Re: AI-Box Experiments

From: Eliezer S. Yudkowsky (sentience@pobox.com)
Date: Mon Jul 08 2002 - 19:11:23 MDT


Anders Sandberg wrote:
> I think it could be run as a roleplaying scenario to 1) enhance the
> motivation to keep the AI in by some setting/scenario, 2) have a
> public and/or a gatekeeper committee. Doing it over the net has the
> advantage of easy documentation and easy access, but doing it as a
> group face-to-face produces faster and often more complex
> interaction. Maybe one should run it with the AI interacting through
> a computer link, and the rest (committee and public) together.
>
> Sounds fun. I might try it with my gaming group. Although that may be
> too easy - they accepted a proposal from a posthuman with bad
> reputation last session. Kids, *never* accept femtotech devices from
> superintelligences you don't know! :-)

I'd be interested in the results of any AI-Box roleplaying, but be sure
to specify it's roleplaying when reporting it, since the dynamics of
roleplaying are different from the dynamics of the experiment. In
roleplaying you want both sides to have a fair chance of winning. In
the experiment, if the AI's side had turned out to have no chance of
winning, that would have been a fair answer to the question. Also, I
suspect that in roleplaying you would want to permit the AI to make its
way out by planting a Trojan Horse as you discuss above.

-- 
Eliezer S. Yudkowsky                          http://singinst.org/
Research Fellow, Singularity Institute for Artificial Intelligence


This archive was generated by hypermail 2.1.5 : Sat Nov 02 2002 - 09:15:15 MST