Re: The chinese finger trap

From: Rolf Nelson (
Date: Sat Apr 26 2008 - 19:40:30 MDT

Keep in mind, the only possible hypothetical simulator that could prod a
UFAI into acting somewhat friendly would (by definition) be a Friendly AI.
Any action we take which reduces the probability of an FAI would also reduce
the probability that a UFAI would be motivated by the possible existence of
an FAI.

I think we're better off building an FAI, but it's a valid technical
question that should ideally be discussed more. If you (or anyone else) want
to create a model that shows in what universes and under what assumptions
building an FAI is a good or bad idea given AI Deterrence concerns, feel
free to submit it to my AI Beliefs blog, or better yet write a peer-reviewed
paper about it.


On Sat, Apr 26, 2008 at 4:02 PM, Rebecca <> wrote:

> Remember rolf's idea from a while back, the cooperation with hypothetical
> simulators via newcomb's paradox such and such theory?
> It recently occurs to me that if this thinking is true, it may actually be
> better NOT to attempt friendly AI.
> As was noted earlier, if friendliness is something that will crop up
> spontaneously through cooperation with a hypothetical average of possible
> simulators, some of which are friendly via being correctly programmed, then
> it naturally follows FAI isn't necessary. What went unnoticed is that it may
> actually be more dangerous to attempt it than not to.
> If we make an unfriendly AI that wants to make pancakes, and doesn't care
> about humanity, then it will (theoretically) cooperate with possible
> simulators by being friendly anyway. It's no skin off its nose if humanity's
> problems get solved and they live in blissful ascension and get free
> t-shirts, as long as they don't get in the way of pancake production much.
> But if we try to make a friendly AI, and we get it a bit wrong, such that,
> say, it's built to eliminate suffering but its definition of suffering is
> too extensive and includes virtually all forms of thought, then it might
> very well NOT cooperate, because letting humanity survive is in direct
> violation of its primary goal.
> So, it seems to me we have a chinese finger trap situation, where it's
> safer to make an indifferent AI that will have nothing against cooperating
> with possible higher friendly AIs than it is to try to make an actual
> friendly AI and fail.
> As things currently stand, I think this is something that should be taken
> very seriously.

This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:01:02 MDT