Re: Singularity and AIs (was: We're stuck with each other)

From: Eugene Leitl (Eugene.Leitl@lrz.uni-muenchen.de)
Date: Sun Jan 27 2002 - 03:36:22 MST


On Sat, 26 Jan 2002, Samantha Atkins wrote:

> Of all the curves we care about the software curve is the
> nearest to flat in terms of actual advances and production of
> high-quality tools. In my opinion, it will be software that is
> the most significant technological bottleneck to Singularity.

A lot of Singularity tech is about autonomous, adaptive fabbing, on all
scales. One of the critical components is a sensor-actuator coordination
loop, a subsystem for robust, adaptive realtime control. It could be a
pipeline, but the control loop must be not deeper than 20 ms, or less,
while processing quite complex stimuli. If you look at this in terms of
conventional machine vision, you will have trouble even pushing around two
megapixels worth of stereo frame buffer, even not mentioning processing,
especially complex processing.

So there is a very real hardware problem, namely that 2 GOPS of braindead
sequential operations don't do you a lot of good, if you have to use a lot
of them under realtime constraints. And remember, you never, ever may lose
tracking, or you're wrecking the entire fab, including setup time (a
single shuttle worth of wafers in a fab is worth 20 M$, and it can take
*days* for the process to settle in).

Things become even more difficult, if you want to grow that algorithm,
since you would have to resort to an emulation layer which can easily eat
up two orders of magnitude of performance, because your native hardware
layer is allergic to mutation.

So, don't assume hardware is a given. It isn't.



This archive was generated by hypermail 2.1.5 : Fri Nov 01 2002 - 13:37:36 MST