2018-06-25 RRG Notes
- Interpersonal morality should be constructed out of personal morality
- Example:
- 5 people come across a pie
- One of them wants the whole pie, regardless of the arguments the other three put forth
- The correct thing to do is to give this person one fifth of the pie, regardless of their objections, and prevent them (by force if necessary) from taking more
- Does the concept of indvidual rights have any meaning?
- Example:
- Suppose there are two people, a "mugger" and a "muggee"
- The mugger wants the muggee's wallet
- In this situation, what use is it for the muggee to appeal to morality?
- If a third party comes along, they are more like to side with the muggee than the mugger
- Oh really? I can name numerous instances in which that is not the case
- The example that comes to mind is the pogrom in Kishniev in 1905 -- Jews ran into the houses of gentiles they considered friends, looking for sanctuary... and were turned away, while those who sheltered Jews were often attacked and beaten alongside the Jews they sheltered
- I suspect that Eliezer's response to this would be that the people committing the pogrom were in moral error. I agree with that conclusion. However, I don't see how Eliezer's framework can lead to that conclusion
- And in any case, it still doesn't answer the question. It just pushes the question back a level - why should the third party side with the mugger than the muggee?
- If a fourth party comes along, they are likely to side with the third party intervening on behalf of the muggee
- Again, this is just not true! In any kind of real scenario, a fourth party is going to have no idea who the is in the "right"
- What will happen is that the first, second and third parties will have to talk to the fourth party and attempt to prove that they are acting on the side of right
- This is why we have standards of evidence and criminal trials... and even then we often get things wrong
- When we talk about individual rights, we're talking about violations upon which is is obligatory for a third party to intervene, and for a fourth party to support the third party's intervention
- However, this does not work as a meta-ethics
- If you fell back in time to a time in which slavery was accepted, it would still be morally right for you to help an escaped slave, even if that was not considered acceptable to society
- What? Doesn't that immediately contradict what he wrote above, with regards to invidividual rights being that which a third party ought to step in to enforce?
- Individual rights only exist in relation to other people
- Example: saying that "everyone has a right to food" is meaningless, because the statement isn't directed at other people
- On the other hand, saying that people who have excess food should give up their food to those who don't have enough food is a meaningful moral statement, because it imposes obligations on people
- Interpersonal morality is a special case of individual morality, not the other way around
- Groups can disagree on the morality of actions, but individuals have to come to a decision and take an action (even if that action is to do nothing)
- "But generally speaking, neurologically intact humans will end up doing some particular thing. As opposed to flopping around on the floor as their limbs twitch in different directions under the temporary control of different personalities." -- I don't know, have you seen some of the Bay Area rationalists?
- However, because we humans have been arguing about interpersonal morality for so long, it's not surprising that we have specific arguments for this special case
- One of these adaptations is universalizability
- Desires have to frames in a form that enable them to leap from one mind to another
- I still don't understand his example. What is the difference between Dennis claming, "I want the pie," and Dennis claiming, "As high priest, God wants me to have the pie?"
- I suspect that Eliezer would claim that both were morally incorrect, but how can a simple change in phrasing make a morally incorrect claim into a morally correct claim?
- Some of our moral arguments have transcended specific tribes and context and and made the jump to being truly transpersonal arguments
- Transpersonal moral arguments are moral arguments that reflect the psychological unity of humankind
- Even the most transpersonal moral argument won't work on a rock or a paperclip-maximizer, but that doesn't really matter -- rocks and paperclip maximizers aren't really things that you can have a moral conversation with
- The question of how much actual agreement would exist among humans on matters of morality is difficult to say in today's world
- Moral disagreements might be dissolved by looking at the world in a different way or by considering different arguments
- Knowing more might dispel illusions of moral agreement
- My Thoughts
- So as far as I can tell, Eliezer is claming:
- Moral arguments only make sense between people
- There is such a thing as moral rightness and wrongness with regards to people that is independent from what any one person thinks
- This morality is the result of an extremely complex computation upon which we have very little introspection
- It's helpful to remember that the differences in morality between people, as vast as they seem, are mere hair-splitting differences when compared to the differences in morality between humans and rocks, or humans and runaway AI
- Suppose you build an AI that tries to "do what you want"
- One of the things this AI can do is modify you to strongly want something that the AI can easily provide
- If you try to make it so that the AI can't modify the programmer, then the AI can't talk to the programmer, because to communicate is to modify
- I disagree with this. I think there are degrees of modification, and that being allowed to show words on a screen to a person is different than being allowed to conduct brain surgery on the person
- Unlike Eliezer, I don't think words are magic; if someone believes something with a high degree of conviction, it's quite possible that no sequence of words will convince them otherwise
