Path: utzoo!attcan!uunet!ogicse!uwm.edu!rpi!zaphod.mps.ohio-state.edu!mips!apple!amdahl!kp
From: kp@uts.amdahl.com (Ken Presting)
Newsgroups: comp.ai
Subject: Re: Another letter to the New York Review
Summary: Why we should want to define "intelligence"
Keywords: emergence, defintions, intelligence
Message-ID: <2c7R02Sr8d8b01@amdahl.uts.amdahl.com>
Date: 1 Mar 90 17:36:54 GMT
References: <12085@venera.isi.edu>
Reply-To: kp@amdahl.uts.amdahl.com (Ken Presting)
Organization: Amdahl Corporation, Sunnyvale CA
Lines: 69

In article <12085@venera.isi.edu> smoliar@vaxa.isi.edu.UUCP (Stephen Smoliar) writes:
>...We have a tendency to assume that every word can be associated with an
>OBJECT, but we may be deliberately confusing ourselves if we try to assume that
>intelligent behavior has the sort of object properties we associate with a loaf
>of bread or a jug of wine.  I would argue that the temporal nature of behavior
>makes it quite a different "thing" to talk about (to the extent that WHAT
>HAPPENS over an interval of time--as opposed to the temporal interval
>itself--really constitute a "thing").  I suppose what I am trying to
>argue is that we tend to talk about properties which emerge from the
>dynamics of a complex system AS IF THEY WERE CONCRETE ARTIFACTS when
>we really have no justification for doing so.  Rather than agonizing
>over the "nature of intelligence"--as if it were a "thing" which had
>a "nature"--perhaps we should be reviewing our vocabulary as it pertains
>to talking about ANY form of behavior.

Stephen, I agree wholeheartedly with your ultimate suggestion, but I'd
like to put in a few words in favor of hairsplitting.  (This will give
me a temporary respite from agonizing over my reply to your terrific
article an the Kronos Quartet performance.  You made me wish I'd been
there.)

In the first place, human bodies (and computers) *are* objects.  After
all, that is how Penrose gets into the debate at all.  One way to think
about intelligence is to try to figure out what "object properties" the
body has which make possible intelligent behavior.  The neurologists are
working on this from one angle.  AI has been working on it from another.
So have philosophers.

In the second place, the dynamical properties of a running computer are  
neatly determined by its program.  The program is of course a static
object (modulo the contents of its data structures, but you see what
I mean).  The $64K question is: what properties does the program need
to have, so as to allow the emergence of intelligent behavior?

In the third place, its easy to give lots of necessary conditions for
the intelligence of behavior.  This is a worthwhile enterprise, but
until we can state a sufficient condition, those goal posts are gonna
keep on movin'.  I doubt that Turing-style imitation conditions will
fill the bill, and in any case they are useless as design guides.

Finally, words like "conscious", "intelligent", and "rational" have
very important roles in the conceptual framework of our moral and
legal system.  What if our machines demand suffrage?  Freedom of speech?
Freedom of movement?  These are not pressing problems, I'll admit.
But I think that the social and moral impact of AI deserves at least
as much investigation as the same issues in Biotechnology.  Someday, a
computer is going to kill somebody, and it won't look like a "glitch".
Questions like "Does the machine understand right and wrong" and "was
it intentional" will determine whether it's the machine or the programmer
that gets powered off.  

A lot of philosophers have gotten badly burned by the concept of 
emergent properties.  I don't think Searle is relying on that concept,
but his arguments about programs not determining any "causal powers"
are similar to some arguments based on emergence.  In my own thinking,
I have found that the concept of "normative property" does a much
better job of capturing the relation between (eg) rationality and
goal-seeking behavior.  Dan Dennett's "Intentional Stance" is an OK
example of a normative approach, although he's not always as thorough
as someone like Davidson.  If you find that the concept of emergence
is useful in formulating some ideas, you will probably find
normativity both more useful and more reliable.

That's what I'd recommend in response to your suggestion for a vocabulary
review.  I was looking for a way to bring in normativity in response
to your Kronos point about the expressibility of information.  That
is a Very Difficult problem.  I think that reading *anything*, even
reading meters, is immensely complex.  Much harder to understand than
pain or pleasure, which are awful.