From: Eugen Leitl (eugen@leitl.org)
Date: Sat Apr 13 2002 - 04:26:30 MDT
On Sat, 13 Apr 2002, Colin Hales wrote:
> I've been following this thread with interest. Here's another link to a
> group that appeared to get all enthused
> http://www.cybernetics.demon.co.uk/VBM.html and then did a few sums. They
> concluded that they'd need 700 PCs to implement a human brain that operated
> at 1/5000th the speed of a human. After that not much seems to have
> happened. It looks like they were overwhelmed.
I've just skimmed over that page. Even not looking at the numbers (which
they've probably copied from Kurzweil or Moravec), they've got a few
things wrong. The most glaring problem I see with their approach is that
they're using rotating magnetic bits for state storage. Even if they would
get the disks to mostly stream (a major assumption), this is many orders
of magnitude away of what modern chipsets achieve to affordable (DDR
SDRAM) memory (and please try to make it stream, at least most of the
time). The rule in numerical application is that once you start to swap
you lose. So, you shouldn't. It is very possible to make a completely
solid-state cluster, using ~1 GHz CPU ~1 GByte RAM, and only use a big
hard drive/100 nodes or so to record the state snapshouts (hourly, or so),
in case there is a power blackout (individual node failure will be caught
by the self-healing nature of a distributed neuronal system).
They're using 4 FastEthernet NICs/node. It is very possible to build a 3d
lattice out of this, if using a diamond lattice. It's also very possible
building a cubic primitive lattice if intercalating the nodes with cheap
8-port switches (in fact, one should probably go to higher port numbers by
now, since 16 and 24 porters with full backplane bandwidth have become
very affordable, FNNs might be useful even for such small switches). This
would scale to about 10 kNodes, though here you'll get very high
operational costs due to power and air conditioning requirements, plus a
number of operators. Small shops are restricted to about 64 nodes.
> My own feeling is that the massive interconnectedess (large
> associative memories resonating with each other) and learning needs
> preclude any realistic widely-distributed WWW type mind (the MURG
Look at the hardware in your head. It works, doesn't it? Local bandwidth
is cheap and plentiful, here's your recipe for spatial tesselation.
> Amygdala may work, but it's not doing much cognition) - at least as
> far as things go at the moment. If you want to implement a mind then I
> think it may have to be a monolith at first.
I'm not a big fan of Grady Booch, but he said "every complex working
system evolved from a simpler working system" which strikes me as true.
> Risks from AI would stem from a monolithic AI sequestering distributed
Monolithic AI doesn't happen.
> resources to embark on self enhancement. Eg someone has a prototype,
> non-working nano-assembler shared on the WEB for group research reasons. The
> AI uses backgrounded modelling a la folding@home on a million PCs to fix it
If it's a superhuman AI, it can figure out a lot from first principles.
> up and then uses it to remake itself by pretending to be a human to the
> local guy with the machine. Damn wicked these AIs!
>
> Back in the real world - In really don't think any standard CPU will do it.
Right now, nothing will do.
> Being an electrical/electronic Engineer, I see maybe a special board using
> supervised FPGAs in dozens of 2000mm high, 19" racks filling rooms initially
FPGA is good, but there are better ways to organize the silicon real
estate. Have a gander at http://www.cellmatrix.com/
> (I think, If my calcs are right), connected in as massively parallel way as
> possible. One can emulate and debug the boards on a PC, but ultimately it
Crossbars don't scale, so you have to settle for a hypergrid, or
hyperlattice.
> becomes a hardwired monolith. If desktop PCs evolved from that massive valve
> based 'ENIAC' or was it 'UNIVAC' (?) back in the 50's, this FPGA based mind
> is 1950's version of a future desktop version that would probably arrive
> very much sooner.
>
> Well that's where my brain is at on the issue, FWIW.
The hardware will come, way before we're ready for it.
This archive was generated by hypermail 2.1.5 : Sat Nov 02 2002 - 09:13:29 MST