Okay. “Morality”’s banned too, as I use it as a synonym for ethics.
As a sub-component of my total preferences, which are predictors of my actions, I consider a kind of “averaged preferences” where I get no more stake in deciding what constitutes a better world than any other mind. The result of this calculation then feeds into my personal preferences, such that I have a weak but not inconsiderable desire to maximise this second measure, which I weigh against other things I want.
It seems to me that you don’t do this second loop through. You have your own desires, which are empathically sensitive to some more than others, and you maximise those.
As a sub-component of my total preferences, which are predictors of my actions, I consider a kind of “averaged preferences” where I get no more stake in deciding what constitutes a better world than any other mind. The result of this calculation then feeds into my personal preferences, such that I have a weak but not inconsiderable desire to maximise this second measure, which I weigh against other things I want.
It seems to me that you don’t do this second loop through.
Oh I do that too. The difference is that I apply a appropriately reduced selfish factor for how much I weigh minds that are similar or dissimilar from my own in various ways.
You can implement the same thing in your total preferences algorithm by using an extended definition of “me” for finding the value of “my personal preferences”.
Edit: I’m not quite sure why this is getting down voted. But I’ll add three clarifications:
I obviously somewhat care about minds that are completely alien to my own too
When I said not that different I meant it, I didn’t mean identical, I just mean the output may not be that different. It really depends on which definition of self one is using running his algorithm, it also depends what our “selfish constant” is (it is unlikely we have a the same one).
By “extended LWish dentition of “me”, I meant the attitude where if you make a perfect copy of you, they are both obviously you, and while they do diverge, and neither can meaningfully call itself the “original”.
To me, that second loop through only has value to the extent that I can buy into the idea that it’s non-partisan—that it’s “objective” in that weaker sense of not being me-specific.
This is why I was confused. I assumed that the problem was, when you talked about “making the world a better place”, “better” was synonymous with your own preferences (the ones which are predictors of your actions). In other words, you’re making the kind of world you want. In this sense, “making the world a better place” might mean you being global dictator in a palace of gold, well stocked harems, etc.
To me, putting that similarity factor into your better world definition is just a lesser version of this same problem. Your definition of “better world” is coloured by the fact that you’re doing the defining. You/ve given yourself most of the stake in the definition, by saying minds count to the extent that they are similar to your own.
To me, that second loop through only has value to the extent that I can buy into the idea that it’s non-partisan—that it’s “objective” in that weaker sense of not being me-specific.
The components that don’t have any additional weight of you are still there in my implementation. If you feel like calling something objective you may as well call that part of the function that. When I said somewhat under 1, that was in the context of the modifier I give to people/entities that are “part-me” when applying the selfish multiplier.
Konkvistador:
1 Me selfish multiplier + 0.5 Me 0.5 * selfish multiplier + …. + 0 Me + 0 Me + 0 Me
syllogism
1 Me * selfish multiplier + 0.5 Me + 0.3 Me + …. 0 Me + 0 Me + 0 Me
As you can probably see there are trivial ways to make these two equivalent.
Define P[i][j] as the preference-weight for some outcome j of some mind i. P[me][j] is my preference weight for j.
To decide my top-level preference for j—ie in practice whether I want to do it, I consider
S P[me][j] + E sum(P[i][j] for i in minds)
Where S is the selfishness constant, E is the ethics constant, and S+E=1 for the convenience of having my preferences normalised [0,1].
In other words, I try to estimate the result of an unweighted sum of every mind’s preferences, and call the result of that what a disinterested observer would decide I should do. I take that into account, but not absolutely. Note that this makes my preference function recursive, but I don’t see that this matters.
I don’t think your calculation is equivalent, because you don’t estimate sum(P[i][j] for i in minds). To me this means you’re not really thinking about what would be preferable to a disinterested observer, and so it feels like the playing of a different game.
PS In terms of FAI, P[i][j] is the hard part—getting some accurate anticipation of what minds actually prefer. I unapologetically wave my hands on this issue. I have this belief that a pig really, really, really doesn’t enjoy its life in a factory farm, and that I get much less out of eating bacon that it’s losing. I’m pretty confident I’m correct on that, but I have no idea how to formalise it into anything implementable.
I don’t think I’m misunderstanding since while we used different notation to describe this:
S P[me][j] + E sum(P[i][j] for i in minds)
1 Me * selfish multiplier + 0.5 Me + 0.3 Me + …. 0 Me + 0 Me + 0 Me
We both described your preferences the same way. Though I neglected to explicitly normalize mine. To demonstrate I’m going to change the notation of my formulation to match yours.
1 Me * selfish multiplier + 0.5 Me + 0.3 Me + …. 0 Me + 0 Me + 0 Me
My notation may have been misleading in this regard, 0.5 Me isn’t 0.5 Me it is just the mark I’d use for a mind that is … well 0.5 Me*. In your model the “me content” dosen’t matter when tallying minds, except when it hits 1 in your own, so there is no need to mark it, but the reason I still used the fraction-of-me notation to describe certain minds was to give an intuition of what your described algorithm and my described algorithm would do with the same data set.
