NSG/ Meeting Announcement (fwd)

From: Eugen Leitl (eugen@leitl.org)
Date: Wed Jun 05 2002 - 13:14:50 MDT


-- 
-- Eugen* Leitl leitl
______________________________________________________________
ICBMTO: N48 04'14.8'' E11 36'41.2'' http://www.leitl.org
57F9CFD3: ED90 0433 EB74 E4A9 537F CFF5 86E7 629B 57F9 CFD3
---------- Forwarded message ----------
Date: Mon, 3 Jun 2002 11:28:42 -0400 (EDT)
From: Fred Hapgood <fhapgood@TheWorld.com>
Reply-To: nsg@europe.std.com
To: nsg@shell.TheWorld.com
Subject: NSG/ Meeting Announcement
Meeting notice: The 02.June.04 meeting will be held at 7:30 p.m. at the Royal 
East (782 Main St., Cambridge), a block down from the corner of Main St. 
and Mass Ave.  If you're new and can't recognize us, ask the manager. He'll 
probably know where we are.  More details below.
<-><-><-><-><-><-><-><-><-><-><-><-><-><-><-><-><-><->
Suggested topic:
Good AI is crucial to the development of NT.  Thus it is 
disappointing at how little progress has been made over the last 
forty years  in finding a set of robust, generalizable, and 
broadly useful AI techniques.   How a skilled professional 
manages to see a path through a complex problem at a glance, 
often even before the situation has been fully explained to him 
or her (sometimes it seems that the less talented people are told 
the better their solutions are), is still very much a mystery in 
any detailed sense.
There is really only one path that is open now to general purpose 
AI -- the "long way around" of brute force simulations: 
generating and comparing multiple, slightly varying, models of 
the problem domain.  While we cannot as yet build a chessplaying 
algorithm that can see the right move without "looking ahead" (at 
least not one that plays well) we can definitely build a pretty 
good machine that gets there by generating and scoring billions 
of positions. While we do not have a program that can deduce a 
new molecular compound from a set of desired properties, we can 
build one that generates and tests millions of formulas until it 
hits on something interesting.  Neural nets and genetic 
algorithms are both variants of this idea.
Brute force simulations have the advantage that for the most part 
the science required to do the programming is known; we do not 
(for instance) have to figure out how our brains work first. 
Their disadvantage is that they require truly astronomical 
amounts of computation, memory, and storage, far more (for most 
purposes) than we can count on Moore's law supplying (at 
practical prices) in our lifetimes.  
Fortunately, it is possible that there is a fix.  Simulations are 
naturally parallelizable in that only the distribution of the 
data units (for any one run) and the collection of the output 
values need be centralized.   Very few computing resources are 
efficiently used. Thus to the degree that unused resources and 
simulation or analysis projects can communicate, huge quantities 
of currently unused resources can be recruited to those projects. 
The most famous case, that of SETI@home, has acquired almost *a 
million years* of CPU time by this means.  There are many other 
examples and more to come.  
For instance currently most Massively Multiplayer Online Role-
Playing Games are not unified, despite their appearance.  They 
are broken up into N copies of the game world, where N is as much 
as any one server can deal with. So when there are 75,000 players 
who think they are in Norrath at a given time - - - I am relying 
on The Harrow Report here -- there might be 25 separate Norraths, 
each independent from the others. If a player on one server tried 
to interact with a player on one of the other 24, he would find 
that the Norath on that other server was not a perfect mirror 
image of his own. 
Sony is exploring whether it might be possible to unify all these 
"shards" (that's what they are called)  by taking unused cycles 
from idle PlayStation2s (a network upgrade is in the works) 
everywhere on the planet.  Some calculations suggest that this 
system would give enough computing power to support more than one 
million players simultaneously, each able to go anywhere in 
confidence that the world it would find there was entirely up to 
date. 
The resources available to support brute force simulations are 
predicted by two exponential functions multiplied together: 
Moore's Law, and the quantity of networked but underutilized 
resources.  While the former is probably a constant, the latter 
seems bound to increase as broadband wireless connections, home 
networks, and embedded devices spread into the population.  It 
seems possible that the huge computational problems lying in the 
path to NT might well find their solvent in a few years. 
<-><-><-><-><-><-><-><-><-><-><-><-><-><-><-><-><-><-> 
NSG founder Chris Fry recently co-authored "Water Programming" 
with Mike Plusch.  According to www.waterprog.org, Water is a 
native Web service programming language that supports ConciseXML 
syntax. It is an open, object- oriented language designed to 
simplify the creation of new Web services. Water adheres to a 
"Learn Once, Use Everywhere" philosophy, in which data, logic, 
and presentation have a uniform representation. 


This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:39 MDT