Re: Please Re-read CAFAI

From: Jef Allbright (
Date: Wed Dec 14 2005 - 00:37:17 MST

Jef (to Michael)
Do you mean that to the extent an agent is rational, it will naturally
use all of its
instrumental knowledge to promote its own goals and from its point of
view there would be no question that such action is good?

Well, the CI is by definition an objective thing. I do *not* believe
that it's obvious that all AIs would be relativists about morality.
However, I believe it's most likely so we don't need to argue about

I certainly don't see any reason to argue about that particular topic
at this time.

The intended question has a double-entendre. The first meaning is that
"all my goals are good from my point of view, by definition, so I
don't need to question it". The second meaning is "I have no
conception of morality, therefore I don't need to question it".

My intent was closer to your first interpretation, but perhaps not
exact. When I say "good" I don't mean it in a moral sense or as in
"good" vs. "evil". I am trying (and not doing a very good [oops, that
word again] job of it) to show that any agent (human, AGI, or other
form) must necessarily evaluate actions with respect to achieving its
goals, and that actions which promote its goals must be considered
good to some extent while actions that detract from its goals must be
considered bad. I am arguing that what is called "good" is what
works, and while such evaluations are necessarily from a subjective
viewpoint, that we can all agree (objectively) that for each of us,
what works is considered good.


I think it's concievable that an AGI might have no moral sense, and be
bound only by consequential reasoning about its goals. I also think
it's concievable than an AGI would have a moral sense, but due to its
mental model have varying beliefs about good and evil, despite its
conclusions about what is objectively true.

I would argue, later, that any moral "sense" is ultimately due to
consequential effects at some level, but I don't want to jump ahead
yet. I would also argue, later, that any moral sense based on
"varying beliefs about good and evil, despite its conclusion about
what is objectively true" (as you put it) would be an accurate
description of part of current human morality, but limited in its
capability to promote good due to its restrictions on its own
instrumental knowledge that must be employed in the promotion of its

Jef's question #2 to Michael
If this [question #1] is true, then would it also see increasing its
objective knowledge in support of its goals as rational and inherently
good (from its point of view?)

Not necessarily. It may consider knowledge to be inherently morally
neutral, although in consequential terms accumulated knowledge may be
morally valuable. An AGI acting under CI would desire to accumulate
objective knowledge as it related to its goals, but not necessarily
see it as good in itself.

I wasn't making any claim about something being good in itself. I was
careful each time to frame "good" as the subjective evaluation by an
agent with regard to whether some action promoted its goals.

There are a lot of abstractions here. Subjectivity doesn't destroy my
point. Even from its own point of view, it might be that some
knowledge is good, and some is not good, and some is neutral. An AGI
might regard knowledge as good, if it contributes to its goals, for
example. If that were the litmus test, then subjectively speaking,
some knowledge would be good, and some would be morally neutral.

Perhaps you missed what I see as the obvious logical opposite -- that
the AGI adopts, subjectively speaking, the "belief"
(goal/desire/whatever) than Knowledge Is Good. In this case, the AGI
desires knowledge *as an end in itself* and not *solely for its
contribution to other goals*.

Jef's question #3 to Michael
If I'm still understanding the implications of what you said, would
this also mean that cooperation with other like-minded agents, to the
extent that this increased the promotion of its own goals, would be
rational and good (from its point of view?)

Obviously, in the simple case.

Interesting that you seemed to disagree with the previous assertions,
but seemed to agree with this one that I thought was posed within the
same framework. It seems as if you were not reading carefully, and
responding to what you assumed might have been said.

Well, you just said something that was true this time :). I'm not
going to change my mind because you said it in the context of a wider
flawed argument ;).

Let's suppose than an AGI assesses some possible course of action --
in this case one of interaction and co-operation. It works out that
pursuing it will contribute to the furtherance of its own goals. It
certainly isn't going to think that such a thing is either evil or
bad. It may have no sense of morality, but in the sense of
advantageous, such an action will be good.

Let me then clarify: yes, such a thing will always be advantageous.
Insofar as an AGI has a moral system, such a thing will also be
morally good.

Yes, I consistently mean "good" in the sense of advantageous. It
seems that making this clear from the beginning is key to more
effective communication on this topic. My difficulty with this is I
see "good" as *always" meaning advantageous, but with varying context.

I think we are in agreement that for any agent, those actions which
promote its values will necessarily be seen by that agent as good.

Thanks for taking the time to work through this, and thanks to the
list for tolerating what (initially, at least) appeared to be very low
signal to noise.

- Jef

This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:54 MDT