[extropy-chat] SIAI seeking seed AI programmer candidates

[Eugen Leitl]

On Wed, Jun 02, 2004 at 08:21:29AM -0400, Eliezer Yudkowsky wrote:
> Eugen Leitl wrote:
> 
> >On Wed, Jun 02, 2004 at 06:50:24AM -0400, Eliezer Yudkowsky wrote:
> >
> >>If we cannot do it on off-the-shelf hardware, we should not be in the 
> >>business.  This is not about computing power, never has been.
> >
> >You're being unrealistic. I guess we should be thankful for that.
> 
> Want to show me the math, Eugen?  Real math, calculations describing 
> human-level intelligence and the lower bound for hardware, not silly 
> analogies to biology?  I could be wrong about my own guess.  I'm just 

Any precise calculation must be silly. Less silly is an educated guess.
Most current projections assume linear semi-log plots for future *performance*.
While this is true for some, carefully selected benchmark subsets a spiking
code is not well-behaved. Using Moore as baseline, the disparity grows with each
year. So any projection based on Moor-ish extrapolations is toast, the
farther, the worse.

I think (based on my layman's understanding of current state of the art) 
a connectionist, very probably a spiking architecture is vital. It doesn't need
to be biologically realistic -- then we can just give up.  Though it 
would scale very well with the number of nodes, if done right (spatial
decomposition, 3d lattice or torus physical node connectivity as Blue Gene
does, and maybe Cell boxes will), you can't do it in software on legacy
architecture systems. Connectivity can be emulated in hardware by a 
packet-switched asynchronous network, processing limited to few-bit 
integers (i.e. far shorter than an int, time stamps (network is 
asynchronous) could be deeper) with a low physical connectivity. 
You trade space for speed here, approaching realtime and superrealtime speeds
of biology on a substrate faster but far less connected that biology.
 
You need some ten hardcoded automaton types, probably some hundred, though. As the
discrete response function is smooth, their representation can be compact.
Connectivity weight is probably as simple, or simpler (you might get away
with just +/-/0, but this wastes packet payload, even using relative
addressing, so see few-bit integers above).

Average connectivity is in range of 10^2..10^4, probably around 10^3. Total
number of automata cells is somewhere in 10^11..10^12 range, probably. 

A single instance of above in dedicated hardware is not very far removed from
today. But it is not cheap, and it *is* dedicated hardware. Much depends
about how future commodity hardware looks like -- if we're (un)lucky, we'll
get a close approximation of above off the shelf a decade or two downstream
(see Cell vaporware).

To make matters worse, above is a framework, with most of parameter space
filled with duds. If we're lucky/unlucky, the framework might be
insufficiently flexible, and contain only duds. Then one has to step back a
few steps, and start from scratch.

If it's not all duds, you need lots of educated guesses as points of
departure, and ridiculous amounts of crunch for sufficient co-evolution
rounds. Which is why we very probably need many moles worth of molecular 
circuitry cells to make it work.

Of course, your wild-assed guess is as good as mine. The only way to figure
out what is really required, is to try.

> wondering why you think you can give hardware estimates for intelligence 
> when you claim not to know how it works.  I used to do that too, convert 
> synaptic spikes to floating-point ops and so on.  Later I looked back on my 
> calculations of human-equivalent hardware and saw complete gibberish, 
> blatantly invalid analogies such as Greek philosophers might have used for 
> lack of any grasp whatsoever on the domain.  People throw hardware at AI 
> because they have absolutely no clue how to solve it, like Egyptian 
> pharaohs using mummification for the cryonics problem.

Many orders of magnitude more performance is a poor man's substitute for
cleverness, by doing a rather thorough sampling of a lucky search space.

-- 
Eugen* Leitl <a href="http://leitl.org">leitl</a>
______________________________________________________________
ICBM: 48.07078, 11.61144            http://www.leitl.org
8B29F6BE: 099D 78BA 2FD3 B014 B08A  7779 75B0 2443 8B29 F6BE
http://moleculardevices.org         http://nanomachines.net
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 198 bytes
Desc: not available
URL: <http://lists.extropy.org/pipermail/extropy-chat/attachments/20040602/19c9556f/attachment.bin>