Maybe I'm missing something here, but it just seems intuitive to me that the reference class should be all relevant beings who I could possibly be if I knew nothing more than the fact that I existed. And then in that case, isn't it just an empirical question as to whether or not a certain being like a chimp or a lobster or a lobster-human hybrid is a candidate for that? I guess it depends on your larger views about personal identity and relative sorts of cognition between species, but I definitely feel like, if I were to somehow wake up with no memories in a sensory deprivation tank that erased all knowledge of the outside world beyond my bare self-awareness, I would think it was possible that I was a woman, or an old man, or really any human being on the planet - whereas I'm pretty confident I could still be sure I wasn't, like, a fish. Couldn't a sufficiently advanced neurobiological model give us at least some guidelines for determining who is and isn't in our reference class in that case?
But whether someone could have been you isn't a question that has a precise answer. Could you have been a woman? A neanderthal? In what sense, if you don't believe in souls, could you have been someone else?
Well, the question isn't a counterfactual one about who I "could have been" - it's a question about who my existence is compatible with me being. Would you agree that, even if you had no knowledge whatsoever about anything other than your bare first-person perspective, you could know immediately that you weren't a fish (assuming, I guess, you also know basic facts about what humans are like and what fish are like)? It seems to me that a fish's first-person perspective would be qualitatively different than mine in such a way that my mere "experience of experiencing" rules it out as a possibility. Whereas I definitely think I would have no way of knowing if I was a human woman in that same situation. Neanderthal is in the middle, and I currently have no idea - but it seems plausible to me that, at some point in the future, we could know enough about neurobiology to determine that!
I think even if the reference class for SSA is set to "anyone I could possibly be" (which is to be understood in whichever sense that SIA'ers mean the term), there's still is an asymmetry here. That is, even with this understanding, on SSA you'll overwhelmingly reject L(T) over L(T-1) (where T is the precise threshold level at which the lobsters become "someone you could possibly be") given your observed evidence for your own humanity (and given your perfect knowledge that T is the exact threshold), and on SIA you won't.
Having thought about this more, though, I think you can get rid of some of the discontinuity by taking into account our level of uncertainty about where exactly the threshold T lies. This might not get rid of all the discontinuity, though (I'm less sure). And I'm also not sure an SSA'er would really mind biting the bullet here in any case.
That's true, but that's also an extraordinarily unrealistic assumption which is probably partly serving as the intuition pump (although maybe not wholly). Realistically, if we discover an alien species, we'll probably never have substantially more confidence that having a billion neurons gets them into our reference class than that having 999,999,999 neurons does.
Well, SSA reference classes have the problem of being totally made-up bullshit, so it's hard to know how we'd find out what they are. But it implies that an ideal reasoner who could figure out her refrenece class should be that presumptuous.
Additionally, this doesn't solve the problem because there will necessarily be a hard break between things that are maybe in the reference class and things that are definitely not.
>Well, SSA reference classes have the problem of being totally made-up bullshit
Well, if you reject "person I could/couldn't be" as a valid binary distinction, it's not clear how you're going to get SIA off the ground; if by contrast you grant it, it seems like something the SSA advocate should be able to appeal to to ground their reference class.
>But it implies that an ideal reasoner who could figure out her refrenece class should be that presumptuous.
Yes, I agree, but I think someone might just say that it's fine for sufficiently implausible assumptions to lead to implausible conclusions, and this would be a case of such. In reality, there's always going to be a level of certainty in reference-class-fulfilling-ness that depends continuously on the physical facts.
>Additionally, this doesn't solve the problem because there will necessarily be a hard break between things that are maybe in the reference class and things that are definitely not.
I'm not sure how this objection works. Everything gets a certain probability of being in the reference class (= "creature you could've been," whatever that means), and this varies smoothly with other kinds of information you have about them. No hard cutoff between definitely-maybe-in-our-reference-class and definitely-not-in-our-reference-class need be drawn.
