On Sun, Jul 01, 2007 at 12:44:09AM -0700, Tom McCabe wrote:
> > They also need knowledge, which is still largely
> > secret.
>
> Knowledge of *what*? How to build a crude gun to fire
> one block of cast metal into another block of cast
> metal?
How about gas centrifuges, materials resistant to
UF6, casting uranium metal; or build nuclear reactors
and separate transurans via ion chromatography?
(And if you thought processing uranium is hard, boy
will plutonium drive you to exasperation).
> So? The number of people who could construct a working
> radio if you sat them down at a table and gave them a
> pile of spare parts is very small. This does not mean
A radio is a simple device. So is a nuke, but an AI is not.
> that it requires some kind of special expertise to
> Google up some basic electronics knowledge and order a
> soldering gun, circuit board and other necessary
> tools.
Right. So why can't we just Google up a supercomputer cluster,
and download the rest from Sourceforge. Or, write it
from scratch, all done by a pimply teenager in his parents'
basement? Doesn't sound that realistic, doesn't it.
> > We still need some 20 years to come into that range.
>
> Does anyone know how much computing power you need to
More importantly, how much computing power do you need
to bootstrap an AI? I don't know yet, but it's some
six to nine orders of magnitude more.
> run AGI? Of course not, nobody has ever built or even
> described a coherent, complete blueprint for building
> AGI.
Really? How about mapping by co-evolutionary-algorithms the
parameter space of integer automaton networks and their
morphogenetic genome driving embodied agents in a superrealtime
reality simulator, selecting for task accomplishement?
That's a plan of four lines, and you know, I have a hunch it
would work. And you wouldn't even need all the ops the
biosphere has been crunching for GYrs, since you take off
with an educated guess, and a very monomaniac fitness function.
> > > catch up to these supercomputers, stopping it
> > would be like trying to
> >
> > Add another couple decades.
>
> You know, there is this thing called the "Internet",
> which any script kiddie can exploit to get additional
> computing power.
To recap, you're at that stage where supercomputers
are barely enough to bootstrap AI (some 20-30 years
from now), and you assume that a crypt kitty can 0wn
a cluster (it's hard, but it can be done, at least
for a brief while), and do something meaningful with it?
Pardon if this strains my credulity a bit.
And you do realize the rather critical difference of
a signalling mesh with <us latency and internode
bandwidth enough to connect major continents? It would
be interesting to see what you could do with the whole of Asia
full of millions of PS3 on FTH broadband, but the current topology
(a tree, not a mesh) is going to put some severe clamping
on your style. I do think it's a real possibility, and
we should get diagnostics and segmentation hardware
into the network fabric.
> > > stop file sharing and software piracy.
> >
> > Everyone knows how to wripte a P2P application.
>
> Do you know how to write a P2P application? Do you
$/=$_;$,=shift;$w=$a=shift;$k{+shift}=1;socket S,2,1,6;bind S,&a;for(listen
S,5;$SIG{ALRM}=\&i;m! (\S+) ([e-i])([^/]*)/!s&&($k{$w=$1}=$,eq$`)&&&$2){alarm
9;(accept(C,S),alarm 0)?read C,$_,1e6:($_="$, $a f".shift)}sub i{}sub t{socket
C,2,1,6;$k{$w}&&=(connect C,&a)?print C"$, ".pop:0;close C}sub h{t"$_ i/"for
keys%k}sub a{$w=~/:/;pack'CxnC4x8',2,$',split'\.',$`}sub f{$w=$_,t"$1 $3/"for
keys%k}sub e{open C,'>',$3;print C $'}sub g{open(C,'<',$3)&&t"$a e$3/".<C>;&h}
It's not my code, and it's bare-bones, but I should be
able to do about that or better in a month or two starting from scratch.
There's some pretty hairy stuff in BitTorrent where MNet
didn't quite reach, but MNet had a very good engineering
basis going for it.
> even know how to program? I do know how to program,
I don't know what knowing to program has to do with AI
(it's rather specialist physics, neuroscience, hardware design,
or at least large-scale number crunching, which is quite
domain-specific).
But, yes, I do get by, enough to pay the bills.
> but I have no idea how to write a P2P application; I'd
> have to figure it out as I go.
Good. Do you think you can figure out how to do AI
given enough time, in about the same manner you can
do for P2P?
And I don't have to tell you that studying nuclear
engineering will land you on a list for life, and if
you start looking for information online on nuclear
weapon design, you will make an even more select list.
Human experts don't come from a vacuum, and don't
operate in a vacuum. Assume that MIBs will be
breathing down your neck by the time you've done
your initial hiring, assuming AI is a watched
technology. It's not, not yet.
> > Nobody knows how to
> > build an AI. If it's a large-scale effort the
> > knowledge can be controlled
> > for a long time.
>
> How exactly is anyone going to *get* the knowledge in
> the first place? Physicists had to actively lobby the
> government to start the Manhattan Project, which had
> immediate military applications and had *the physics
Intelligent weapon control and battle management has very immediate
military applications. DARPA funds some of the best stuff
there is. I would put the Urban Challenge right at the
cutting edge; but I'm no AI expert.
> already tested in the lab*. A nuclear physics reaction
> is scalable; you can do the reaction with one particle
> before you try it out on a whole mass of them. An AGI
Of course it's scalable, look at ANNs.
> is not scalable; you cannot get anything interesting
> to show up until you already have a mostly-functioning
> full-scale system.
That's not quite correct, though there are some new effects
when you approach criticality.
--
Eugen* Leitl <a href="http://leitl.org";>leitl</a> http://leitl.org
______________________________________________________________
ICBM: 48.07100, 11.36820 http://www.ativel.com http://postbiota.org
8B29F6BE: 099D 78BA 2FD3 B014 B08A 7779 75B0 2443 8B29 F6BE
-----
This list is sponsored by AGIRI: http://www.agiri.org/email
To unsubscribe or change your options, please go to:
http://v2.listbox.com/member/?member_id=4007604&user_secret=7d7fb4d8
