From: Gordon Worley (redbird@rbisland.cx)
Date: Sun Feb 09 2003 - 16:13:03 MST
On Sunday, February 9, 2003, at 05:14 PM, Michael Roy Ames wrote:
>> Worse yet, what if it learns how to be deceptive while playing games.
>>
>
> Yes, what if? This issue shares similar difficulties to that of
> competition, and I would suggest similar solutions. Deception is a
> useful technique is certain situations, as long as it is used as a
> subgoal in service of a larger positive-sum goal - such as the poker
> example. As with competitiveness, deception should be explained and
> taught only when the Seed AI can understand its utility in the larger
> positive-sum context.
Deception is lying, one of the last things we want a seed AI to start
doing. If it's capable of lying, we're in trouble and need to shut the
AI down, if that's still possible. Any games a Friendly seed AI plays
must be non competitive and, more so, have no room for immoral
techniques. The only game an AI can play is the be-moral game (and its
subgames, like the make-everybody-happy game), and even then we must be
careful not to expose it to deception being employed by other players
and pieces. If you think about it more you'll see the only kind of
game a Friendly seed AI can safely play isn't much of a game in human
terms, but is still recognizable as a game.
I think an adult FAI will be able to fully handle being exposed to
people who lie without fear that it will start lying itself. Until
such time, demonstrating the effectiveness of lying could be deadly.
-- Gordon Worley "It requires a very unusual mind http://www.rbisland.cx/ to undertake the analysis of redbird@rbisland.cx the obvious." PGP: 0xBBD3B003 --Alfred North Whitehead
This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:41 MDT