Simplifying AI

From: ben goertzel (
Date: Tue Apr 02 2002 - 09:46:30 MST

Eliezer wrote:

Ben Goertzel wrote:
> Actually, our estimate of the size of the problem of creating an AI
> in mid-2000.
> Throughout 2001 we kept discovering how to reduce various mind functions
> special cases of a moderately small set of mind functions, in what I
> is a pragmatic way.

Okay, this sounds very exciting, but can you be more specific? Which mind
functions? What special cases did you reduce them to? Special cases of
what? Are you talking about a wide variety of high-level behaviors
by a bounded set of interdependent subsystems, or are you talking about
aspects of the subsystem design surprisingly turning out to be a set of
special cases rather than big general cases?

Basically what I mean is that in the Webmind AI Engine we had a huge number
of specialized mind modules, and we now have a much smaller number, used to
perform the same functions as the large number used to.

I'll give just a few examples. Unfortunately, to give a complete set of
details, or to get into these in detail, would require a load of detailed
background in Novamente design.

1) A (we hope!) passably efficient way to do "linguistic feature structure
unification" (a kind of language parsing) as a case of general logical
unification... this basically eliminates a separate NLP module, reducing it
to a set of special parameter settings for general cognition methods... as
it should be...

2) Complex inference (inference dealing with quantifiers, variables, etc.)
is reduced to "simple inference on inheritance links" + "complex
procedures". This reduction is carried out using combinatory logic. I
suspect the human brain carries out a similar reduction in a very different

3) Causal inference is reduced to "predictive implication" plus a special
case of "procedure learning" (the latter being the learning of "plausible
causal mechanisms")

4) Association-finding is reduced to simple probabilistic inference plus a
Hebb rule variant

5) Goals and feelings are represented as simple wrappers around generic
CompoundRelationNodes (representing complex combinations of simple


Overall, what we found is that a lot of mental functions are obtainable as
combinations or specializations of other mental functions, but not in
immediately obvious ways.

-- Ben G

This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:37 MDT