From: James Ratcliff
To: [email protected]
Sent: Monday, June 12, 2006 3:52 PM
Subject: Re: [agi] Re: Four axioms (WAS Two draft papers . . . .)

>> you mentioned in a couple of responses the volition of the masses as your overall formula, I am putting a couple of thoughts together here, and that the intitial formula and rules for this would be hard-coded in the beginning. 
 
    This initial "formula" would be the four axioms and the rules would be the rules for extrapolation which I'll be posting once I think I've got initial agreement on the axioms (and yes, there is a substantial amount of voodoo in the rules for extrapolation that is going to need to be gone over VERY carefully and the rules will probably need to be eventually codified as axioms as well).
 
>> Using this as a starting point then, How do two AGI's develope along a different line of beliefs?
 
    If experience teaches the AGI's different things about the world, then the AGI's will develop different beliefs.  An AI whose experience base only includes a tribe where ritualistic cannibalism is performed after death to make survivors feel happier because the dead will forever remain a part of them (a la Robert Heinlein's Stranger In a Strange Land) and who never sees any of the diseases (primarily parasitic) caused by cannibilism will conclude that cannibalism is friendly and good .  An AI whose experience base mainly includes people who are strongly revulsed by any cannibalism and who has seen the diseases caused by cannibalism will conclude that cannibalism is unfriendly and bad.  Both AIs will be "correct" within their respective worlds and will have to work out what is friendly and good when their worlds collide.
 
>> My other main problem with your theory of volition and my own Value formula is the future states.  How can we possibly program them to look so far forward and attempt to glean everyones future volition's.  That seems like an insurmountable task in itself.
 
    My theory of volition has no problem with future states because it ignores them as unnecessary and insufficiently calculable.  I agree fully with your assessment that looking forward and attempting to glean everyone's future volition is an insurmountable task.  It is my current volition that causes me to take actions like flossing my teeth because I know that it will benefit me in the future.  Trying to extrapolate my future volition would require guessing what I will experience/learn and how I will react to it (and further, it is my strong contention that any such extrapolation is very likely to be very biased).  It is my firm belief that Eliezer has unnecessarily headed down a dead-end rabbithole with his CEV.  I don't think that guessing future volition is necessary for friendliness.
 
>> Another take on that is a large set of value formulas, that maybe we Could take a snapshot of, or a glimpse of, with trying to look directly into everyones mind and seeing thier volitions.
 
    The AI is going to have to have a good idea of what people's volitions are.  The best way to get this is to ask them (Most of the time, the simplest solutions are the best as long as they are sufficient.) 
 
>> I think we may have to simulate this with a formula that takes into account majority opinion, and protection for the minorities.
 
    Exactly.  The rules for extrapolation do exactly this. 

>> Thought:  I want a piece of candy, but it is bad for my teeth, my weight, etc,  does the robot give me a piece of candy when I ask?  At what point though does he say no vs yes.  How much of free will do the AI's attempt to put forward on us to 'protect' us, and our and the masses future volitions.
 
    The robot's decision as to whether or not to give you the piece of candy is probably going to be based upon a) how well it believes that you are capable of determining what is best for yourself, b) whether it believes that it actually is good or bad for you, and c) how much happiness giving you candy in this one instance is likely to cause.  If you are a mature adult, the AI will generally expect that you know that large amounts of candy are bad for you and that you will therefore self-regulate so it will probably give you candy when you ask if it is easy for it to do so (i.e. there are not substantial other trade-offs involved).  Note, however, that the robot is not in any way at all compelled to give you candy unless it wants to (because it thinks it will cause the most happiness for itself, you, and everyone else) because <teaser>one of the rules of extrapolation is that no entity is required to take action to fulfill the desires of another entity</teaser>.  If you are a child and it has seen you eat far too much candy, it will likely decline to give you any.  If you are a child and it knows that you haven't had a lot of candy and that your parents wouldn't mind, it may well give it to you.  Note, however, that a robot declining to give you candy is VERY different from a robot attempting to block your acquisition of candy.  The latter is a clear case of a conflict between two volitions (your desire for candy and it's desire to prevent you from having candy) that are explicitly declared as equal by axiom 2. 
 
    I am EXTREMELY leery of any entity which trys to "protect me" because they "know what is better for me than I do".  That is the path to forcible "uplifting" (guaranteed to drive many people absolutely berserk and to trigger DeGaris's "artilect war").  It is also the path of Leon Kass telling me how I can or can not modify myself.

>> You mentioned once in one of the posts, that your volitions would be represented by rules, I dont remember exactly your wording, but it seems more likely that we would have to define variations of these rules, and allow the AGI to learn more rules, as oppsed to giving them a very devious task of gleaning ones volition.
 
    No, my volitions are assumed to be exactly what I say that they are or, in dire cases where I can't be consulted, what is normal and common as long as there is no evidence to the contrary.  Other people's volitions are what they say they are or, in dire cases where they can't be consulted, what is normal and common as long as there is no evidence to the contrary. 
 
    Whether or not these volitions are friendly is extrapolated/derived by rules from the initial four axioms.  I believe that the necessary set of rules for proper expansion is pretty small (yet will produce infinite extrapolation based upon experience -- as in the case of the two AIs considering the morality of cannibalism).
 
        Mark

To unsubscribe, change your address, or temporarily deactivate your subscription, please go to http://v2.listbox.com/member/[EMAIL PROTECTED]

Reply via email to