//Yes, I agree, but I think someone might just say that it's fine for sufficiently implausible assumptions to lead to implausible conclusions//
Idk, doesn't seem that weird. Also, this is structurally identical to the presumptuous philosopher argument, so if that's a problem for SIA, this is a problem for SSA.
//I'm not sure how this objection works. Everything gets a certain probability of being in the reference class (= "creature you could've been," whatever that means), and this varies smoothly with other kinds of information you have about them.//
Presumably there are some things that you should think don't have any chance of being in the reference class--e.g. rocks. So then somewhere from rocks to people, there's a hard jump to things that could be in your reference class.
One way around this is to appeal to a multiverse theory. If the landscape multiverse theory is correct, then there will be an infinite number of lobster-abominations and an infinite number of humans, so it becomes unclear that we should privilege the lobster-abominations over the humans.
Second, we need to take into account meta-credences. I think it’s coherent for a proponent of SSA to argue that they have a prior for SSA that’s about .9, and that in the lobster-abomination case, they would renounce belief in SSA. It’s not as though SSA proponents have to regard belief in SSA as sacrosanct. It’s perfectly cogent to maintain that one is pretty sure that SSA is right, but that hypothetical evidence like the lobsters (maybe coupled with defeaters for the multiverse) would change one’s mind.
You have not understood the argument. The argument is that SSA implies that an ideal reasoner should, in the case described in the article, have extreme presumptuousness about the number of neurons had by the lobsters.
1) In the multiverse case, the SSA proponent thinks the reference class of observers consists of almost infinite (e.g. 10^500) humans + almost infinite (more neurons) lobsters, so whether this sextillion class of entities ends up in the reference class (whether they have more or less neurons) makes no substantive difference, and
2) The credences for more or less neurons affects the credences in SSA itself. If we have good prior grounds (99%), as you state in the article, to think the more neuron hypothesis is true, then the ideal reasoner can rationally abandon SSA. In other words, unless their prior for SSA is absurdly high, there is no reason to think that SSA believers have to stubbornly cling to the less neuron hypothesis
The intention of this comment isn’t to defend SSA -- let alone against such gruesome human-lobster combos as might exist -- but (this also seems like one of Ape in the coat’s points) wouldn’t an SSAer simply deny that one’s reference class would include such creatures, e.g. because only humans (or, if SSA isn’t to rule out the likes of simulated people and brains in vats, entities subjectively indistinguishable from humans) are candidates for inclusion in one’s reference class? To be sure, this might strike you as unprincipled or ad hoc reference-class gerrymandering, but that’s a charge the SSAer has to contend with in any case and they could reasonably complain that it doesn’t add to the objection to bring it against them in a roundabout way.
Well, SSA doesn't say that only people with your exact experiences are in your reference class--if it did, SSA would lack any updates, because it would maximize the share of people with your experiences with your experiences.
Would humans missing 10 neurons be in our reference class? However you draw the boundary around humans, you get hypersensitivity.
It’s plausibly a vague matter what counts as a human. (If you like, read ‘human’ throughout as ‘entity subjectively indistinguishable from a human’. Taking humans as the boundary of one’s reference class is of course just a plausible-sounding suggestion and not something I think the SSAer needs to commit to.) If you think that vagueness is an epistemic phenomenon, then there might still be hypersensitivity, but happily no-one would ever be in a position analogous to that of the presumptuous philosopher in your article. This hypersensitivity may or may not still be a big cost to SSA -- I’m inclined to think that if it’s troubling here, it’s at least no more troubling than the hypersensitivity that, on the epistemicist’s view, all ordinary vague predicates display, modulo the residual problem of reference-class ad hockery. If you instead take vagueness (in this instance; I suppose someone might be an epistemicist about some vague predicates and not others) to be a non-epistemic phenomenon, then SSA doesn’t give you hypersensitivity, but it might be genuinely indeterminate whether or not someone is included in the SSAer’s reference class. Again you may or may not take this to be a big problem for SSA, but fwiw my view is that wouldn’t be as significant a problem as hypersensitivity. (I’m setting aside degree theories of vagueness here, if only because I haven’t thought through how exactly the SSAer should modify their account to accomodate them.)