Konkvistador:
1 Me selfish multiplier + 0.5 Me 0.5 * selfish multiplier + …. + 0 Me + 0 Me + 0 Me
syllogism
1 Me * selfish multiplier + 0.5 Me + 0.3 Me + …. 0 Me + 0 Me + 0 Me
So if syllogism and Konkvistador where using the same selfish multiplier (let us call it S for short as you do) the difference between their systems would be the following
0.5 Me 0.5 (S-1) + 0.3 Me 0.3 (S-1) + …. really small fraction of Me really tiny number (S-1)
This may be a lot or it may not be very much, it really depends on how big it is compared to:
1 Me * S + 0 Me + 0 Me + 0 Me + … 0 Me
In other words if “Me” is very concentrated in a universe, say you drop me in a completely alien one, my algorithm wouldn’t produce an output measurably different from your algorithm. Your algorithm can also consistently give the same result if your S and Me embrace an extended self-identify, rather than just your local feeling of self. Now this of course boils down to the S factor and me being different for the same person when using this algorithm (we are after all talking about how something is or isn’t implemented rather than having silly sloppy math for fun), but I think people really do have a different S factor when thinking of such issues.
In other words if for S * P[me][j] you use dosen’t force your P[me][j] to necessarily a value of one. To help you understand a bit more by that imagine there is a universe that you can arrange to your pleasure and it contains P[you] but not just any P[you] it contains P[you] minus the last two weeks of memory. Does he still deserve the S factor boost? Or at least part of it?
Readers may be wondering that if the two things can be made mathematically equivalent, why I prefer my implementation to his (which is probably more standard among utilitarians who don’t embrace an extended self). Why not just adopt the same model but use a different value of Me or a different S to capture your preferences? This is because in practice I think it makes the better heuristic for me:
The more similar a mind is to mine, the less harm is done by my human tendency towards anthropomorphizing (mind projection fallacy is less an issue when the slime monster really does want our women). In other words I can be more sure of my estimation of their interests, goals and desires is are likley to be influenced by subconscious rigging “their” preferences in my favour because they are now explicitly partially determined by the algorithm in my brain that presumably wants to really find the best option for an individual (the ones that runs when I say “What do I want?”). Most rationalist corrections made for 0.5 Me * 0.5 also have to be used in Me and vice versa.
I find it easier to help most people, because most people are pretty darn similar to me when comparing them with non-human or non-living processes. And it dosen’t feel like a grand act selflessness or something that changes my self-image, signals anything or burns “willpower” but more like common sense.
It captures my intuition that I don’t just care about my preferences and some averaged thing, but I care about specific people’s preferences independent of “my own personal desires” more than others. This puts me in the right frame of mind when interacting with people I care about.
Edit: Down-voted already? Ok, can someone tell me what I’m doing wrong here?
Okay. “Morality”’s banned too, as I use it as a synonym for ethics.
As a sub-component of my total preferences, which are predictors of my actions, I consider a kind of “averaged preferences” where I get no more stake in deciding what constitutes a better world than any other mind. The result of this calculation then feeds into my personal preferences, such that I have a weak but not inconsiderable desire to maximise this second measure, which I weigh against other things I want.
It seems to me that you don’t do this second loop through. You have your own desires, which are empathically sensitive to some more than others, and you maximise those.
I think our positions may not be that different.
Oh I do that too. The difference is that I apply a appropriately reduced selfish factor for how much I weigh minds that are similar or dissimilar from my own in various ways.
You can implement the same thing in your total preferences algorithm by using an extended definition of “me” for finding the value of “my personal preferences”.
Edit: I’m not quite sure why this is getting down voted. But I’ll add three clarifications:
I obviously somewhat care about minds that are completely alien to my own too
When I said not that different I meant it, I didn’t mean identical, I just mean the output may not be that different. It really depends on which definition of self one is using running his algorithm, it also depends what our “selfish constant” is (it is unlikely we have a the same one).
By “extended LWish dentition of “me”, I meant the attitude where if you make a perfect copy of you, they are both obviously you, and while they do diverge, and neither can meaningfully call itself the “original”.
To me, that second loop through only has value to the extent that I can buy into the idea that it’s non-partisan—that it’s “objective” in that weaker sense of not being me-specific.
This is why I was confused. I assumed that the problem was, when you talked about “making the world a better place”, “better” was synonymous with your own preferences (the ones which are predictors of your actions). In other words, you’re making the kind of world you want. In this sense, “making the world a better place” might mean you being global dictator in a palace of gold, well stocked harems, etc.
To me, putting that similarity factor into your better world definition is just a lesser version of this same problem. Your definition of “better world” is coloured by the fact that you’re doing the defining. You/ve given yourself most of the stake in the definition, by saying minds count to the extent that they are similar to your own.
The components that don’t have any additional weight of you are still there in my implementation. If you feel like calling something objective you may as well call that part of the function that. When I said somewhat under 1, that was in the context of the modifier I give to people/entities that are “part-me” when applying the selfish multiplier.
