RE: Defining Right and Wrong

From: Billy Brown (
Date: Mon Nov 25 2002 - 10:33:43 MST

Samantha Atkins wrote:
> Hmm. Better check this out with Computability Theory and existence in a
> finite universe. Also consider some of the logical conundrums implied
> by "perfect knowledge". There will never be an entity with "perfect
> knowledge" and "infinite processing power". So why base an argument on
> such even as a useful model. Models containing logial contradictions
> are not generally useful except to prove the model is faulty.

For exactly the same reason that thinkers in the physical sciences usually
start with idealized models that only roughly correspond to the real world.
The first step to constructing a fundamental theory in a given field is to
strip away all the messy details of the real world and try to isolate the
essential problem you are trying to grapple with. Once you've got a handle
on how to deal with the simplified, idealized version of the problem, then
you can go back and try to solve it again with the realistic complications.

In this case, it is interesting to note that many of your objections to the
idealized model also point to failings of real human ethical systems. After
all, how many existing systems explicitly acknowledge and try to deal with
issues like limited information, uncertainty about the effects of actions
and finite planning time? It also becomes obvious, in the face of the
unrealistic complexity of fully solving even the ideal case, that any system
useable by humans has to be more like a set of handy heuristics than a list
of inviolate universal principles.

>From a Friendly AI perspective, especially, I think this is a more
productive way of approaching the problem of ethics than the traditional
human "let's lay down some universal rules that sound good" method. Arguing
about what list of rules is best doesn't work out well, whether the
programmers are doing it or the AI is trying to figure it out internally,
because there is no clear way to compare them. OTOH, trying to build a real
system that approximates the performance of the ideal one as well as
possible is a clear-cut engineering problem that allows different approaches
to be tested against each other, and their pros and cons weighed in a
meaningful fashion.


This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:41 MDT