I'm not sure I quite understand how SIA avoids this issue. I have certain evidence about myself - for example, that I have hands instead of claws, skin instead of an exoskeleton, etc. Call this evidence V. It seems plausible that P(human|V) ought to be ~1, where "human" there is shorthand for the statement "I am a human."
SIA states that P(human|V) = E[# of human observers with V] / E[# total observers with V]. At L(0), the lobsters will be different enough from us that they won't have V, hence they won't figure into the numerator or denominator at all. (For example, they won't have perceptual evidence that they have hands instead of claws.) Around L(max), they will have V, so they will factor in to the denominator. At some point in the middle, something has to change, right? Maybe you have less ambiguity about reference class, but you still have ambiguity about "evidence."
SIA only tells you to think there are more observers that *you currently might be*. I has nothing to say about how many lobsters there are--or even human lobster combos that you know yourself not to be.
I'm not sure I understand. It seemed like your objection was that on SSA, there's some threshold T such that we should automatically highly favor the hypothesis that the creatures are L(T-1) rather than L(T), because otherwise the probability that we ourselves would be humans would be too small. And that's supposed to be absurd, because the difference between L(T-1) creatures and L(T) creatures is so small. Did I understand correctly?
Because if so, it seems like SIA gets you something similar based on the continuum of evidence the creatures could have about themselves, which is very dissimilar to ours on the low end of L and perhaps identical on the high end. Is that not the case? Do the creatures' evidence very severely differ from ours even at L(max)? Because in *that* case, SSA will not have the consequence you say it has (that we should reject L(T) in favor of L(T-1) due to the surprise of being human).
Well, on SIA the only people who count are people who we might currently be. Now, we'll probably not be sure of how advanced one will have to be in order to be people we might currently be--and that will depend on our evidence. Let's consider three scenarios:
1) We don't know where the threshold is for beings being sufficiently similar that we might currently be them. Then you don't get hypersensitivity because we have a credence split among the lowest threshold for what we might be (maybe you do get it, but that would just be for Sorites paradox reasons--the point is if we think that there's a higher probability that we currently are at higher ends of the spectrum, then we'll get a bias in favor of bigger ends of the spectrum). But this is importantly different from the result for SSA because it's about beings that we currently might be, rather than far away aliens. SIA doesn't care about what far away aliens are like, because they're not people you might currently be.
2) We know exactly what threshold of creature we are--say we're L(1 billion). Well then SIA says that you should be confident the galaxy being populated by creatures of L(1 billion)--but that's perfectly non-objectionable.
Maybe the worry for SIA is supposed to be this: let N be the minimum number such that you might currently be L(N). SIA tells you to be very confident that the universe is populated with creatures of L(N) rather than L(N-1). But this isn't implausible at all. Here's an analogy. Suppose I don't know exactly what color my shirt is. Imagine a plain white shirt is S(0), plain black is S(10,000), and each extra number darkens it a bit. I think the lowest possible value of S my shirt might be is S(1,000). There are two possibilities: one is that the universe is populated with people with shirts of S(999) and the other S(1,000). Assume they have equal priors. Well, it's perfectly reasonable to think that S(999) is ruled out by this--it flatly doesn't explain the data.
Well, before I respond, I wanted to get clarification on my question about whether the creatures are supposed to have "our evidence" V near L(max) because they're so similar to us. My main point is that if the answer to this is "no," then it doesn't follow on SSA that P(lobsters are L(T-1) | V) is much higher than P(lobsters are L(T) | V)! And if the answer is "yes," then SIA also has an issue.
They don't have our evidence. But that doesn't matter on SSA. The people in blue shirts don't have the same evidence as the people in red shirts, but nevertheless, SSA reasons as if you're randomly drawn from among all the people both in red and bleu.
I actually think I changed my mind a bit. If V is our evidence of our being human, and V is essentially coextensive with *being* human regardless of the lobsters' L-level, then it's surprising on SSA that we get V on L(T) and not on L(T-1), assuming that's where the reference class line is drawn. So I think you might be right, unless there's something else I'm missing.
