RE: AI and Moore's Law redux

From: Ben Goertzel (
Date: Sat Jan 26 2002 - 13:00:41 MST

There is a distinction between explicit and implicit knowledge that
has to be made here. N items in memory have 2^N combinations, and so
there *is* exponential increase of knowledge implicit in linear
increase in memory.

However, when knowledge is reprsented only implicitly, it takes
processing time to sort it out, so one has a time-space tradeoff.

This very crudely put point is really at the heart of AI design.


> -----Original Message-----
> From: []On Behalf
> Of Alan Grimes
> Sent: Saturday, January 26, 2002 11:29 AM
> To:
> Subject: Re: AI and Moore's Law redux
> Ben Goertzel wrote:
> > For sure, you need a lot of RAM. You need a lot of knowledge readily
> > accessible -- not just atomic bits of knowledge but combinations of
> > and relations between atomic bits of knowledge.... And each of these
> > bits of knowledge needs to be continually re-configuring itself in
> > terms of the new bits of knowledge that have come into the mind, and
> > the changes in other bits of knowledge in the mind. This does require
> > a lot of memory and processing power, and my strong sense is that
> > it ain't gonna run on a Commodore VIC-20.
> A dude on has proposed a learning system that
> would express what would ammount to a LOGARIHMIC growth function of
> memory unilization as a function of knowlege.
> The function would be something like:
> [Memory required] = LOG_k([Total Knowlege])
> I can't say I know what K would be; It would likley depend on the design
> of the system employed.
> His proposal was based on a notion of "abstraction" where a higher-level
> abstractions would be specified, like FORTH words, by simpler
> abstractions. As these higer level abstractions would overlap to a great
> extent, the cost of additional abstractions would decrease as more base
> abstractions become available.
> In other words, for a linear increase in memory, an exponential increase
> in knowlege should be possible.
> The process for automating this would probably solve the translation
> problem.
> --
> <my website.

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