DRogers@SUMEX-AIM.ARPA@sri-unix.UUCP (09/14/83)
From: David Rogers <DRogers@SUMEX-AIM.ARPA>
[This continues a discussion on Human-Nets. My original statement,
printed below, was shot down by several people. Individuals certainly
derive satisfaction from hobbies at which they will never excel. It
would take much of the fun out of my life, however, if I could not
even imagine excelling at anything because cybernetic life had
surpassed humans in every way. -- KIL]
From: Ken Laws <Laws@SRI-AI.ARPA>
Life will get even worse if AI succeeds in automating true
creativity. What point would there be in learning to paint,
write, etc., if your home computer could knock out more
artistic creations than you could ever hope to master?
I was rather surprised that this suggestion was taken so quickly
as it stands. Most people in AI believe that we will someday create an
"intelligent" machine, but Ken's claim seems to go beyond that;
"automating true creativity" seems to be saying that we can create not
just intelligent, but "genius" systems, at will. The automation of
genius is a more sticky claim in my mind.
For example, if we create an intelligent system, do we make it a
genius system by just turning up the speed or increasing its memory?
That"s like saying a painter could become Rembrandt if he/she just
painted 1000 times more. More likely is that the wrong (or uncreative)
ideas would simply pour out faster, or be remembered longer. Turning
up the speed of the early blind-search chess programs made them
marginally better players, but no more creative.
Or let's say we stumble onto the creation of some genius system,
call it "Einstein". Do we get all of the new genius systems we need by
merely duplicating "Einstein", something impossible to do with human
systems? Again, we hit a dead end... "Einstein" will only be useful in
a small domain of creativity, and will never be a Bach or a Rembrandt
no matter how many we clone. Even more discouraging, if we xerox off
1000 of our "Einstein" systems, do we get 1000 times the creative
ideas? Probably not; we will cover the range of "Einstein's" potential
creativity better, but that's it. Even a genius has only a range of
creativity.
What is it about genius systems that makes them so intractable?
If we will someday create intelligent systems consistently and
reliably, what stands in the way of creating genius systems on demand?
I would suggest that statistics get in our way here; that genius
systems cannot be created out of dust, but that every once in a while,
an intelligent system has the proper conditioning and evolves into a
genius system. In this light, the number of genius systems possible
depends on the pool of intelligent systems that are available as
substrate.
In short, while I feel we will be able to create intelligent
systems, we will not be able to directly construct superintelligent
ones. While there will be advantages in duplicating, speeding up, or
otherwise manipulating a genius system once created, the process of
creating one will remain maddeningly elusive.
David Rogers DRogers@SUMEX-AIM.ARPA
[I would like to stake out a middle ground: creative systems.
We will certainly have intelligent systems, and we will certainly have
trouble devising genius systems. (Genius in human terms: I don't want
to get into whether an AI program can be >>sui generis<< if we can
produce a thousand variations of it before breakfast.) A [scientific]
genius is someone who develops an idea for which there is, or at least
seems to be, no precedent.
Creativity, however, can exist in a lesser being. Forget Picasso,
just consider an ordinary artist who sees a new style of bold,
imaginative painting. The artist has certain inborn or learned
measures of artistic merit: color harmony, representational accuracy,
vividness, brush technique, etc. He evaluates the new painting and
finds that it exists in a part of his artistic "parameter space" that
he has never explored. He is excited, and carefully studies the
painting for clues as to the techniques that were used. He
hypothesizes rules for creating similar visual effects, trys them out,
modifies them, iterates, adds additional constraints (yes, but can I
do it with just rectangles ...), etc. This is creativity. Nothing
that I have said above precludes our artist from being a machine.
Another example, which I believe I heard from a recent Stanford Ph.D.
(sorry, can't remember who): consider Solomon's famous decision.
Everyone knows that a dispute over property can often be settled by
dividing the property, providing that the value of the property is not
destroyed by the act of division. Solomon's creative decision
involved the realization (at least, we hope he realized it) that in a
particular case, if the rule was implemented in a particular
theatrical manner, the precondition could be ignored and the rule
would still achieve its goal. We can then imagine Solomon to be a
rule-based system with a metasystem that is constantly checking for
generalizations, specializations, and heuristic shortcuts to the
normal rule sequences. I think that Doug Lenat's EURISKO program has
something of this flavor, as do other learning programs.