I think you've convincingly disproved the type of SSA which assumes that in this kind of probability theory problems correct reference class can be constructed this way. You've proven that there is, in fact, no such L for which lobster people are in my reference class. This doesn't disprove all SSA reasoning, but narrows the available reference class quite a lot.
This is not exactly a new insight. Doomsday inference and Save The Puppy already very strongly hint that drawing the boundaries of reference class across time produce absurd results. And as soon as we outlaw this kind of shenanigans everything adds up to normality.
Of course, the actual rule for "drawing the border of the reference class", or about the anthropic reasoning in general, should be based on the available information about the causal process that led to my existence. The disagreement between scientists and anthropic-reasoners shouldn't happen at all - the latter are supposed to use the up-to-date scientific knowledge to determine the correct reference class instead of just assuming that it's L = 700 millions out of nowhere. And if I do this here, well seems that my reference class is trivial - there is exactly one person born as a result of sexual encounter between my parents at the specific date of my birth. So I'm a random sample of one, and my existence, unsurprisingly enough, tells me absolutely nothing about the human-lobster hybrids.
Sure. Consider how we treat a coin toss as a random sample from {Tails, Heads}, P(Tails) = P(Heads) = 1/2. Why do we do that? It's not that coin tosses are physically equiprobable. If we knew all the physical forces affecting the coin at the moment of the toss we would be able to predict the outcome via deterministic model. But usually we do not know it. And so we *approximate* actual physical process this way via a mathematical model based on our limited state of knowledge about the coin toss. This model still corresponds to reality, but it's less accurate. And both changes to the actual causal process that makes the toss, and our knowledge state affect the applicability of the model.
For example if the coin isn't tossed but always put Heads, and we know it then the model isn't applicable anymore. Instead we have a different model where P(Heads) =1 and P(Tails) = 0. But suppose we do not know that the coin isn't tossed. Now there is a sense in which the random sample model is applicable - it correctly represents our knowledge state about the problem, and a sense in which it is not - it doesn't correctly approximate the causal process that produces Heads/Tails outputs - our knowledge state is itself inadequate. Therefore, to make correct inferences about reality our knowledge state has to correctly approximate causal process that produces the outcomes.
Is the idea just that of the people who you might be based on your current evidence, you reason as if you're randomly drawn from them, and then do probability judgments about theories based on the probability of creating those peopel?
The idea is that the whole notion of "people who you might be" depends on the setting of probability experiment and your knowledge state of it. Just like when the coin is tossed it might be either Heads or Tails. But when the coin is always put Heads, then it might be only Heads.
Suppose that you lived in the universe where all people come into existence with their full fledged personalities in a puff of smoke directly created by God himself. And the holy scripture, that is completely correct in every other way, precisely describing the working of this universe, claims that God intends to create some limited number of very specific people, but in no particular order. In such universe Doomsday Argument is correct. According to your best knowledge about the mechanism of your creation, it was possible for you to be created at any time, to have any birth rank, therefore the fact that you has the specific birth rank that you have tells you something about the total amount of people in the universe. Here your "reference class" is all people in this universe through time.
However, if the infallible scripture claimed that God cares about the order of people created and intends to create every person at their particular place and time, the situation would be very different. Now, according to your knowledge, there is no way for you to be created with any other birth rank, but the one you have. It's like having the coin being put Heads, instead of tossing. Therefore your "reference class" is only you yourself, and you do not get any information from learning your birth rank.
I fell asleep on my beach towel and when I woke I was a 6 foot long half lobster man.
Maybe I'm missing something here, but it just seems intuitive to me that the reference class should be all relevant beings who I could possibly be if I knew nothing more than the fact that I existed. And then in that case, isn't it just an empirical question as to whether or not a certain being like a chimp or a lobster or a lobster-human hybrid is a candidate for that? I guess it depends on your larger views about personal identity and relative sorts of cognition between species, but I definitely feel like, if I were to somehow wake up with no memories in a sensory deprivation tank that erased all knowledge of the outside world beyond my bare self-awareness, I would think it was possible that I was a woman, or an old man, or really any human being on the planet - whereas I'm pretty confident I could still be sure I wasn't, like, a fish. Couldn't a sufficiently advanced neurobiological model give us at least some guidelines for determining who is and isn't in our reference class in that case?
