Robotics

[The following is an excerpt from the introduction of ROBOTICS,
a collection of essays edited by Marvin Minsky on the latest 
generation of robots being "born" in research laboratories 
around the world.  The book was published by Anchor Press/
Doubleday in 1985.]

Artificial intelligence, commonly called AI, is an inquiry not 
just into computers and new kinds of machines but also into the 
nature of intelligence itself -- a subject no one understands 
very well.  Why not?  Perhaps because no one's ever had a chance
to look at any other kind of intelligence except our human kind!
Why is this important?  Because it is very hard to understand 
anything until you have other things to compare with it.  

At last we are about to have a chance to meet some really alien 
minds!  Our first encounters will not be in the forms of highly
evolved intelligences from other planets.  It seems much more 
likely now that our first encounters with alien minds will be 
with ones that we ourselves have built -- our very own AI 
machines.  

The only trouble is that we don't know how soon we're going to
be able to meet them, because we don't know how to make 
machines intelligent.  What, exactly, is the problem?  Most 
people outside the field of AI think they already know what 
it is:  It has something to do with the mysteries of Inspiration,
Creativity, Intuition, Originality, and Emotion.  However, 
if there's one thing we have surely learned in research on 
artificial intelligence, it is that this is not the best way 
to describe the problem.  In AI research, we are learning 
how amazingly complicated are the simple things we do in  
everyday life.  

When we admire the outstanding performances of our greatest 
thinkers, athletes and the like, the trouble is that we look 
right past the wonderful things we -- all of us -- do whenever
we walk, talk, see, reason, and plan.  It is not necessary to 
start AI research by trying to write computer programs that 
will paint great paintings or write brilliant plays.  At 
this stage, we can learn more by trying to develop programs 
that can distinguish a dog from a cat by sight or carry on 
a simple, childish coversation.  

Why is it so hard to give machines common sense?  What is the
difference between how expert computer systems work and how 
human common sense proceeds?  This is a matter of the range
and varieties of knowledge they use.  The expert system 
manages to work using only a few varieties of highly 
specialized knowledge about its subject matter.  Within 
each of those knowledge categories, the program may "know" 
thousands of items, but they are all essentially of the same
type.  However, the knowledge that a sensible person must 
have to get around the ordinary world is not anything like 
that:  One has to know thousands of different kinds of things.
So, there is much more complexity in how things are represented 
in our minds than in the computer programs.  

It would be a mistake to try to study commonsense reasoning
without also studying the learning process.  Even after we 
understand how to do it, it will still be an enormous job 
to program into a machine all the knowledge a reasonable 
person must have.  In the end, it would be easier, and better,
to program our machines to acquire such knowledge themselves:
by watching what happens, by having conversations with 
knowledgeable people, by asking questions and making experiments,
reading books and doing all the other things that people do to 
educate themselves.  Once we discover an adequate set of 
principles for this process, there is no reason that 
intelligent computers could not learn much faster than we do.

What do we mean by "intelligence"?  We mean the ability to solve
problems that people would say require intelligence.  Still, if
intelligence is a single thing, could there be many clever ways
for brains -- or machines -- to think?  It certainly would be 
surprising if there was one single way.  It could turn out that 
there are different ways for machines to solve problems.  
Indeed, some researchers in AI hope to adopt some methods from
logic and mathematics, in which everything else can be deduced
from only a few basic principles.  

Yet another approach to how to make intelligent machines is to
copy human psychology.  We still don't know enough about 
psychology; nevertheless, since we're the only creature 
around that does the kinds of intelligent things we want our 
computers to do, we may as well use ourselves as examples. 
 
One thing few outsiders are told is just how long it takes 
to solve these kinds of problems.  The new sciences of 
robotics and artificial intelligence are very young, barely 
25 years old.  A mere handful of people were working on these
profound problems in 1960.  In the new science of artificial
intelligence, only a hundred or so people are involved in what
one would call basic research.  

As progress continues, we'll reap the fruits of our research 
and start to see machines that display more genuine signs of 
having minds.  We'll start to give them learning skills to 
organize their little minds, so that they can learn from us, 
and from each other, as we do.  

[by Marvin Minsky]