From: Eliezer S. Yudkowsky (firstname.lastname@example.org)
Date: Sat Sep 29 2001 - 15:23:24 MDT
Xavier Lumine wrote:
> Such a number is impossible to estimate to any closer than two magnitudes. I
> can only say that such processing power is available today for a reasonable
> price (thousands, not millions, of dollars).
The amount of processing power you need depends on the current stage of
the project. Ironically, the farther the project advances, the less
processing power you need for a given amount of intelligence. In a
perfect world we would start out with Blue Gene and gradually move to a
Beowulf network; the reality, of course, is likely to be the reverse.
But I don't think that you could buy the Final Piece of Hardware today for
a few thousand dollars. The Initial Piece of Hardware, maybe.
Unless it really does turn out that, in ultimate terms, a few extra orders
of magnitude here and there count for very little, and that software is
basically the sole concern. I don't think this is likely to be the case,
though. I think the first AI research will occur on hardware that is
inadequate for real intelligence, but which is adequate to help build the
> The only concrete numbers I can give relate to the possible hardware
> configuration of the cluster which we plan to run the seed AI on:
> 27 boards, each with 1 or 2 Celeron or Pentium 3 processors, at 1 GHz each.
> Not sure what the ops/clock on the respective processors are (4? 6?)
News to me, actually... I'm guessing that this is a reference to hardware
that Dmitriy knows he can get his hands on. The usual figure I toss out
is 32x2GHz some time down the road. For the initial project, you use
whatever you have available, and you look for funding to buy more hardware
when you start running up against the walls.
But remember, seed AI is a long-term project, and the hardware used will
change with time, funding, and Moore.
-- -- -- -- --
Eliezer S. Yudkowsky http://singinst.org/
Research Fellow, Singularity Institute for Artificial Intelligence
This archive was generated by hypermail 2.1.5 : Sat May 18 2013 - 04:00:24 MDT