From: Eugen Leitl (eugen@leitl.org)
Date: Mon Dec 30 2002 - 07:36:21 MST
On Mon, 30 Dec 2002, Rüdiger Koch wrote:
> This wouldn't be very practical - a global network has latencies with
> lower bounds limited by the speed of light. That would make a global
Yes, but all the hardware on the local loop is tightly coupled. Here's a
clear analogy to the wet neuron, which is also mostly locally interacting
over a ms time window. The switches introduce negligible latencies to the
raw wire speed. You hardly see them even with userland pingpong code.
Latencies will go down even for WAN type networks as we go to 10 GBit/s
hardware, which can (and at some point necessarily will) be geographically
routed. A photonically geographically switched ~40 GBit/s optical network
of the future doesn't introduce latencies other than essentially
relativistic. A ns is worth some ~1 m, a us some km. Given that high-order
processes of people involve seconds on a substrate that is liter in size
and ~100 m/s signalling constrained it's clearly more handicapped than a
city-size assembly which signals over 2*10^6 times faster. Besides,
latency is not very relevant, as the pipe acts as a FIFO even now. You
just pack your slices encoding spikes (to address, global tick when
issued) into packets and pack them into the pipe, and do the reverse as
the pipe is running full duplex. You don't have to twiddle your thumbs
until the response arrives, so it's all clearly nicely decoupled.
Given that it takes a very fat machine to saturate a GBit/s pipe
sustainably while not doing much else besides the networks are clearly not
today's bottleneck. Sure, current phone wire kBit pipes to the end user
don't look very hot, but there's technology in the pipeline to address
that. There's about a GBit/s worth aggregate wireless bandwidth within a
(small) cell, not counting line of sight laser and fiber. We have people
deploying sub-ms enduser latencies over GBit/s Ethernet fiber backbone.
That's today's technology, out in the field already in some blessed
patches of the geoid.
> mind an extremely sluggish mind regardless of it's architecture.
> Remember our talks about Amygdala clustering and SP/2 switches,
> Firewire, Myrinet and SCI links? I guess the physical size of an AI
Given a 1 kCPU Beowulf using Myrinet it's clear the bottleneck is CPU. It
takes seconds to even stream through a GByte worth of core, not counting
semirandom access and doing I/O, while crunching some numerics heavy code.
The fatter the node, the less load is there on the network. It's the usual
volume/surface scaling ratio which holds true for relatively low
dimensionalities (N<10, very true for N=3 which is the implementation
layer for our wetware and hence sufficient for an in machina model, though
memory accesses are temporally rather flat even for high dimensionalities;
node-locally).
> can't be much bigger than the maximum cable lengths of these network
> types. Else you'd get a society of multiple AIs, a sluggish AI or
> nothing at all. To get a society you'd need to find collections of
> machines that are linked with high bandwidth / low latency such as the
> racks at large ISPs.
Currently, that would be a rogue code's chiefest point of interest.
Especially, that internally the traffic is unmetered, though a sudden
traffic spike and a worm-like pattern might trigger some specific
safeguards. Demo codes for scanning very large networks in seconds exist.
As the ISP node landscape is homogenous, you can 0wn it in seconds through
a single vulnerability. Models of realtime global worms (minutes globally)
exist as well.
Depending on the payload riding the worm things could be rather quick.
> There really is no need to put consciousness / evolution issues into a
> TCP/IP stack. You want to implement that *skyhigh* above of any
A worm would just break through to highes privilege and run directly,
allowing OS some time slices (if it be stealthy) or killing it altogher
(fast burn, sacrificing stealth or speed).
> application protocol of your liking, such as Corba/IIOP, SOAP, MPI or
> whatever. No job for the IETF. There is no consciousness in our neural
> communications, too (unless you believe Penrose/Hameroff ;)
No one believes Penros/Hameroff! :)
This archive was generated by hypermail 2.1.5 : Wed Jan 15 2003 - 17:58:57 MST