This tracks what I would strongly expect to be the church's position, and I can't imagine that we would get to the point of passing any kind of laws protecting the status of digital minds without the church having a position. This strikes me as reason more to work to figure out how our goals align with the church, rather than try to fight them on this.
To manage its own limited processing power, the brain uses attention to focus on specific signals.
What do you think about AI systems that don't have the kinds of limitations we do, and that might not need to restrict focus as much, or might be able to devote cognitive resources to many different problems in parallel. It is plausible that they wouldn't count as having attention in the way that Graziano thinks we do, and that they wouldn't therefore satisfy his theory of consciousness. But it seems weird to suggest that the reason our preferences matter is because of our cognitive limitations, no?
From your perspective now, given your uncertainty, you might have no reason to think that the drive changes things in a better or worse way, and so the expected value derived from taking the expected value of each possible consequence considered piecemeal and summed up might be neutral.
I think if you accept some degree of objective chance, this might completely resolve the problem. (I explore this idea, without direct connection to incomparability, here: https://philpapers.org/rec/SHICAT-11)
It is consistent with a purely subjective neutral expected value, however, that you might be quite sure that the results, if revealed in their full detail, would be incomparable.
This creates an intuitive tension. There is something called the "Principle of Reflection" -- it is a bit controversial, but very intuitive -- that says roughly that if you know you would believe something if you were put into a better epistemic position, you should believe it now. The challenge Bulldog raises relies on taking reflection seriously and thinking that we would almost surely see incomparability if we looked closely at all the details, so we should accept incomparability now.
What I defend is that, in the absence of a good argument based on welfare ranges and not p(sentience), we don't know if the welfare range of simpler animals is below or above the bar above which their welfare would dominate over that of more complex animals (not that it is below!).
I suspect we disagree about how it is reasonable to apportion probabilities in the absence of empirical evidence. I think that conditioning on p(sentience) is sufficient to justify a non-negligible probability of similar levels of welfare range in the absence of any further empirical evidence. The empirical evidence that shrimp have small brains detracts from this probability, but not by much. That evidence makes it much more likely that they are not sentient at all, not much more likely that they only feel small feelings.
Which kind of functionalism? I am very sceptical of at least computational functionalism (CF). Any algorithm run by a digital computer can be executed with pen and paper (although it may take a super long time), and I have a hard time imagining how such process would itself be conscious.
I'm skeptical of functionalism about consciousness (though I don't know any alternative that fares better.) But functionalism about valence seems much harder to avoid. Maybe if you have a benevolent God? Or some sort of dualism? Otherwise, it seems to me that you're going to be hard-pressed to explain why it is that the functional role of valence aligns with whatever properties constitute the fact that valence matters. Why is it that pain is bad and we avoid it, or pleasure is good and we seek it, if it is not just the case that the things we're inclined to avoid count as pain, and the things we're inclined to seek count as pleasure (very roughly).
However, I think you mean that the expected welfare range will be significant (for example, at least 1 % of that of humans) as long as there is one plausible model (for example, which gets 10 % weight) which predicts a significant welfare range (for example, 10 % of that of humans).
Yeah, basically.
I wonder whether electromagnetic (EM) field theories of consciousness could shed some light on it. I assume the maximum intensity of the EM fields generated by brain activity depends on the number of neurons, at least when assessed across species (there is little variance in the number of neurons of humans, which means the maximum intensity of EM fields may not vary much in humans).
This makes me think we're inclined towards a different basic perspective on the determinants of valence. This kind of sounds like you're thinking of pain as a sort of physical magnitude, like weight or charge. Then it is reasonable to think it is likely to scale with size, so that much smaller brains are likely to have much smaller magnitudes. I'm more inclined towards functionalist interpretations of welfare, on which something like relative functional significance determines welfare levels. E.g. something's attention-grabbing capacity helps to determine its welfare significance. In that case, you might be deeply skeptical that small animals have the right functional role at all, but once you grant they do, it is much more plausible that welfare ranges are similar to humans. However, for my point to be right, I think you just need to treat these kinds of functionalist views as in the running. You don't have to be confident that they're true.
