[ExI] Smallest human-equivalent device

[Eugen Leitl]

On Thu, Oct 10, 2013 at 11:35:56PM -0400, Rafal Smigrodzki wrote:
> On Wed, Oct 9, 2013 at 3:37 PM, Eugen Leitl <eugen at leitl.org> wrote:
> > On Wed, Oct 09, 2013 at 12:16:05PM -0700, Keith Henson wrote:
> >> > From: Rafal Smigrodzki <rafal.smigrodzki at gmail.com>
> >>
> >> > How many Eugens could you fit in the head of a pin (make it a sphere
> >> > 3/16 inch diameter)?
> >>
> >> 27 years ago Eric Drexler worked this out and got around a 10 cm cube
> >> "volume of a coffee cup" for a human capacity hardware.  With enough
> >> power and cooling it would run a million times faster than a meat
> >> state human.
> >
> > Nanosystems uses an engineering analysis based on diamond rod logic.
> > Deliberately conservatively, in order to be easy analyzable. There
> > are a few problematic assumptions there as well. So it is an answer,
> > but not an exhaustive one.
> 
> ### I am surprised that the estimate is so, well, bulky. 1 liter
> volume is very close to actual human brain volume and I simply don't
> believe our brain is anywhere near the limits of miniaturization.

I agree. However, there's a widespread tendency to underestimate
what evolutionary-driven biology has managed to accomplish in 
a few gigayears. A synapse is pretty damn small

http://www.nature.com/nrn/journal/v12/n7/box/nrn3025_BX2.html

Characteristic sizes of the synaptic complex

Synaptic active zone diameter: 300 ± 150 nm
Synaptic vesicle diameter: 35 ± 0.3 up to 50 nm
Synaptic cleft width: 20 ± 2.8 nm
Number of docked vesicles: 10 ± 5
Total number of vesicle per synaptic bouton: 270 ± 180
 
> There should be large improvements just from removing metabolism from
> the brain. Human brain already does a little bit of that: Neurons

Metabolism is dual-use here, because the elements are active.
You can complain about hydration, but diffusion is very efficient
on microscale, and there *is* active transport. We can complain
about homeostasis, but hardware configurability depends on
the same mechanisms. In nanoscale solid-state you're pretty much
limited to arrays of static elements with very low intrinsic 
connectivity, so you have to state the network layer in terms
of such syntax. Look at 10^4 connectivity, and look at the
space it takes. There are different computational paradigms
than this, but they're incredibly complex and opaque to
top-layer us. Useful for ALife from scratch, but very hard
to compile digitized neuoanatomy/connectome into such a wildy
different paradigm.

> offload a lot of their metabolism (i.e. energy generation) onto glia
> through lactate exchange. Instead of using a lot of real estate to

Glia are far from being just glue, so anyone who thinks glia can
be thrown away will experience an unexpected surprise.

> extract all chemical energy from glucose, neurons do a quick-and-dirty
> glycolysis and let the glia pick up the pieces, allowing the neural
> cytoplasm to do more computationally relevant chemistry, such as
> processing of neurotransmitters, adjustment of synaptic strength, etc.
> 
> A designed neuromorphic device would be fed energy in a highly
> computation-friendly form (DC current, light) just like today's
> computers rather than using chemical precursors like our brain does,
> and that alone should bring the volume down by a large percentage.

100-1 cm^3 vs. ~1400 cm^3 is a very large percentage. The exact
number is not known, because the mode of computation is very
different. The only way to know for sure is to run benchmarks
on your wet/hardware targets, which you can't. A way to assess
would be to compare ALife and wet life solutions to the same,
simple problem. E.g the retina or cochlea would be a good playground.