Konkvistador:
1 Me selfish multiplier + 0.5 Me 0.5 * selfish multiplier + …. + 0 Me + 0 Me + 0 Me
syllogism
1 Me * selfish multiplier + 0.5 Me + 0.3 Me + …. 0 Me + 0 Me + 0 Me
As you can probably see there are trivial ways to make these two equivalent.
Hmm either I don’t understand or you don’t.
Define P[i][j] as the preference-weight for some outcome j of some mind i. P[me][j] is my preference weight for j.
To decide my top-level preference for j—ie in practice whether I want to do it, I consider
S P[me][j] + E sum(P[i][j] for i in minds)
Where S is the selfishness constant, E is the ethics constant, and S+E=1 for the convenience of having my preferences normalised [0,1].
In other words, I try to estimate the result of an unweighted sum of every mind’s preferences, and call the result of that what a disinterested observer would decide I should do. I take that into account, but not absolutely. Note that this makes my preference function recursive, but I don’t see that this matters.
I don’t think your calculation is equivalent, because you don’t estimate sum(P[i][j] for i in minds). To me this means you’re not really thinking about what would be preferable to a disinterested observer, and so it feels like the playing of a different game.
PS In terms of FAI, P[i][j] is the hard part—getting some accurate anticipation of what minds actually prefer. I unapologetically wave my hands on this issue. I have this belief that a pig really, really, really doesn’t enjoy its life in a factory farm, and that I get much less out of eating bacon that it’s losing. I’m pretty confident I’m correct on that, but I have no idea how to formalise it into anything implementable.
I don’t think I’m misunderstanding since while we used different notation to describe this:
S P[me][j] + E sum(P[i][j] for i in minds)
1 Me * selfish multiplier + 0.5 Me + 0.3 Me + …. 0 Me + 0 Me + 0 Me
We both described your preferences the same way. Though I neglected to explicitly normalize mine. To demonstrate I’m going to change the notation of my formulation to match yours.
1 Me * selfish multiplier + 0.5 Me + 0.3 Me + …. 0 Me + 0 Me + 0 Me
P[me][j] * S + P[1] + P[2] + …. P[i-2] + P[i-1] + P[i]
S P[me][j] + E sum(P[i][j] for i in minds)
My notation may have been misleading in this regard, 0.5 Me isn’t 0.5 Me it is just the mark I’d use for a mind that is … well 0.5 Me*. In your model the “me content” dosen’t matter when tallying minds, except when it hits 1 in your own, so there is no need to mark it, but the reason I still used the fraction-of-me notation to describe certain minds was to give an intuition of what your described algorithm and my described algorithm would do with the same data set.
So if syllogism and Konkvistador where using the same selfish multiplier (let us call it S for short as you do) the difference between their systems would be the following
0.5 Me 0.5 (S-1) + 0.3 Me 0.3 (S-1) + …. really small fraction of Me really tiny number (S-1)
This may be a lot or it may not be very much, it really depends on how big it is compared to:
1 Me * S + 0 Me + 0 Me + 0 Me + … 0 Me
In other words if “Me” is very concentrated in a universe, say you drop me in a completely alien one, my algorithm wouldn’t produce an output measurably different from your algorithm. Your algorithm can also consistently give the same result if your S and Me embrace an extended self-identify, rather than just your local feeling of self. Now this of course boils down to the S factor and me being different for the same person when using this algorithm (we are after all talking about how something is or isn’t implemented rather than having silly sloppy math for fun), but I think people really do have a different S factor when thinking of such issues.
In other words if for S * P[me][j] you use dosen’t force your P[me][j] to necessarily a value of one. To help you understand a bit more by that imagine there is a universe that you can arrange to your pleasure and it contains P[you] but not just any P[you] it contains P[you] minus the last two weeks of memory. Does he still deserve the S factor boost? Or at least part of it?
Readers may be wondering that if the two things can be made mathematically equivalent, why I prefer my implementation to his (which is probably more standard among utilitarians who don’t embrace an extended self). Why not just adopt the same model but use a different value of Me or a different S to capture your preferences? This is because in practice I think it makes the better heuristic for me:
The more similar a mind is to mine, the less harm is done by my human tendency towards anthropomorphizing (mind projection fallacy is less an issue when the slime monster really does want our women). In other words I can be more sure of my estimation of their interests, goals and desires is are likley to be influenced by subconscious rigging “their” preferences in my favour because they are now explicitly partially determined by the algorithm in my brain that presumably wants to really find the best option for an individual (the ones that runs when I say “What do I want?”). Most rationalist corrections made for 0.5 Me * 0.5 also have to be used in Me and vice versa.
I find it easier to help most people, because most people are pretty darn similar to me when comparing them with non-human or non-living processes. And it dosen’t feel like a grand act selflessness or something that changes my self-image, signals anything or burns “willpower” but more like common sense.
It captures my intuition that I don’t just care about my preferences and some averaged thing, but I care about specific people’s preferences independent of “my own personal desires” more than others. This puts me in the right frame of mind when interacting with people I care about.
Edit: Down-voted already? Ok, can someone tell me what I’m doing wrong here?
Can you see how mathematically the two algorithms could create the same output?