From: Ben Goertzel (ben@webmind.com)
Date: Tue Jul 31 2001 - 14:24:53 MDT
> Do you think giving a human-level Webmind more resources (not just memory,
> but ops) will increase its raw intelligence, or can this only be done
> through it rewriting its code into new designs?
I think giving it more resources will increase its intelligence. Of course,
some aspects of the system benefit more from increased firepower than
others.
The good news here is that schema learning seems to be a problem that scales
well as one adds more memory and processing power. More is always better --
the rate of improvement of intelligence with addition of memory & processing
can be bumpy, but there's definitely no "ceiling" beyond which adding
firepower doesn't help significantly for schema learning.
(This conclusion refers only to the WM approach to schema learning; I can't
speak about other approaches as confidently. Schema learning is done in WM,
very roughly, using an evolutionary computing approach with inference based
on the system's knowledge as the fitness function. Hebbian learning and
inference based approaches historically have traditionally scaled badly as
compared to evolutionary-style approaches, but maybe Peter and Eli will come
up with clever ways to circumvent the traditional problems with these sorts
of techniques.)
A phase transition in WM intelligence will occur when it has enough
processing power to effectively learn schema of the same size as its various
AI control schema. At this point, through improving schema learning,
additional firepower will indirectly improve *all* aspects of the system's
intelligence. (I'm sure this won't be the only phase transition in the
system's evolution.) (We may somewhere have data that will allow us to make
estimates of how much processing power this will be, but I don't have it at
my fingertips.)
-- ben g
This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:37 MDT