Novamente Uber Alles?

From: Ben Goertzel (
Date: Sun Apr 21 2002 - 17:24:02 MDT


Just in case anyone on this list actually believes an evil Novamente is
going to conquer the world in the next couple years (or hours ;), I guess I
should make some cautionary remarks...

If things work out well, I think it's plausible we'll have the current
Novamente design *completely implemented* in 18 months or so. It could end
up taking double that long, however, for a couple reasons: team members
spending nearly all their team working on practical applications of the
partially-complete system; or unforeseen difficulties in "tuning" (i.e.
getting to actually function as intended) various system components.

What we will have then is, hopefully, not just a bunch of AI code, but a
baby mind. Teaching this baby mind anything may take a long time. And the
teaching process may reveal significant design shortcomings, leading to more
& more work. It is rather hard to predict how time-consuming the teaching
of a baby digital mind will be, since no one has yet seriously attempted
such a thing.

I do believe we're actually on the path to creating a seed AI. Of those who
are familiar with the system design, some agree with this assessment and
some disagree. (The members of the Novamente engineering team are on the
"agree" side, not surprisingly; the Webmind folks who were more skeptical
moved on to other things after the paychecks stopped coming in!) Right now
there is no strong proof that we're on the right track, so we're going to a
large extent on our collective intuition. Our practical experimentation
with Webmind was mainly been in the vein of "datamining", not in the vein of
teaching an autonomous system how to cooperatively experience the world.
The same is true of Novamente, so far. We are naturally very eager to get
to the "baby mind" phase, but there's still a bunch of implementation in the
way. We view pursuing these datamining apps (currently in the
bioinformatics domain), not only as a way to generate revenue and as a
service to the world, but also as a good way to tune many of the system's
cognition algorithms on tasks that are difficult, but easier than autonomous
experiential interactive learning.

As Novamente is being coded for efficiency in a very aggressive way, I'm
fairly confident that this time around, when we do complete the
implementation, we'll have a system that can be tested and taught. The
Webmind AI Engine was much more complete than Novamente is now, when Webmind
Inc. folded in March 2001, but its inefficiencies were such that its most
advanced intended abilities could not be pragmatically tested in a really
significant way.

-- Ben G

This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:38 MDT