RE: AI-Box Experiments

From: Colin Hales (colin@versalog.com.au)
Date: Sun Jul 07 2002 - 17:43:35 MDT


Eliezer S. Yudkowsky wrote:
<snip>
>
> Based on my experiences in the above cases, I've created a
> page with a suggested protocol for AI-Box experiments:
> http://sysopmind.com/essays/aibox.html
>

So, somehow we built the genie in a bottle, that the bottle has no
connectivity with the universe, and that I am sitting outside, talking
through the bars, as it were, of the cell, dangling the keys provocatively.

I am having trouble imagining how the poor >H AI serial number#0000001 ended
up boxed-in in the first place. There is a presumption that the AI can be
'programmed' into it's prison isolation sensory deprivation tank. It's a
chicken/egg thing that makes the scenario moot. If the scenario is realistic
then humans should be able to learn to play the piano by correspondence and
without owning any sort of keyboard or having ears or fingers, without ever
hearing a piano. If it _can_ be programmed into the bottle then it's not
AI#0000001, it's a copy of a previously created AI the genie has been cloned
out of its sensory envelope (anaesthetised) into the bottle from outside and
hence is already free.

So I guess it comes down to a matter of whether you believe

a) >H intelligence level - an understanding of causality of the universe in
all its subtleties, including the intricate meta-causality of human
behaviour- can be achieved through abstract reasoning alone from axiomatic
input. (Piano learned and understood by a sensorially mute AI by
correspondence).

or b) That the road to the requisite understanding for >H intellect is by
'being' in the universe you are meant to understand. (Piano learned and
undertood by a sensorially equipped AI by playing the piano).

I'm a fan of b).

An AI capable of understanding the beliefs and experience of its jailer to
the point of being able to argue its way out would never be in the bottle in
the first place. Or have I missed something?

Colin Hales
*give me a good reason to take your $20 :-) *



This archive was generated by hypermail 2.1.5 : Sat Nov 02 2002 - 09:15:13 MST