From: Rüdiger Koch (rkoch@rkoch.org)
Date: Sun Apr 07 2002 - 14:37:39 MDT
It seems very difficult to me to run an AI distributed like this. Massively
parallelizing neural nets is possible but not as trivial as distributed.net
stuff. If the NN is a spiking network you need to sync it. It doesn't need to
be exactly synced but the different part may not vary in simTime speed much
because timing is what the spiking is all about.
Another issue is latency. The latency requirements to run an Amygdala NN
massively parallel is around 100us UDP roundtrip. This is achievably on a 1GB
Ethernet, Myrinet or a bus system like the SP2. The Internet cannot get lower
than a certain limit because of speed of light limitations. Currently I am
getting around 650ms roundtrip latency to javien2-3.spots.ab.ca
(www.extropy.org) over 17 hops in spite of having DSL.
The only way to run an AI distributed on the Net is to have autonomous agents
which talk to each other with a high level protocol (such as the English
language or a formalized subset thereof). But then it would really be many
AI, not only one.
I am not even sure if this is even worthwile to try - the bandwidth
requirements are quite excessive. The network costs would by far exceed the
CPU/memory costs.
-Rudiger
On Sunday 07 April 2002 22:13, Samantha Atkins wrote:
> Robert J. Bradbury wrote:
> > On Sun, 7 Apr 2002, Alfio Puglisi wrote:
> >>On the contrary, even with a good human model in hand, you
> >>would be hard-pressed to find the hardware to run it.
> >
> > Not really. That was the precise point of my posts on Kazaa DC effort.
> > You could buy access to sufficient processing capacity ~1 petaflop
> > by buying a controlling share of the company (currently ~$3 million).
> > It would most likely be a slower-than real time simulation due to
> > the fact that many of those computers are probably on dial-ups.
> > But as soon as the balance shifts in favor of DSL, cable, optic
> > (company or university), or even 3-4G wireless (wouldn't it be
> > ironic if the AI emerged from our handhelds and cellphones...)
> > you are going to have the bandwidth to ratchet up the simulation
> > clock rate.
>
> There are a lot of buried "ifs" in there. At dialup or even
> ADSL speeds distributed all over the country (or broader) such a
> beast would run so atrociously s-l-o-w as to be useless as far
> as I see. It is difficult to do decidely mundane computational
> task on such a network with adequate performance. Just adding
> more computers does not drastically increase the amount of
> useful work accomplished for all types of problems.
>
> > It is going to become necessary to only install software which
> > has been vetted by a "reliable" (presumably 3rd party) security
> > focused organization and only accept licenses where the software
> > distributor accepts liability for any damage the software might do.
> >
> > We live in very scary times.
>
> Not half so scary as this and certainly not justifying such
> draconian methods. Please do not give aid and comfort to the
> control freaks at large in position to do considerable damage
> today. This is a much greater *immediate* danger to all of us
> than the scenario you paint.
>
>
> - samantha
-- Rüdiger Koch http://rkoch.org Mobile: +49-179-1101561
This archive was generated by hypermail 2.1.5 : Sat Nov 02 2002 - 09:13:18 MST