From: Jeff Bone (email@example.com)
Date: Sat Dec 08 2001 - 15:38:47 MST
Clearly I've got some issues with the rather fuzzy definition of Friendliness
that's been offered, and the differences between the more rigorous definitions
and the use of that term in practice in some of the communiques around these
parts. Let me offer up a strawman definition that I think is, perhaps, more
rigorous and useful.
A "Friendly" SI is one which strives to ensure an optimum
distribution of its constituents across a "satisfaction landscape."
Okay, so what does that mean? First, constituents: the constituents for any
Friendly SI are any other intelligences (self-aware, self-interested, active
things) which exist and act within the SI's sphere of influence, i.e.
temporal-spacial range of activity.
Satisfaction landscape: assume all constituents have uniquely-defined
self-interests, and these self-interests may conflict with with the
self-interests of other constituents in various ways. The satisfaction
landscape may be visualized as a three dimensional volume. The proximity of
two individuals within the XY plane relates to the probability that their
self-interests may overlap, conflict, interfere, or otherwise negatively
effect each other. The distance along the Z dimension relates to how closely
achieved, observable reality relates to the idealized, desired reality for any
Optimum distibution: an optimum distribution across the satisfaction
landscape is that one which results in the highest average "height" (placement
along Z of individuals) throughout the model. Additional considerations:
given two allocations across the landscape with identical average height
values, the preferred allocation is the one for which the average slope
relative to the XY plane between points in the landscape is minimized, i.e.
it's the "smoothest."
Now, that's just a strawman. The latter definition approximates Pareto
optimality, but like many questionable social systems (e.g. socialism) it
strives for equality of outcome rather than equality of opportunity. I'm not
sure that's the best option.
I'd like to see a similar --- perhaps even more rigorous --- "official"
definition of the expectations we have for a Friendly SI. Eli, Brian?
This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:37 MDT