From: Mike Linksvayer (ml@gondwanaland.com)
Date: Thu Apr 11 2002 - 20:46:27 MDT
On Wed, 2002-04-10 at 01:34, Eugen Leitl wrote:
> On 9 Apr 2002, Mike Linksvayer wrote:
> > Apart from its relevance to the potential for an [ab]useful
> > distributed AI, I think the roles speed and intelligence play in terms
> > of getting "intelligent work" done is an interesting question. An
> > uploaded rat running at many times the speed of a physical rat
> > presumably won't get any more intelligent work done. How much work
> > could a IQ300-equivalent AI that runs at 1/1000 human speed get done?
> > Perhaps not much, but it might be extremely interesting work.
>
> I've missed where the 1:1000 time base was coming from. It is a completely
> off-the-wall number, especially if not linked to a specific technology,
> and specific stage of development.
I made it up, in response to discounting of distributed AI because it'd
be S-L-O-W. So I threw a number out there in an attempt scare up some
thoughts along the lines of "Ok, distributed AI running over existing
networks can't happen, but if speed is the issue, at some point our
networks are no longer too S-L-O-W. What is that point?"
> There's a considerable gap between what even current hardware can do in
> principle, and what we are able to do with it with current methods. This
> gap is unlikely to become any narrower, so there's definitely a
> phase-change scenario involved here. A lot of early-stage embarassingly
> parallel tasks are about finding out how to utilize those dormant
> resources. And the network bandwidth becomes an increasingly less
> important issue as the node size goes up.
I see software ("what we are able to do with [hardware]") as by far the
biggest challenge, one reason I don't think protesting that current
hardware is too slow is very useful. If we had hardware 1000x faster
than current hardware right now what could we do with it (along a
critical path to AI) that we _can't_ do with current hardware. NB: we
now have hardware 1000x faster than that available during the mid-80s AI
hypelet.
Mike Linksvayer
http://gondwanaland.com/ml/
This archive was generated by hypermail 2.1.5 : Sat Nov 02 2002 - 09:13:25 MST