In the limit, we can imagine a system with nearly infinite computing
power that builds models of its environment in its memory. It carries
out experiments on this model, and verifies the experiments by
carrying them out in the real world when it can. It can solve
ordinary problems through various applicable rule invocations,
unifications, planning, etc. Problems requiring creativity can often
be solved by applying inappropriate rules and techniques (i.e.,
violating their preconditions) just to see what will happen --
sometimes it will turn out that the preconditions were unnecessarily
strict. [The system I have just described is a fair approximation to
a human -- or even to a monkey, dog, or elephant.]
True genius in such a system would require that it construct new
paradigms of thought and problem solving. This will be much more
difficult, but I don't doubt that we and our cybernetic offspring will
even be able to construct such progeny someday.
-- Ken Laws ]POURNE%MIT-MC@sri-unix.UUCP (01/11/84)
From: Jerry E. Pournelle <POURNE @ MIT-MC>
I should have thought that if you can make a machine more or
less intelligent; and make another machine ABLE TO RECOGNIZE
GENIUS (it need not itself be able to "be" or "have" genius)
then the "genius machine " problem is probably solved: have the
somewhat intelligent one generate lots of ideas, with random
factors thrown in, and have the second "recognizing" machine
judge the products.
Obviously they could be combined into one machine.andree@uokvax.UUCP (01/24/84)
#R:sri-arpa:-1568100:uokvax:900008:000:548 uokvax!andree Jan 22 00:59:00 1984 This sounds very similar to the "Truth Recognizers" in Hofstadter's and Rucker's writing - machines that will tell if an arbitrary statement (or theorem) is true. Hook them up to a statement/theorem generator, and you can find out if any statement/theorom is true by waiting long enough. The problem is, defining "true" is a hard problem - rucker claims you can't do it in any finite amount of space/time. [I tend to believe him.] Something tells me "genious" will suffer from the same problem. Something to do with context-sensitivity... <mike
simon@psuvax.UUCP (Janos Simon) (01/27/84)
It would be useful if people thought or learned something about subjects they write about. Theorem-proving (as in proving any theorem of a sufficiently string theory) is impossible - it is not recursive for theories as weak as Peano Arithmetic, it cannot be done in less than exponential time for almost trivial decidable theories (like 1st order theory of real addition), and it is co-NP complete (and therefore probably not doable in polynomial time) even for the propositional calculus. Obviously this has nothing to do with context sensitivity. What is "genious"? *Down with intellectual pollution!* js
sew@minn-ua.UUCP (01/30/84)
#R:sri-arpa:-1568100:minn-ua:2300001:000:756 minn-ua!sew Jan 30 10:30:00 1984 An article in Nature (June 83? Also a few months later in Psychology Today) suggested that mammal brains might use a related approach to help determine reality. The idea is that when mammals dream, the "lower" brain is feeding random signals into the "higher" brain and whatever responds is inhibited. Thus many false connections would get broken, along with some infrequently reinforced good connections. The concept seems to be simple. If a connection responds to a random signal, it is probably a random connection. Remove random connections what what will be left is connections which are part of patterns. "To make a statue of an elephant, remove everything which does not look like an elephant." Scot E. Wilcoxon ...ihnp4!umn-cs!minn-ua!sew
andree@uokvax.UUCP (01/31/84)
#R:sri-arpa:-1568100:uokvax:900010:000:653 uokvax!andree Jan 29 13:53:00 1984 Sorry, js@psuvax, but I DO know something about what I spoke, even if I do have trouble typing. I am aware that theorom-proving machines are impossible. It's also fairly obvious that they would use lots of time and space. However, I didn't even MENTION them. I talked about two flavors of machine. One generated well-formed strings, and the other said whether they were true or not. I didn't say either machine proved them. My point was that the second of these machines is also impossible, and is closely related to Jerry's genius finding machines. [I assume that any statement containing genius is true.] Down with replying without reading! <mike