--- Bob Mottram <[EMAIL PROTECTED]> wrote:
> > SIAI has not yet solved the friendliness problem.
> 
> I've always had problems with the concept of "friendliness" spoken
> about by folks from SIAI.  It seems like a very ill-defined concept.
> What does "friendly to humanity" really mean?  It seems to mean a lot
> of different things to a lot of different people (observer relative).

Eliezer S. Yudkowsky has written a fairly precise definition, but again this
is not a solution.  http://www.singinst.org/upload/CEV.html

We should not ignore the problem, but that is precisely what we are doing. 
Once machines are smarter than us, there will be an intelligence explosion and
humans will have no control over it.  It will be directed by evolution.

> > A human brain sized neural network requires about 10^15 bits of memory and
> > 10^16 operations per second.
> 
> Direct comparisons between computing speed and brain activity I also
> find problematic.  People often quote numbers like this without having
> any idea how they were arrived at.  As far as I can discern all roads
> lead back to Moravec, who based his figures upon the retina observing
> a TV screen, and admitted that this was a very wobbly estimate
> potentially subject to a wide margin of error.  I think until the
> "essential function" of a neuron is known it's really hard to make
> direct comparisons between what computers do and what brains do.

My estimate is based on 10^11 neurons, 10^15 synapses, and an information rate
of 10 bits per second per axon.  The number of synapses per neuron is based on
studies by the IBM Blue Brain project (8000 synapses per neuron in mouse
cortex).  In most neural models, the information carrying signal is the firing
rate, not the individual pulses.

But you are right that it is a crude estimate.  Cognitive studies of long term
memory estimate 10^9 bits.  That is also about the quantity of language input
by an average adult since birth.

One reason for the wide uncertainty (10^6) is that we don't really understand
how the brain works.  Another is that machines are going to be doing different
tasks.  Their purpose is not to behave like humans, but to serve humans (at
least initially).  Many of those tasks (like arithmetic) don't take a lot of
computing power.

The message posting service I have proposed does not address friendliness at
all.  It should be benign as long as it can't reprogram the peers.  I can't
guarantee that won't happen because peers can be arbitrarily configured by
their owners.


-- Matt Mahoney, [EMAIL PROTECTED]

-----
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=8660244&id_secret=75050384-f1d45d

Reply via email to