Re: Something more interesting, please!

Tim Tyler <seemysig@xxxxxxxxxxxxxx> wrote:
Curt Welch wrote:

[materialism]

This is the whole complexity of the topic
and why you see it debated ad nauseam here (and elsewhere in
philosophy).

I understand that some people are not materialists, and think
there's a soul - or some other metaphysical mystery - at the
root of minds.

But AI scientists need pay these folk no more attention than
the evolutionists pay creationists.

Debating creationists is mostly a useless time sink.

Similarly AI practitioners can mostly ignore the anti-materialist
naysayers, and get on with their real work. Hofstadter, Dennett,
etc. put the case for materialism in AI long ago.

It's not as though the anti-materialists are in charge of funding,
and really need a rude awakening.

How about some posts on something interesting:

e.g. will the first AI arise out of a proofreading project, a customer
service shop, a search assistant, a robot project, a NPC in a
game, out of something to do with pornography, or in some other way?

I vote for a government funded robotics project for the military.

Woo, scary ;-)

or: we have a whole WWW of textual training data - how best to use
that to make an AI that does more than proofreading?

You can't understand English to the level needed if you can't ground
the meaning of the ontology to our existence in 3D space/time. Words,
pictures, and videos from the Internet are not enough for an AI to
understand what it all means to the level needed to do constructive
proofreading.

IMO, you could build an /extremely/ useful proofreader from all
that data. ***-loads better than today's spell checkers, anyway.
But proofreading is a rather narrow AI task - you couldn't
hold much of a conversation with such a beast.

AI can produce better spell checkers and
grammar checkers, but it runs out of steam when it can't understand the
content of anything it's reading because it has no grounding. The
information it needs to understand all the text on the internet is not
actually on the internet, it only exists in the heads of the humans who
built the internet.

The only way to get that grounding is to build a learning machine and
give it sensors and effectors connected to the real world (eyes and
ears, arms and legs).

I'm not expecting to build a mind that lives in the real world
directly - that would take far too long, and looks like an
impractical approach to me.

IMO, the immediate aim should be to build a mind that lives in
cyberspace. Such a virtual robot should be able to sense and act -
but not with eyes and arms, but rather with I/O ASCII streams.

Cyberspace is a real and important environment, and I *definitely*
think that the first decent AIs will spend most of their time in it.

It's true that many cyberspace tokens have associated real-world
concepts, that it would be /nice/ to understand in the context
of some applications - but I don't think building robots is
a good answer to this in the short term - since it would be too
time-consuming a diversion. Rather such AIs should figure
out as much as they can from the context, and do the best
they can. If there's some important concept that they
know that they don't understand, they can try asking.

Of course AIs will /eventually/ acquire sensors and
actuators that extend beyond cyberspace - but they /probably/
don't need to do much of that in order to reach the point
where they can collectively redesign, reprogram and
improve themselves - and that seems like the main thing
to aim at. Once we have self-improving AI, it ought to
be able to help a great deal with nailing down molecular
nanotechnology - and /then/ we will be able to build some
proper robots.

Well, I do find it fascinating to wonder what type of intelligent agent
could evolove from only having access to the internet. As you say, the
internet is very much an environment and if you can build an intelligent
agent, what would happen if it's sensors and effectors only gave it direct
access to the internet? How much of all the human information could it
learn to understand in some sense?

It could certainly learn about things like blue skys because there are many
pictures and movies on the internet that would allow it to develop the
concept about that. There's lots of audio on the internet so it could
learn about sounds. But it couldn't learn how making a body movement made
a sound when you interacted with an environment.

Without the help of at least being connected to a virtual simulation, it
would have great difficulty in understanding 3D space and great difficulty
in understanding the 3D space nature of moves of pictures (it it really
could understand it at all). If you can't understand a picture of an
iPhone is a picture of an object in 3D space it will be extremly difficult
for this intelligence to understand what all these pictures and movies were
really about, which means it would have a very hard time grasping what many
of the the words were about.

It would have no understanding of taste and smell, which means ideas
associated with food would be extremely hard for it to grasp.

It's just hard for me to grasp what type of behavior such an intelligent
agent would tend to be able to create.

However, I've also always been interested in the idea of giving an
intelligent agent access to only a virtual tty like interface where all it
could do, is interact with with people though a simple keyboard interface.
I've always thought that such an intelligent agent should be able to learn
to talk and act very human like, even if they had no grasp of the real
meaning of most the words since it would have no concept of 3D space, or
light, or vision, or hearing, or taste, or touch, etc. My guess however is
that it could learn to play the Turing game fairly well (by simply learning
to use diversion tactics if asked about subject it didn't really
understand).

Full access to cyberspace is just a high bandwidth information rich version
of the same problem - except I always thought of my problem as a real time
interaction with humans game and information to cyberspace doesn't require
real time human interaction, even though it's available.

Of course, being connected to cyberspace does give one arms and eyes
already. There are many web cam sites and a small collection of sites that
allow you to manipulate things things the real world (like point a camera,
or operate a robot). So access to cyberspace is only one small step from
giving such a device real access to the real world.

The other idea from above is the idea of giving them access to virtual 3D
worlds. Having a body in a virtual 3D world with high-dimension links into
the brain of a human-level AI could give it the experience it needs to
understand most of human existence. So, with tools like that, maybe an AI
in cyber space alone could actually understand much of what the information
on the internet was all about.

It's all very interesting.

--
Curt Welch http://CurtWelch.Com/
curt@xxxxxxxx http://NewsReader.Com/
.