. For the conclusion to follow, (ii) needs to be: evidence of a welfare range that is not too insignificant to crucially undermine the importance of their large numbers.
I disagree with this. You don't need evidence of a welfare range that is not too insignificant -- you need a presumption of a reasonable probability of a welfare range that is not too small and no significant evidence against it. Without a theory of how you get from neurons to welfare, it seems to me to be very unreasonable to be confident that simpler brains most likely have much smaller welfare ranges. And we don't have a good theory of how to get from neurons to welfare.
One thing I've found for myself: it is easy to think of lifestyle changes across the board, but different areas have very different impacts on your budget. Opting for the slightly nicer apartment may end up costing way more than switching to the fancy vegan yogurt. Insofar as lifestyle creep is reasonable (it is probably healthy to be increasing your quality of life a bit over time), it is better to focus on the little things than the big ones.
AIs could have negligible welfare (in expectation) even if they are conscious. They may not be sentient even if they are conscious, or have negligible welfare even if they are sentient. I would say the (expected) total welfare of a group (individual welfare times population) matters much more for its moral consideration than the probability of consciousness of its individuals. Do you have any plans to compare the individual (expected hedonistic) welfare of AIs, animals, and humans? You do not mention this in the section "What’s next".
This is an important caveat. While our motivation for looking at consciousness is largely from its relation to moral status, we don't think that establishing that AIs were conscious would entail that they have significant states that counted strongly one way or the other for our treatment of them, and establishing that they weren't conscious wouldn't entail that we should feel free to treat them however we like.
We think that it estimates of consciousness still play an important practical role. Work on AI consciousness may help us to achieve consensus on reasonable precautionary measures and motivate future research directions with a more direct upshot. I don't think the results of this model can be directly plugged into any kind of BOTEC, and should be treated with care.
Do you have any ideas for how to decide on the priors for the probability of sentience? I agree decisions about priors are often very arbitrary, and I worry they will have significantly different implications.
We favored a 1/6 prior for consciousness relative to every stance and we chose that fairly early in the process. To some extent, you can check the prior against what you update to on the basis of your evidence. Given an assignment of evidence strength and an opinion about what it should say about something that satisfies all of the indicators, you can backwards infer the prior needed to update to the right posterior. That prior is basically implicit in your choices about evidential strength. We didn't explicitly set our prior this way, but we would probably have reconsidered our choice of 1/6 if it was giving really implausible results for humans, chickens, and ELIZA across the board.
The right conclusion would be that the respondants have no idea about the right exponent, or how to weight the various models because they would not be able to adequately justify their picks.
There is a tension here between producing probabilities we think are right and producing probabilities which could reasonably act as a consensus conclusion. I have my own favorite stance, and I think I have good reason for it, but I didn't try to convince anyone to give it more weight in our aggregation. Insofar as we're aiming in the direction of something that could achieve broad agreement, we don't want to give too much weight to our own views (even if we think we're right). Unfortunately,among people with significant expertise in this area, there is broad and fairly fundamental disagreement. We think that it is still valuable to shoot for consensus, even if that means everyone will think it is flawed (by giving too much weight to different stances.)
This last part carries a lot of weight; a simulacrum, when dormant in the superposition from which it can be sampled, is nonexistent. A simulacrum only exists during the discrete processing event which correlates with its sampling.
There seems to me to be a sensible view on which a simulacrum exists to the extent that computations relevant to making decisions on its behalf are carried out, regardless of what the token sampler chooses. This would suggest that there could conceivably be vast numbers of different simulacra instantiated even in a single forward pass.
One odd upshot of requiring the token sampler is that in contexts in which no tokens get sampled (prefill, training) you can get all of the same model computations but have no simulacra at all.
This tracks what I would strongly expect to be the church's position, and I can't imagine that we would get to the point of passing any kind of laws protecting the status of digital minds without the church having a position. This strikes me as reason more to work to figure out how our goals align with the church, rather than try to fight them on this.
