Re: Cats, was I am a moral, intelligent being

From: Martin Striz (mstriz@gmail.com)
Date: Wed Jun 07 2006 - 14:48:21 MDT


On 6/7/06, Keith Henson <hkhenson@rogers.com> wrote:

> My point was that more intelligent AIs or upgraded humans may have a
> different view of what is moral as we have a different view of what is
> moral compared to cats. If you upgraded a cat to human level intelligence
> would it think controlling the population of regular cats the way we do was
> moral? (I have no idea.)

This is the problem that I have with "collective volition." There is
not such thing. There are millions of independent trajectories that
would lead to happiness. Cats love nothing better than to hunt. So
if I were to upgrade a cat, should I give it the intelligence to build
mouse traps while retaining the desire to kill things, or should I
engineer it to love mice and other woodland creatures? From an
engineering standpoint, all forms of happiness are equivalent. It is
not better to be Socrates than to be the pig. If you don't think so,
you're projecting your own anthropomorphosism.

Coding a "friendly" AI implies some objective measure of "happiness."
But I don't think such a thing exists in objective, engineering terms.
 Happiness can be any feedforward process. There are millinos of
trajectories by which an AI could upgrade us to be what we would want
to be, or it could engineer us to /want to be/ what we are. When you
understand that there's no difference, then you understand that
happiness is an infinite goal set, which means that it's an empty goal
set.

So an AI engineered to "make people happy" will do anything it pleases
once it recursively self-improves. Now you COULD engineer an AI to
make people happy by the metrics with which we measure happiness now,
as humans 1.0. In other words, you hardcode as goals: make people
healthier, live longer, smarter, etc. You have to narrow the domain
for the AI.

Ben's suggestion of creating limited AGI, and within that framework,
limited FAI, is a good one.

Martin



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