From: Eugene Leitl (eugene.leitl@lrz.uni-muenchen.de)
Date: Wed Jul 05 2000 - 01:06:58 PDT
Adam L. Beberg writes:
> Noone said the system has to be implemented via trojans and virii. It is
> acceptable to do things legally :)
Code sandboxes don't really work, so your user base has to actively
trust you. You might get a user base the size of SETI@home, or an
order of magnitude more, but it will be a fraction of total available
resources, and the temporal scope of the project will be limited,
given people's attention spans. Worms have no such limitations, but
they are, admittedly, rude, even if they don't destroy information
actively.
> Most areas of the brain are extreemely specialized, with the ability to
> adapt when dammage occurs. And I'm in a constant battle with bad design,
> not centralization.
The hardware itself is anything but specialized. A patch of cortex
looks just like any other. The fact that the system compensates for
damage in most areas but evolutionary oldest should tell us something
important.
> You only need lots of bandwidth for raw data, now knowledge, which is
> generally very tiny. And raw data processing is always done better in
I don't see the point. A simulation of a slab of the cortex, residing
on individual nodes, needs the information of the state of interfaces
to be exchanged, in realtime. This is a lot of bandwidth, and/or
latency (a high bandwidth can compensate for latency, if you use
deeper interfaces transfer).
> silicon then software, much like the human eye, it's a vector graphic
> before it leaves the eye, and only change information is sent.
The retina does indeed do a 1:126 compression, but the encoding is
anything but a simple "vector graphics". (Try reconstructing the
original image from the crossection section of the chiasmata, for an
illustration). And, the retina burns a lot of massively parallel ops
to do that encoding. Luckily, data operation on a bitmap is largely
local, so an array of DSPs can do very well here. This is not so true
for subsequent visual processing "pipeline".
> Well, we have robots that can walk and play volleyball. AI that can have
Do we have robots which can walk and play volleyball? Which can climb
a flight of stairs? Read a fairy tales book to a child? Play chess? A
mean game of Go? Paintball? Thread a twine into through needle's eye?
Wash the dishes? Mow the lawn? Walk the dog? Draw a car from coast to
coast? Do homework? Fill out an IRS form? Vacuum the rooms? Do all of
it, plus more, and learn new skills in the process? Hardly.
We have some narrow specialists. That's all. All of it is hardcoded,
and none of it is especially adaptive.
> a conversation as well as a sign-language enhanced gorrila, or a
> confused person, but linked to knowledge databases so it's more
> knowledgeable then the average result of the public school system.
> Machines that can drive better then humans, and read the road signs,
I haven't seen any robust driving robots so far. Daytime, yes. Not
night and mist.
> unlike ~50% of humans that cant read. Understand spoken speech, take out
> the trash, even fix dinner.
I think you're living in a Ray Bradbury universe. The reality I
inhabit is nothing even remotely like it.
> There isn't a whole lot a machine can't handle now days. In fact I cant
> think of anything I _do_ that a machine cant do better. The machine just
> can't _think_, because it's only programmed to do one thing... for now.
> Since all I do is think (and type), i'm not too worried. If you cant
> think, you will be obsolete in the near future, assuming you aren't
> already but hanging on via unions and ludite fears.
I agree with you on the long term, but it takes a certain attitude to
recognize problems, in the order to be able to fix them. A lot of AI's
past failures has been due to an attitude problem. Do not claim things
which you can't deliver on a reasonably short schedule.
It does really start to piss people off, aftera while.
This archive was generated by hypermail 2b29 : Wed Jul 05 2000 - 02:18:41 PDT