- We don't imagine any future in which we want something and we have it is a good future
- However, we don't explicitly say this, which we must in order to build a "safe" AI
- There is a duality between this problem in AI and moral philosophy, which is to say that merely "wanting" something doesn't make the thing "right"
- We don't have introspective access to our own morality, we can only look at situations, plug them into the black box of our psychology, and then get an answer, moral or not-moral
- We don't ask ourselves, "What will I decide to do?" in the abstract sense
- We ask ourselves what we will do in order to maximize our goals, which we don't have full introspective access to
- What we name "right" is a "fixed framework" that grows out of a particular starting point that we all share by virtue of being human
- My Thoughts
- Actually, this essay makes the previous one make a lot more sense
- Morality, according to Eliezer, is a computation embedded in the human psyche; an enormously complex computation, but a computation all the same
- People can have erroneous outputs to this computation, just as a malfunctioning calculator can sometimes output 2+3 = 6 if it's miswired
- Or better analogy, it can appear that 2047 is a prime number, but it factorizes to 23 x 89
- However, all people are fairly similar, by virtue of sharing a particular mind design, and, as a result, we can come up with a "universal" morality that applies to all people
- I think this is basically Eliezer's argument for CEV
- The one-sentence summation of Eliezer's theory of metaethics is: there is no pure ghostly essence of goodness apart from things like truth, happiness and sentient life
- Whenever people think about "goodness", it's in relation to specific things, such as truth, or beauty or human life, or any one of a large number of other things
- However, it doesn't seem like this; because of the way our brains work, it feels like a property like "goodness" has an independent existence of its own
- You can't replace "goodness" with "utility function" since you can construct a mind that embodies any utility function
- The moment you ask "What utility function should I use," you're back to thinking about what you value, and which utility function would best preserve those values
- Your values can change in response to arguments, but there is no form of justification that can work independently of human minds
- So is morality merely a quirk of human psychology? Should it be abandoned?
- Of course not -- we value each other because we're humans, and if the only justification we have for this valuing is that we're human, then so be it
- Worrying about the lack of universalizability of moral arguments just causes existential angst with nothing to show for it
- An allegorical story about aliens whose ethical system revolves entirely around whether heaps of pebbles have prime numbers of pebbles
- Like humans and human morality, these aliens don't really have any justification for why prime numbers of pebbles are "right" and composite numbers are "wrong"
- They fight wars over whether certain large numbers are prime or composite
- Have moral philosophers who deny that any pebble counting progress has occurred - that all developments in determining whether a large number of pebbles is a prime-number of pebbles is a "random walk"
- When they build an AI, they naively think that the AI will automatically learn that prime numbers of pebbles are correct and that composite numbers are incorrect
- The purpose of this story is to illustrate the orthogonality thesis which states that any intelligence can be combined with any goal
- More intelligence doesn't automatically imply more morality
- So if all morality in humans has the same starting point, then why are there so many moral disagreements between humans?
- People do not have complete access to the output of their own morality
- Disagreement has two prerequisites
- Possibility of agreement
- Possibility of error
- I'm not so sure about that. Yes, according to Aumann's Agreement Theorem, it's not possible for ideal Bayesian reasoners to disagree
- However, Aumann's Agreement Theorem only applies to matters of fact, to matters of value
- Eliezer is claiming that matters of value are matters of fact - that morality is a fixed computation, and that moral disagreement means that one side is wrong about the results of that computation, not that they're attempting a different computation entirely
- This is where I disagree: I think it is possible that one side is running a different computation entirely
- There are numerous ways in which a person could be mistaken about their own morality
- They might have mistaken beliefs about the world
- They might have a mistaken meta-ethics
- They might have cultural influences which cause them to make a moral error
- What I don't understand is how this enables Eliezer to conclude that one side is correct whereas the other side is not. He keeps talking about moral progress. But what about moral regress? Maybe every day, we do stray further from the light of God
- Assuming that fellow humans share an entirely different reference frame, and thus, an entirely different morality, is an extreme position to take
- Even psychopaths share largely the same moral frame as non-psychopaths and it's plausible that a psychopath would take a pill that would turn them into a non-psychopath if it were offered
- This is empirically false; people refuse psychotropic drugs all the time, if they think that said drugs will change their values
- This is why I have not and will never take LSD
- Moreover, this makes about as much sense as a paperclip maximizer accepting a code change to turn it into a staple maximizer
- Eliezer claims that saying, "There is nothing to argue about, we are merely different optimization processes," is something that we should use only on paperclip maximizers and not fellow humans
- I 100% disagree with that, because another way of phrasing that statement is, "I guess we have different values, let's agree to disagree and move on." That's something that people do all the time
- But of course, as we all know, Eliezer can't agree to disagree