But whether someone could have been you isn't a question that has a precise answer. Could you have been a woman? A neanderthal? In what sense, if you don't believe in souls, could you have been someone else?
Well, the question isn't a counterfactual one about who I "could have been" - it's a question about who my existence is compatible with me being. Would you agree that, even if you had no knowledge whatsoever about anything other than your bare first-person perspective, you could know immediately that you weren't a fish (assuming, I guess, you also know basic facts about what humans are like and what fish are like)? It seems to me that a fish's first-person perspective would be qualitatively different than mine in such a way that my mere "experience of experiencing" rules it out as a possibility. Whereas I definitely think I would have no way of knowing if I was a human woman in that same situation. Neanderthal is in the middle, and I currently have no idea - but it seems plausible to me that, at some point in the future, we could know enough about neurobiology to determine that!
That makes SSA completely lack content https://joecarlsmith.com/2021/09/30/sia-ssa-part-1-learning-from-the-fact-that-you-exist#iii-can-t-we-just-use-the-minimal-reference-class
I think even if the reference class for SSA is set to "anyone I could possibly be" (which is to be understood in whichever sense that SIA'ers mean the term), there's still is an asymmetry here. That is, even with this understanding, on SSA you'll overwhelmingly reject L(T) over L(T-1) (where T is the precise threshold level at which the lobsters become "someone you could possibly be") given your observed evidence for your own humanity (and given your perfect knowledge that T is the exact threshold), and on SIA you won't.
Having thought about this more, though, I think you can get rid of some of the discontinuity by taking into account our level of uncertainty about where exactly the threshold T lies. This might not get rid of all the discontinuity, though (I'm less sure). And I'm also not sure an SSA'er would really mind biting the bullet here in any case.
But this implies that if you knew where the reference class is you get extreme presumptuousness.
That's true, but that's also an extraordinarily unrealistic assumption which is probably partly serving as the intuition pump (although maybe not wholly). Realistically, if we discover an alien species, we'll probably never have substantially more confidence that having a billion neurons gets them into our reference class than that having 999,999,999 neurons does.
Well, SSA reference classes have the problem of being totally made-up bullshit, so it's hard to know how we'd find out what they are. But it implies that an ideal reasoner who could figure out her refrenece class should be that presumptuous.
Additionally, this doesn't solve the problem because there will necessarily be a hard break between things that are maybe in the reference class and things that are definitely not.
>Well, SSA reference classes have the problem of being totally made-up bullshit
Well, if you reject "person I could/couldn't be" as a valid binary distinction, it's not clear how you're going to get SIA off the ground; if by contrast you grant it, it seems like something the SSA advocate should be able to appeal to to ground their reference class.
>But it implies that an ideal reasoner who could figure out her refrenece class should be that presumptuous.
Yes, I agree, but I think someone might just say that it's fine for sufficiently implausible assumptions to lead to implausible conclusions, and this would be a case of such. In reality, there's always going to be a level of certainty in reference-class-fulfilling-ness that depends continuously on the physical facts.
>Additionally, this doesn't solve the problem because there will necessarily be a hard break between things that are maybe in the reference class and things that are definitely not.
I'm not sure how this objection works. Everything gets a certain probability of being in the reference class (= "creature you could've been," whatever that means), and this varies smoothly with other kinds of information you have about them. No hard cutoff between definitely-maybe-in-our-reference-class and definitely-not-in-our-reference-class need be drawn.
