Re: [SL4] Parallelism in AI (WAS: _New Scientist_)

From: Brian Atkins (
Date: Mon Jul 03 2000 - 15:05:35 MDT

I think you should take into account that people are working on ways to
get Beowulfs and such to act like a single machine. At first it is projects
like Cosm and Multix (sp?) that are merging the memory and filesystems for
you behind the scenes. People are working on ways to also have single
processes automagically take advantage of parallel CPUs (see Tera Computing
Inc.). This is all very good for us since taking advantage of the cheapest
computing will allow us to reach human-equiv hardware much sooner than
if we have to wait for single processors/machines to reach that point.

"Eliezer S. Yudkowsky" wrote:
> Brian Atkins wrote:
> >
> > "Eliezer S. Yudkowsky" wrote:
> > >
> > > Artificial Intelligence at any given point. Also, I deleted a written
> > > and edited section on distributed processing from "The Plan to
> > > Singularity" because I decided that distributed AI wasn't workable.
> >
> > Is this due to bandwidth/latency issues? I mean, likely any real AI is
> > going to be running on some kind of parallel machine with a distributed
> > architecture like a Beowulf or the ASCI series of supercomputers, right?
> > The only difference between those and computing over the Net is the
> > internode link characteristics.
> Once an AI gets up to the stage where it can handle the level of
> code-rewriting ability needed to usefully absorb parallel processors -
> which isn't *necessarily* all that much, comparatively speaking - we'll
> have the option of running it on parallel processors. Before that
> point, the AI will only be able to run on multiple processors to the
> extent that the human programmers, i.e. me, can figure out cognitive
> methods that employ parallelism. AI will be able to usefully run
> distributed (as opposed to parallel) to the extent that multiple
> instantiations of the AI can accomplish multiple tasks or subtasks and
> then rendezvous, or to the extent that autonomous tasks can be farmed
> out, calculated, and returned with reasonable latency.
> In other words, a Beowulf network is useful in two ways. You can either
> exploit it for truly parallel cognitive tasks (this might cover quite a
> lot of territory), or you can run multiple instantiations of the AI. In
> either case, though, it seems likely that any given instance of the AI
> will maintain a serial, non-parallel, non-distributed ego. AI doesn't
> become native to parallelism until it can understand its own code.
> "Code", not "architecture" or "design" or "philosophy", but it does
> still have to understand the code.
> The original version of _Plan_ relied on the brute force of a planetary
> distributed network to run the AI. I'd like to point out that this
> isn't totally impossible, in the sense that if it was do-or-die we might
> be able to do it. However, it added a rather significant obstacle, with
> not much benefit over-and-above private supercomputers, so I took that
> out.
> --
> Eliezer S. Yudkowsky
> ------------------------------------------------------------------------
> Free, Unlimited Calls Anywhere!
> Visit - click below.
> ------------------------------------------------------------------------

Free, Unlimited Calls Anywhere!
Conference in the whole family on the same call.
Let the fights begin! Visit - Click below.

This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:35 MDT