Nobody came back on my suggestion for a much simpler AGI test. Let's call it
the Neo-Maze Test.
You program a robot rover or simulation robot with flexible rules to run
fairly basic-type mazes, including mazes with multiple solutions.
The test is then whether it can run very different kinds of neo-mazes (to be
set by other programming teams) - let's say, mazes with holes-in-the-wall,
large open squares, roundabouts, spaghetti junctions, underpasses, ramps,
traffic lights and traffic problems etc. etc. - for which its rules have
NOT equipped it, but which it does have the basic raw capacity to run.
This will be a test of its ability to fundamentally adapt its approach to an
activity to fundamentally altered environments - to adapt its rules, steps/
moves and recognized paths to its goals. That, I reckon, is the primary
requirement of AGI. (Animals can do it).
From there you can move to ever more complex, and higher-level activity
tests. But setting your first goal as passing a language test is to my mind
absurd.
P.S. Another comparable test would a Video-Game Test, where some gameplaying
agent that is programmed to play something like Pac-Man, will have to adapt
to fundamental variations on that game, with again radical alterations to
the game's maze structure and the type of predators it must deal with, and
treasures it must find.
Or, thinking of the link screen on the Novamente site, you could have simply
a Building Navigation Test - a robot or agent programmed to negotiate fairly
simple rooms with simple furniture arrangements, and fairly simple
corridors - will then have to megotiate ever more complex rooms and building
corridors, ideally the full range of modern architecture. That surely would
be a highly practical test with highly commercial applications.
----- Original Message -----
From: "Derek Zahn" <[EMAIL PROTECTED]>
To: <[email protected]>
Sent: Wednesday, April 25, 2007 2:40 PM
Subject: Re: [agi] The University of Phoenix Test [was: Why do you think
your AGI design will work?]
Ben Goertzel writes:
I don't think there are any good, general incremental tests for progress
toward
AGI. There are just too many different potentially viable approaches,
with
qualitatively different development arcs.
Nevertheless, I wish somebody would try to specify some that are perhaps
not completely general. Without that, the only way to determine whether
any
progress at all is being made is by an analysis of internal structures --
pointing to
a data structure and making claims about its meaning. And the developers
of
the system are notoriously bad at doing this as they are too emotionally
and
intellectually tied to the work.
I wonder at what point our ancesters became "generally intelligent"? Were
humans
of 10,000 years ago generally intelligent? If so, why did it take them so
many billions
of person-years to develop the most rudimentary capabilities that we seem
to expect
our artifical general intelligences to breeze through effortlessly? I
suppose the
real test is at what point an individual from the past would be able to
pass the Turing
test (or some similar thing) if born into our present world and educated
like we were
and I doubt any scientists could make any confident guesses about that.
I think that figuring out a good working definition of general
intelligence and
demonstratable intermediate steps is the single most important missing
piece
of the endeavor.
-----
This list is sponsored by AGIRI: http://www.agiri.org/email
To unsubscribe or change your options, please go to:
http://v2.listbox.com/member/?&;
--
Internal Virus Database is out-of-date.
Checked by AVG Free Edition. Version: 7.5.463 / Virus Database: 0.0.0/0 -
Release Date: <unknown> 00:00
-----
This list is sponsored by AGIRI: http://www.agiri.org/email
To unsubscribe or change your options, please go to:
http://v2.listbox.com/member/?member_id=231415&user_secret=fabd7936