//Well, if you reject "person I could/couldn't be" as a valid binary distinction, it's not clear how you're going to get SIA off the ground//
SIA doesn't reference people I could have been, it references people I might currently be. But your reference class can't be people you might currently be for the reasons Carlsmith explains in part 3 https://forum.effectivealtruism.org/posts/pRBDPFuDoyEPz7hes/sia-greater-than-ssa-part-1-learning-from-the-fact-that-you#III__Can_t_we_just_use_the_minimal_reference_class_
//Yes, I agree, but I think someone might just say that it's fine for sufficiently implausible assumptions to lead to implausible conclusions//
Idk, doesn't seem that weird. Also, this is structurally identical to the presumptuous philosopher argument, so if that's a problem for SIA, this is a problem for SSA.
//I'm not sure how this objection works. Everything gets a certain probability of being in the reference class (= "creature you could've been," whatever that means), and this varies smoothly with other kinds of information you have about them.//
Presumably there are some things that you should think don't have any chance of being in the reference class--e.g. rocks. So then somewhere from rocks to people, there's a hard jump to things that could be in your reference class.
One way around this is to appeal to a multiverse theory. If the landscape multiverse theory is correct, then there will be an infinite number of lobster-abominations and an infinite number of humans, so it becomes unclear that we should privilege the lobster-abominations over the humans.
Second, we need to take into account meta-credences. I think it’s coherent for a proponent of SSA to argue that they have a prior for SSA that’s about .9, and that in the lobster-abomination case, they would renounce belief in SSA. It’s not as though SSA proponents have to regard belief in SSA as sacrosanct. It’s perfectly cogent to maintain that one is pretty sure that SSA is right, but that hypothetical evidence like the lobsters (maybe coupled with defeaters for the multiverse) would change one’s mind.
You have not understood the argument. The argument is that SSA implies that an ideal reasoner should, in the case described in the article, have extreme presumptuousness about the number of neurons had by the lobsters.
So what I’m saying is that:
1) In the multiverse case, the SSA proponent thinks the reference class of observers consists of almost infinite (e.g. 10^500) humans + almost infinite (more neurons) lobsters, so whether this sextillion class of entities ends up in the reference class (whether they have more or less neurons) makes no substantive difference, and
2) The credences for more or less neurons affects the credences in SSA itself. If we have good prior grounds (99%), as you state in the article, to think the more neuron hypothesis is true, then the ideal reasoner can rationally abandon SSA. In other words, unless their prior for SSA is absurdly high, there is no reason to think that SSA believers have to stubbornly cling to the less neuron hypothesis
The intention of this comment isn’t to defend SSA -- let alone against such gruesome human-lobster combos as might exist -- but (this also seems like one of Ape in the coat’s points) wouldn’t an SSAer simply deny that one’s reference class would include such creatures, e.g. because only humans (or, if SSA isn’t to rule out the likes of simulated people and brains in vats, entities subjectively indistinguishable from humans) are candidates for inclusion in one’s reference class? To be sure, this might strike you as unprincipled or ad hoc reference-class gerrymandering, but that’s a charge the SSAer has to contend with in any case and they could reasonably complain that it doesn’t add to the objection to bring it against them in a roundabout way.
Well, SSA doesn't say that only people with your exact experiences are in your reference class--if it did, SSA would lack any updates, because it would maximize the share of people with your experiences with your experiences.
Would humans missing 10 neurons be in our reference class? However you draw the boundary around humans, you get hypersensitivity.
It’s plausibly a vague matter what counts as a human. (If you like, read ‘human’ throughout as ‘entity subjectively indistinguishable from a human’. Taking humans as the boundary of one’s reference class is of course just a plausible-sounding suggestion and not something I think the SSAer needs to commit to.) If you think that vagueness is an epistemic phenomenon, then there might still be hypersensitivity, but happily no-one would ever be in a position analogous to that of the presumptuous philosopher in your article. This hypersensitivity may or may not still be a big cost to SSA -- I’m inclined to think that if it’s troubling here, it’s at least no more troubling than the hypersensitivity that, on the epistemicist’s view, all ordinary vague predicates display, modulo the residual problem of reference-class ad hockery. If you instead take vagueness (in this instance; I suppose someone might be an epistemicist about some vague predicates and not others) to be a non-epistemic phenomenon, then SSA doesn’t give you hypersensitivity, but it might be genuinely indeterminate whether or not someone is included in the SSAer’s reference class. Again you may or may not take this to be a big problem for SSA, but fwiw my view is that wouldn’t be as significant a problem as hypersensitivity. (I’m setting aside degree theories of vagueness here, if only because I haven’t thought through how exactly the SSAer should modify their account to accomodate them.)
