to contribute. (Notebooks are available at )
... Actually, the straightforward definition of AI given, for example, by Roger Schank when he says that "AI means that you are going to make machines intelligent. […] We were interested in the mind and what it meant to be intelligent and how thinking took place" (Drew McDermott et al. 1985) could not go unnoticed and unanswered by philosophers. ...