On Sat, 15 Jul 2000, Emlyn (onetel) wrote:
> > I wrote:
> > >To get the hardware "cheap enough", you need a driver for the research
> > >required to quickly advance the curve on what is likely to be highly
> > >special purpose hardware.
> I'm not sure that I agree that the hardware will need to be all that special
> purpose. You're basically looking at massive massively parallel hardware,
> which has got to be where hardware is heading in any case.
Not completely true. Programming massive parallelism is not a mainstream
application. Human minds can't think that way so writing the software
is very difficult. I would argue massive parallelism will remain a
special purpose application where it is almost always designed as
a model or simulation of some real world process (weather, buckytube
bending, protein folding, etc.). Either that or it will be parallelism
for doing identical stuff in multi-threads (e.g. web serving).
While general purpose hardware can simulate neural nets, to get to the
real time or faster time scales that Robin requires, it has to be
special purpose (at least currently). Ultimately it isn't computation
the brain is good at, it is evolving clever interconnects. Almost
*all* current hardware devices (transistors, FETs, etc) and communications
protocols are not optimal for brain architectures. You need to change
1-to-1, 1 to few or 1-to-many strategies into many-to-many. As
interconnected as the entire WWW is today it is still much less
than a single human brain.
> Is it more likely that we'll get far faster massively parallel systems (at
> cheapish prices) first, or that we'll be able to conquer the problems of
> high bandwidth connections between human & machine, an area of technology
> that could barely be said to have hatched?
We will have special purpose massively parallel systems (I would love
to know what IBM thinks the market for Blue Gene is). I suspect they
will not be particularly cheap (they have to sell 200 @ $1 million each
to recover their investment at a decent profit). We will start with
low bandwidth human-machine communications (think of the retraining you
do to scribble on a palm-pilot) that will jump significantly when it
becomes feasible to implant devices into which neurons can grow
that are linked to exo-GHz transceivers (think cell-phone or
Borgian exohardware). With human "eye" "input" and an equivalent
optic-nerve sized "output", you have the foundation for high bandwidth
So I see them both developing on separate paths, but the market size
for the massively parallel hardware is likely to be much smaller than
the market for high bandwidth exolinks. I can't win at Tetris anymore
because the keyboard will not respond as fast as I can think. Think
about the protein folding market vs. the game playing market.
This archive was generated by hypermail 2b29 : Mon Oct 02 2000 - 17:34:37 MDT