From: Michael Roy Ames (michaelroyames@hotmail.com)
Date: Sat Nov 30 2002 - 21:20:13 MST
Dear Billy,
You wrote:
>
> At this point you've turned the majority of ethics into a
combination of
> software engineering and experimental science, which to my mind is
a
> necessary step for any serious Friendly AI project.
>
I agree. What you have described: facts about existing human ethics,
ways of judging ethical dilemmas, defined measures (and their
limitations) and how to use those measures - is part of Friendliness
content. We can start creating and organizing this sort of content
immediately, because it is not contingent upon the detailed design of
FAI. Once a Seed AI is up and running, and is in the learning phase,
this content can be introduced/taught/shown to the program in easily
digestible lessons. I believe that a lot of groundwork can be laid
in preparing this content, and that such content will be useful for
*any* AI project that wishes to contain and manipulate ethical/moral
data. Such content will obviously be just heuristics, 'rules of
thumb', and will be merely a starting point from which a sentience
can learn through experience of simulated moral dilemmas, and
interaction with other moral sentients.
Creating the FAI content is going to be interesting and important,
and can proceed independently of the programming of the Seed AI. I
imagine it to be like putting together the schooling curriculum for a
child, a child who will probably absorb and integrate the information
very quickly. The simulations and tests presented to a Seed AI will
be as important (if not more so) than the initial content, as *that*
is when the Seed AI will relate the new information to all the other
knowledge it has. The lessons will have to be graded (just like for
human children) starting with very simplistic material, probably
tailored to a 'toy-simulation' or 'microdomain', and gradually
becoming more sophisticated as the Seed AI grows in intelligence.
I'm not sure if SL4 is the place to discuss the Seed AI's curriculum
(Eliezer?) but I have lots of ideas about it (I'm sure many
subscribers here do as well). I'll save them for another time.
Michael Roy Ames
This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:41 MDT