I said:So if you removed all knowledge except that knowledge that referred to a
box or a ball then the text-based program would not be able to figure out that
a "ballbox" was referring to a box shaped like a ball rather than a
cube.-------------------------Well of course if the program had some
information on the shapes of cubes and balls then it might be able to guess
that a "ballbox" was a box shaped like a ball. However, in a developed AGI
program we would expect that a lot of significant information about the how
shapes of things can affect their uses (and definitions) would be kept separate
from the descriptions of various kinds of objects. So, yes, of course an AGI
program could, using only the information about boxes and balls, infer that the
shape was somehow relevant to the new word, if it guessed that the combination
of the two words intended to provide a pointer to possible definitions of what
the word meant. I was just trying to make the point that if concepts were
really primitive and nothing else that could be applied to them had been
defined, then it would be very difficult to make inferences about the concepts
outside of the primitive definitions. It takes a lot of information to
understand something simple. The real problem is not that these inferences
cannot possibly be made (Mike is wrong of course) but that there are so many
possible inferences that could be made that it would easily overwhelm an AGI
program. However, if there was a huge amount of information related to boxes,
balls, shapes and there were other cases that the program could analyze where
two words were conjoined to create a new word based on the features of the two
objects the words referred to, then a few best possible inferences could be
assembled and many of the less likely cases might be ignored.Jim Bromer
From: [email protected]
To: [email protected]
Subject: RE: [agi] A General O.D. (Operational Definition) for all AGI projects
Date: Sat, 11 May 2013 08:26:44 -0400
Mike Tintner said:
Your project must have an E.M. for how BALL + BOX = BALLBOX i.e. you have
to show how with only standard knowledge of two objects, balls & boxes, you
can a) generate and/or b) understand a new, third object, “ball-box” that is
derived from them by non-standard means. In this case, a BALLBOX is a box
shaped like a ball rather than a
cube.------------------------------------------
I am only replying to this as a way to repeat one of my ideas that seems
obvious or commonsensical me.
It takes many 'statements' about a simple idea to understand it. So if you
removed all knowledge except that knowledge that referred to a box or a ball
then the text-based program would not be able to figure out that a "ballbox"
was referring to a box shaped like a ball rather than a cube. And in fact, I
did not realize what Mike was talking about until he made the statement that,
"a ballbox is a box shaped like a ball rather than a cube." So no, an AGI
program would not be able to figure that out without information beyond the
heavily redacted information about balls and boxes. However, once the
definition of a "ballbox" was made, as Mike made it for us, a text only AGI
program would be able to figure it out (just as I was able to figure it out
once I read it,) and use the term intelligibly. And, significantly, if the
text-based AGI program had many statements about different things it would be
able to consider ideas like:
A box that can hold a ball. (That was my first guess.)
A sphere that can hold a box. (That is a simple rearrangement of terms.)
An box that was shaped like a ball.
A ball that was shaped like a box.
A metaphor of something else.
For example, a square line on the ground where balls are put or something
(similar to a "batter's box").
It is easy to see that these could all be generated using computational methods
of rational creativity if the program had general knowledge about many
different things.
The real question is whether or not a text-based AGI program would ever be able
to distinguish what kinds of things words "box" or "ball" referred to without
ever seeing one. I can say that there are human beings who are born blind but
who can use references to things like, "the view of the mountains in the
distance," intelligibly. If the program used terms like this intelligibly
their use would always be removed slightly from our more familiar use of the
terms. But so what? The text-based AGI model is just a step that is being made
to try to discover *how thinking works* in general rather than precise
subprograms that concern the visual shapes of things (as in Mike's unconscious
cherry-picked example.)
Jim Bromer
-------------------------------------------
AGI
Archives: https://www.listbox.com/member/archive/303/=now
RSS Feed: https://www.listbox.com/member/archive/rss/303/21088071-f452e424
Modify Your Subscription:
https://www.listbox.com/member/?member_id=21088071&id_secret=21088071-58d57657
Powered by Listbox: http://www.listbox.com
