RE: Humane-ness (resend due to addressing error)

From: Chris Healey (
Date: Thu Feb 19 2004 - 12:07:47 MST

> On February 18th, Ben G Wrote:
> But what you describe -- creating an AI that will embody a set of
> consistent, meaningfully
> defined goals PRIOR to massive iterated self-modification -- that is
not as hard...

Well, assuming we make a marked point of "intending" humaneness, that
sounds like a good start :)

I didn't get the impression from CFAI that Eliezer thought the
massively iterated pure self-modification should happen before the
concepts and goal system reflected a high degree of convergence. His
position seemed to represent that if we get the structural
consideration correct, then given accurate information, the AI should
voluntarily refrain from such self-enhancement until it's "predicted"
self-enhancements are regularly MUCH better than the best programmer
enhancements, and ALL of the AGI's "practice" design attempts are
analytically approved by the programmers on an ongoing basis. Plus a
margin of safety. If we get to this point at all, it seems a
statistical near-certainty that that convergence has happened.

We probably STILL can't conclude whether or not that convergence will
be maintained into transhumanity , but it seems like the most solid
methodology I've heard proposed. Although, most other specific
recommendations I've seen aren't really an integrated set of ideas,
but rather a disparate set of recommendations.

I think that as Eliezer further develops his ideas, and restates them
in more standardized terminology (i.e. game theory and experimental
economics, etc..), then the precepts of Friendly AI should be more
accessible to a larger body of specialists.

>From my academic experience, most of the minds that COULD be
contributing to the theory are pretty much doing all they can, on
information overload, to filter out the crap from their respective
field. That more or less results in ignoring content not phrased in
readily processible terminology. So as I see it, one of the biggest
challenges to the disemination of Friendliness Theory is a matter of
avoiding an easy misclassification by those individuals' "efficiency

This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:45 MDT