What's your reason for rejecting necessitarianism?
Seems like things could have been otherwise and I don't think it has any advantage.
I'm not sure I quite understand how SIA avoids this issue. I have certain evidence about myself - for example, that I have hands instead of claws, skin instead of an exoskeleton, etc. Call this evidence V. It seems plausible that P(human|V) ought to be ~1, where "human" there is shorthand for the statement "I am a human."
SIA states that P(human|V) = E[# of human observers with V] / E[# total observers with V]. At L(0), the lobsters will be different enough from us that they won't have V, hence they won't figure into the numerator or denominator at all. (For example, they won't have perceptual evidence that they have hands instead of claws.) Around L(max), they will have V, so they will factor in to the denominator. At some point in the middle, something has to change, right? Maybe you have less ambiguity about reference class, but you still have ambiguity about "evidence."
SIA only tells you to think there are more observers that *you currently might be*. I has nothing to say about how many lobsters there are--or even human lobster combos that you know yourself not to be.
I'm not sure I understand. It seemed like your objection was that on SSA, there's some threshold T such that we should automatically highly favor the hypothesis that the creatures are L(T-1) rather than L(T), because otherwise the probability that we ourselves would be humans would be too small. And that's supposed to be absurd, because the difference between L(T-1) creatures and L(T) creatures is so small. Did I understand correctly?
Because if so, it seems like SIA gets you something similar based on the continuum of evidence the creatures could have about themselves, which is very dissimilar to ours on the low end of L and perhaps identical on the high end. Is that not the case? Do the creatures' evidence very severely differ from ours even at L(max)? Because in *that* case, SSA will not have the consequence you say it has (that we should reject L(T) in favor of L(T-1) due to the surprise of being human).
Well, on SIA the only people who count are people who we might currently be. Now, we'll probably not be sure of how advanced one will have to be in order to be people we might currently be--and that will depend on our evidence. Let's consider three scenarios:
1) We don't know where the threshold is for beings being sufficiently similar that we might currently be them. Then you don't get hypersensitivity because we have a credence split among the lowest threshold for what we might be (maybe you do get it, but that would just be for Sorites paradox reasons--the point is if we think that there's a higher probability that we currently are at higher ends of the spectrum, then we'll get a bias in favor of bigger ends of the spectrum). But this is importantly different from the result for SSA because it's about beings that we currently might be, rather than far away aliens. SIA doesn't care about what far away aliens are like, because they're not people you might currently be.
2) We know exactly what threshold of creature we are--say we're L(1 billion). Well then SIA says that you should be confident the galaxy being populated by creatures of L(1 billion)--but that's perfectly non-objectionable.
Maybe the worry for SIA is supposed to be this: let N be the minimum number such that you might currently be L(N). SIA tells you to be very confident that the universe is populated with creatures of L(N) rather than L(N-1). But this isn't implausible at all. Here's an analogy. Suppose I don't know exactly what color my shirt is. Imagine a plain white shirt is S(0), plain black is S(10,000), and each extra number darkens it a bit. I think the lowest possible value of S my shirt might be is S(1,000). There are two possibilities: one is that the universe is populated with people with shirts of S(999) and the other S(1,000). Assume they have equal priors. Well, it's perfectly reasonable to think that S(999) is ruled out by this--it flatly doesn't explain the data.
Well, before I respond, I wanted to get clarification on my question about whether the creatures are supposed to have "our evidence" V near L(max) because they're so similar to us. My main point is that if the answer to this is "no," then it doesn't follow on SSA that P(lobsters are L(T-1) | V) is much higher than P(lobsters are L(T) | V)! And if the answer is "yes," then SIA also has an issue.
