From: pdugan (pdugan@vt.edu)
Date: Tue Dec 13 2005 - 15:33:10 MST
I don't know how I feel about goals. On one hand, I have a sort of goal
hierarchy I pursue in my life, one that I take fairly seriously. On the other
hand, I find in my experience that the finest moments of consciousness are
when I'm not occupied with a goal, and I think apprecaiting that is a key to
having good discretion in one's actions. Its possible that this principle goes
for AGI's without explicit goal systems, and its also possible for goals to
"emerge", for a lack of a better word, from its reasoning, pattern learning
and general interactions with its environment. This isn't allowed in
Yudkowskian FAI, but maybe a goaless AGI would be a way to marginalize the
risk of hard take-off. Tremendous benifits could still be gained by prompting
the AGI interactively and learning/improving as a species based on that
interaction. As an analogy, consider the perspective that, for all intents and
purposes, an SAI is functionally a deity; would you rather have a diety with
an adgenda prompting it to radical action, or a deity willing to be as
indirect and unobvious as possible. Then the people who catch on will start
the transhuman improvement thing going, while society takes the overall
transition at a reasonable pace, maybe a decade.
Patrick
>===== Original Message From Kaj Sotala <xuenay@sci.fi> =====
>> it would be programmed to take data as an input and perform analysis on the
>> data, it also would be given the ability to self-improve, and lets assume
we
>> also succeed in making it self-aware. Would it just sit their idle because
>> no goals had been specified?
>
>Yes.
>
>Also, giving it the ability to self-improve wouldn't do anyone any good
>if it didn't have a goal to improve itself towards. (Maybe it could
>start to perform self-modification at random, if it was specifically
>designed to do so - but random changes would most likely just break
>things, and even "conduct random changes in your own code" is a goal.)
This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:54 MDT