RE: Hardware Progress: $319/GF

From: Ben Goertzel (ben@goertzel.org)
Date: Fri May 31 2002 - 17:17:57 MDT


> The crossover point calculated was $3M given the original assumptions, and
> the $15M figure is for a 100 million mip cluster today. I project that in
> 2006, the cost for that performance will reach the economic crossover
> point and be $3M.

$3M was a figure Webmind could have afforded at its peak, and that Cycorp
could certainly afford, assuming a good enough reason.

>
> Regarding the marshalling of the required computional power I expect that
> to be most easily achieved via peer to peer over the net - a knowledge
> grid. Given the expectation of beneficial results, computer owners could
> donate unused cycles.

I used to be much more bullish on this option before I had a full AGI
design.

According to the Novamente design (like the Webmind design before it),
though, the most intensive thought processing has really got to occur in a
pool of RAM on either one machine or a cluster of very tightly connected
machines (Ethernet, not cable modem or DSL). Also, dealing with machines
that may go down at any time, or that may divert their processing power to
other things sporadically, is very expensive in terms of overhead (a lot of
monitoring is required, and a lot of redundancy).

Thus, what we need for Novamente AGI is a large dedicated cluster or a
single ubercomputer.

However, there are plenty of "background thinking" tasks that could
effectively be outsourced to a globally distributed P2P network of
relatively weak and unreliable machines. We designed a very general
software framework called "Webworld" to support this, after surveying
existing frameworks for Internet distributed processing and finding them too
narrowly specialized. Only a crude prototype of WebWorld was ever
implemented, though, and no one is working on that project at the moment (my
team is focusing our efforts more narrowly on core AGI and practical apps of
the in-progress codebase).

I admit that this is a Novamente-centric perspective, and other AGI designs
could possibly use a P2P global network of home or office computers as their
primary computing substrate. But I'll believe that when I hear a few
details about these hypothetical AGI designs ;>

Also, I understand that eventually we'll have fast Ethernet equivalent into
everybody's house, and at that point a globally distributed Novamente will
become more interesting. But I'm not sure this is gonna advance as fast as
processor speed...

ben



This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:39 MDT