From: Ben Goertzel (ben@webmind.com)
Date: Thu Apr 26 2001 - 14:22:18 MDT
> I have a hard time grasping this. It seems very unlikely to me,
> because I don't
> think we're close to the hardware required, for another 5-10
> years at a _minimum_,
> even with expensive supercomputers.
...
> If you think that hardware that is currently available to you is
> sufficiently
> powerful enough for real AI, I hope you'll explain how this can
> be. This was
> partly the point of my questions last week about Moravec and
> Kurzweil's estimates,
> and James Rogers had a lot of important and interesting stuff to
> say about it.
> None of that discussion, though, pointed me in the direction of
> thinking that
> all that's needed is a couple of years of programming by 30 people.
All that discussion was based on comparing the computing power of existing
machines to the computing power of human brains (as estimated very very
roughly, since no one really understands how the brain does computation)
I think this kind of discussion is not all that relevant, although it's
certainly more interesting than discussing Pamela Anderson's boobs or
what-not...
My estimate is based on this. I have an AI system with about 7 different
"AI modules" all implemented on the same common dynamic semantic networks,
itself implemented on a distributed Java agents platform (now becoming a
mixed Java/C agents platform).
We had a network of 30 quad-processor PC's with 4GB of RAM each. Using
about 3-10 of these (depending on the module), one can get any one of our AI
modules to do pretty interesting things. According to our theory of how the
mind works, when we can run all these modules together, we'll have emergent
intelligence. Exactly how smart it will be is hard to tell -- there's not
really an objective way to compare human and nonhuman intelligence.
I don't mean to imply that if we had enough machines we'd have a thinking
machine working. There are too many parameters in the system and we need to
tune them all to work well with each other (what we call the "homeostatic
control" problem). And our Java system is too slow to let us run all the
needed parameter tuning experiments. So as I've discussed we're recoding
the "cognitive core" of the system in ANSI C at the moment, as an
experiment, currently a very promising one.
So, of course, I don't have a convincing proof that I can produce a real AI
in a year or two with a small team on a practical amount of hardware. The
only proof to convince a skeptic would be a complete working system. But my
intuition on this is grounded in practical experience with the components of
an AI system, rather than based on high-level abstract comparisons of
computers with human brains.
-- Ben
This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:36 MDT