From: Philip Sutton (Philip.Sutton@green-innovations.asn.au)
Date: Sun Jun 13 2004 - 10:30:08 MDT
Eliezer,
You have proposed the collective volition mechanism as the way to
achieve FAI. But the computational power (and other technological
power) that would be needed by a FAI to 'read', understand and
extrapolate 6 billion (or more) people would be massive and technically
demanding (if it was a real exercise with meaningful depth).
So it seems to me that the collective volition mechanism can't be
available to an early stage (ie. less powerful) general AI that is this side
of take off. So how do we ensure friendliness in the lead up to take
off??
I have two other niggling doubts:
1. that human collective volition is so complex (with such swarms of
feedback and feedforward) that it is in effectively uncomputable -
even for a super intelligence.
2. That no large populations of humans will voluntarily cede their
autonomy to a super-AI even if it is supposed to be operating on
human collective volition - so the collective volition idea is
likely to be purely acedemic or hypothetical - unless compliance is
imposed forcibly.
I think it would be more productive to assume continuing human
autonomy and to get a constructive working relationship developed
between humans and early-stage general AIs so that the AIs evolve
technically and morally as part of a community of sentients. And so
that humans can get used to general AIs so that they are not suddenly
spooked by their emergence.
I this constructive partnership exists then as the AIs take-off in
intelligence they will be in a position to feed insights and ideas to
humans about how threats to human wellbeing might be more
effectively dealt with. The advanced AIs are likely to get further by
education and persuasion than they are if they tried to impose their
wisdom.
Where coercion is needed to stop humans activing destructively that
should be able to be handled through the normal human governance
processes (as we do now) - albeit made technically much more
effective as a result of the assistance of advanced AIs.
Cheers, Philip
This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:47 MDT