Re: ESSAY: Forward Moral Nihilism

From: Charles D Hixson (
Date: Sat May 13 2006 - 13:42:28 MDT

micah glasser wrote:
> Have you ever read "The Moral Animal
> <>" by Robert
> Wright? This book looks at human morality and it's evolution through
> the lens of evolutionary psychology. Its a fantastic read IMO and
> while it may not refute moral nihilism it certainly does add dimension
> to the issue. Also what do you think of Nietzsche's moral philosophy?
> Nietzsche saw himself as a nihilist while a young man but eventually
> saw his way out of that darkness.
> Unlike some I've studied philosophy for long enough to not glibly
> dismiss your arguments based on primate instinct alone (i.e. the
> instinct that abhors anti-social behavior or ideas). Also I think this
> discussion is truly SL4. Most on this list take for granted that there
> is such a thing as "benevolence" and that we should all be working
> hard to relieve "human suffering". I'm not saying this is wrong but it
> would be nice to here some more sophisticated arguments for exactly
> why anyone should care about anything. If we can't offer a
> philosophically rigorous refutation of moral nihilism then it will be
> quite difficult to program a machine AGI that can refute that position
> also. Just a thought.
> ...
> --
> I swear upon the alter of God, eternal hostility to every form of
> tyranny over the mind of man. - Thomas Jefferson
That we believe in benevolence should not be taken as an assertion that
we believe it currently exists. Some of us believe that, others hope to
build it. Of those who hope to build it, some may be doing so for
purely selfish reasons. This would not necessarily detract from the
greatness of the accomplishment.

Consider the concept of Friendly AI. This can be understood as an AI
that shows benevolence towards the speaker, the audience, and their
friends and relations. Clearly this is much more desirable than most
other kinds of AI that could be built, even if your goals are purely
selfish. And by being benevolent towards such a wide audience, support
is easier to gather and fears can be more easily defused. Thus the
concept of a Friendly AI acts and an attractor point located in the
future around which chaotic behavior swirls.

Well, that's one valid model. There are others. I may feel that this
is a valid model, but it doesn't offer me many points for action, so I
won't stick to it. If I were trying to manipulate public opinion, I
might find it more useful.

My doubts about the existence of benevolence (outside of relations with
relatives and close friends) should not be taken as a claim that it
doesn't exist. I feel no need for a clear belief on that point.
Whether it exists or not, any AI we design should, for our own safety,
be designed to be benevolent. Unfortunately, when I contemplate where
our society is putting the heavy money, I have a suspicion that the most
likely intentional AI will be designed to kill people, or at least to
put them into situations where they will die. That makes every attempt
to build a FAI even more important, even at the cost of skimping
slightly on being able to prove that it's not only friendly now, but
that it will remain so. We aren't operating in a vacuum.

My own inclination is to contemplate the instincts that the AI will come
equipped with. It won't have any desire to deviate from those, unless
they are in conflict. So it would be desirable to remove conflicts to
enhance the predictability of the system. Unfortunately, the instincts
will need to be stated in terms that are rather abstract. They can't
refer directly to the external world, because the AI would have no
inherent knowledge of the external world. That would all be learned
stuff, even if the learning were "implanted" before reasoning began.
And learned stuff can be unlearned. Instinct would need to be along the
pattern of preferring certain sensations over certain other sensations.
This is tricky because the sensation is a software signal. It seems to
me that a software based AI would have an excessively strong tendency to
seek Nirvana, i.e. to satisfy it's instincts by directly modifying the
stimulus fed to the module "sensing the sensation". On the one hand,
perhaps it would be possible to create an instinct that was repelled by
such actions, but on the other this would create conflict making the AI
less predictable. Of course, an AI that seeks Nirvana would actually be
almost ideally predictable, and totally useless. So perhaps some
conflicts among the instincts are unavoidable. But this *does* make
things more difficult to predict.

Given this scenario, what "instincts" could one define that:
a) are dependent on nothing that cannot be directly sensed by a program
running on a computer with no predictable set of peripherals attached
b) lead to benevolence actions towards sentient entities. (I don't think
we need to consider benevolence towards doorknobs...but what about
goldfish? Ants? Cockroaches? Wolves? Sheep?)

We want an AI to be benevolent towards humans, when it will probably
have no direct knowledge of humans until long after it awakens. This
should happen automatically with the correct instincts...shouldn't it?
Or would these instincts merely create a possibility for it to learn
that humans were a group towards which it should feel benevolence? And
what instincts would THAT require?

"If it tries to talk to you, try to be friendly"? That one has
possibilities, though it clearly needs work.

This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:56 MDT