An AGI program cannot be a programming environment where specific
combinations of keywords, variable values and grammar correspond with
specific output values (of meaning or action) on a one to one basis.  This
goes for the application of numerical processes as well.  This
correspondence would represent a correspondence of data object to reference
or a correspondence of application of method to reference.

So the AGI program has to be able to do some learning for itself.

On the other hand we want to allow programmers to use these kinds of
methods if they choose.

To put this in AI terms, we want the program to be able to make some good
guesses.  If you understand what I am getting at then you might also see
that this seems to suggest that the earliest AI models were on the right
track to AGI.  For example, an AGI program might derive a conclusion based
on categorization of type and knowledge that a variable or a reference
belonged to that category.  However, the program would have to make some
conclusions based on incomplete knowledge.  So it might, for example, make
a conclusion based on an analogous method suggested by a
similar application of the method using other members of the
categories that the references belonged to.  (From familiar category to
novel specific).

Similarly a conclusion based on a numerical method would be acceptable so
long as the application of the numerical method is not fully defined
for all or nearly all methods and situations.

The rule that some references must have a one to many correspondence seems
to suggest that the program has to be capable of making mistakes and since
it has to do some learning for itself it has to be capable of learning from
its mistakes.

It would be possible to design some unique programming environments.  One
of them might involve gradual shaping.  Again, shaping is an important part
of education so it is something that we would not want to exclude. However,
if the program allows the programmer-user to shape the relationships
between combinations of categories, variables, grammars and output so that
all or nearly all of the combinations of output have a one to one
correspondence to reference then it may turn out that the program was a
programming environment.  Some programmers could actually design something
like this to fool us.

So the program has to exhibit the ability to learn for itself, and it has
to have the ability to learn from its mistakes.

The insistence that the program has some ambiguity of reference (or
whatever abstract term should be used to cover the possibility of a one to
many reference or application of method) is an acknowledgement of the
nature of the problem. This rule is based on the opinion that an AGI
program has to use a component system to attain greater knowledge.  As the
program acquires references these references will be used in many different
ways.  This is true for any kind of data environment/modality.  Lets
suppose that someone was trying to fool us with his novel programming
program which used shaping to shape all data references to specific kinds
of outputs for specific combinations of reference and grammar.  As long as
someone who knew how to use the program was using it he could get the
program to work for some of the problems that he was challenged with.
When naive users tried it however, they would not be able to get it to
work.  On the other hand if other users tried any kind of AGI program, they
might try their best to screw it up and since we are still working on such
a primitive level, they would be able to succeed.  So the best compromise
is to ask the programmer to show that his AGI program could deal with
ambiguity.

Suppose that the program used references of different levels of
generality.  This would arise based on learning from similar situations.
Then a reference (of some level of generality) would refer to different
objects in different situations.  (Or different contexts for the same
object.  And categorical conclusions would be applied to different objects
or even different kinds of objects).  The insistence that an AGI test
designed for a simple AGI program had to work with ambiguous references is
merely an acknowledgement of the nature of the problem and the fact that a
programming environment where a programmer user solved all the problems
that were encountered isn't AGI.

Jim Bromer



-------------------------------------------
AGI
Archives: https://www.listbox.com/member/archive/303/=now
RSS Feed: https://www.listbox.com/member/archive/rss/303/21088071-c97d2393
Modify Your Subscription: 
https://www.listbox.com/member/?member_id=21088071&id_secret=21088071-2484a968
Powered by Listbox: http://www.listbox.com

Reply via email to