They don't have our evidence. But that doesn't matter on SSA. The people in blue shirts don't have the same evidence as the people in red shirts, but nevertheless, SSA reasons as if you're randomly drawn from among all the people both in red and bleu.
I actually think I changed my mind a bit. If V is our evidence of our being human, and V is essentially coextensive with *being* human regardless of the lobsters' L-level, then it's surprising on SSA that we get V on L(T) and not on L(T-1), assuming that's where the reference class line is drawn. So I think you might be right, unless there's something else I'm missing.
Ha, nice one.
I think you've convincingly disproved the type of SSA which assumes that in this kind of probability theory problems correct reference class can be constructed this way. You've proven that there is, in fact, no such L for which lobster people are in my reference class. This doesn't disprove all SSA reasoning, but narrows the available reference class quite a lot.
This is not exactly a new insight. Doomsday inference and Save The Puppy already very strongly hint that drawing the boundaries of reference class across time produce absurd results. And as soon as we outlaw this kind of shenanigans everything adds up to normality.
Of course, the actual rule for "drawing the border of the reference class", or about the anthropic reasoning in general, should be based on the available information about the causal process that led to my existence. The disagreement between scientists and anthropic-reasoners shouldn't happen at all - the latter are supposed to use the up-to-date scientific knowledge to determine the correct reference class instead of just assuming that it's L = 700 millions out of nowhere. And if I do this here, well seems that my reference class is trivial - there is exactly one person born as a result of sexual encounter between my parents at the specific date of my birth. So I'm a random sample of one, and my existence, unsurprisingly enough, tells me absolutely nothing about the human-lobster hybrids.
Can you explain what you mean by "should be based on the available information about the causal process that led to my existence?"
Sure. Consider how we treat a coin toss as a random sample from {Tails, Heads}, P(Tails) = P(Heads) = 1/2. Why do we do that? It's not that coin tosses are physically equiprobable. If we knew all the physical forces affecting the coin at the moment of the toss we would be able to predict the outcome via deterministic model. But usually we do not know it. And so we *approximate* actual physical process this way via a mathematical model based on our limited state of knowledge about the coin toss. This model still corresponds to reality, but it's less accurate. And both changes to the actual causal process that makes the toss, and our knowledge state affect the applicability of the model.
For example if the coin isn't tossed but always put Heads, and we know it then the model isn't applicable anymore. Instead we have a different model where P(Heads) =1 and P(Tails) = 0. But suppose we do not know that the coin isn't tossed. Now there is a sense in which the random sample model is applicable - it correctly represents our knowledge state about the problem, and a sense in which it is not - it doesn't correctly approximate the causal process that produces Heads/Tails outputs - our knowledge state is itself inadequate. Therefore, to make correct inferences about reality our knowledge state has to correctly approximate causal process that produces the outcomes.
Is the idea just that of the people who you might be based on your current evidence, you reason as if you're randomly drawn from them, and then do probability judgments about theories based on the probability of creating those peopel?
The idea is that the whole notion of "people who you might be" depends on the setting of probability experiment and your knowledge state of it. Just like when the coin is tossed it might be either Heads or Tails. But when the coin is always put Heads, then it might be only Heads.
Suppose that you lived in the universe where all people come into existence with their full fledged personalities in a puff of smoke directly created by God himself. And the holy scripture, that is completely correct in every other way, precisely describing the working of this universe, claims that God intends to create some limited number of very specific people, but in no particular order. In such universe Doomsday Argument is correct. According to your best knowledge about the mechanism of your creation, it was possible for you to be created at any time, to have any birth rank, therefore the fact that you has the specific birth rank that you have tells you something about the total amount of people in the universe. Here your "reference class" is all people in this universe through time.
However, if the infallible scripture claimed that God cares about the order of people created and intends to create every person at their particular place and time, the situation would be very different. Now, according to your knowledge, there is no way for you to be created with any other birth rank, but the one you have. It's like having the coin being put Heads, instead of tossing. Therefore your "reference class" is only you yourself, and you do not get any information from learning your birth rank.