What do you think about AI systems that don't have the kinds of limitations we do, and that might not need to restrict focus as much, or might be able to devote cognitive resources to many different problems in parallel. It is plausible that they wouldn't count as having attention in the way that Graziano thinks we do, and that they wouldn't therefore satisfy his theory of consciousness. But it seems weird to suggest that the reason our preferences matter is because of our cognitive limitations, no?
From your perspective now, given your uncertainty, you might have no reason to think that the drive changes things in a better or worse way, and so the expected value derived from taking the expected value of each possible consequence considered piecemeal and summed up might be neutral.
I think if you accept some degree of objective chance, this might completely resolve the problem. (I explore this idea, without direct connection to incomparability, here: https://philpapers.org/rec/SHICAT-11)
It is consistent with a purely subjective neutral expected value, however, that you might be quite sure that the results, if revealed in their full detail, would be incomparable.
This creates an intuitive tension. There is something called the "Principle of Reflection" -- it is a bit controversial, but very intuitive -- that says roughly that if you know you would believe something if you were put into a better epistemic position, you should believe it now. The challenge Bulldog raises relies on taking reflection seriously and thinking that we would almost surely see incomparability if we looked closely at all the details, so we should accept incomparability now.
I suspect we disagree about how it is reasonable to apportion probabilities in the absence of empirical evidence. I think that conditioning on p(sentience) is sufficient to justify a non-negligible probability of similar levels of welfare range in the absence of any further empirical evidence. The empirical evidence that shrimp have small brains detracts from this probability, but not by much. That evidence makes it much more likely that they are not sentient at all, not much more likely that they only feel small feelings.
I'm skeptical of functionalism about consciousness (though I don't know any alternative that fares better.) But functionalism about valence seems much harder to avoid. Maybe if you have a benevolent God? Or some sort of dualism? Otherwise, it seems to me that you're going to be hard-pressed to explain why it is that the functional role of valence aligns with whatever properties constitute the fact that valence matters. Why is it that pain is bad and we avoid it, or pleasure is good and we seek it, if it is not just the case that the things we're inclined to avoid count as pain, and the things we're inclined to seek count as pleasure (very roughly).
Yeah, basically.
This makes me think we're inclined towards a different basic perspective on the determinants of valence. This kind of sounds like you're thinking of pain as a sort of physical magnitude, like weight or charge. Then it is reasonable to think it is likely to scale with size, so that much smaller brains are likely to have much smaller magnitudes. I'm more inclined towards functionalist interpretations of welfare, on which something like relative functional significance determines welfare levels. E.g. something's attention-grabbing capacity helps to determine its welfare significance. In that case, you might be deeply skeptical that small animals have the right functional role at all, but once you grant they do, it is much more plausible that welfare ranges are similar to humans. However, for my point to be right, I think you just need to treat these kinds of functionalist views as in the running. You don't have to be confident that they're true.
I disagree with this. You don't need evidence of a welfare range that is not too insignificant -- you need a presumption of a reasonable probability of a welfare range that is not too small and no significant evidence against it. Without a theory of how you get from neurons to welfare, it seems to me to be very unreasonable to be confident that simpler brains most likely have much smaller welfare ranges. And we don't have a good theory of how to get from neurons to welfare.
One thing I've found for myself: it is easy to think of lifestyle changes across the board, but different areas have very different impacts on your budget. Opting for the slightly nicer apartment may end up costing way more than switching to the fancy vegan yogurt. Insofar as lifestyle creep is reasonable (it is probably healthy to be increasing your quality of life a bit over time), it is better to focus on the little things than the big ones.
This is an important caveat. While our motivation for looking at consciousness is largely from its relation to moral status, we don't think that establishing that AIs were conscious would entail that they have significant states that counted strongly one way or the other for our treatment of them, and establishing that they weren't conscious wouldn't entail that we should feel free to treat them however we like.
We think that it estimates of consciousness still play an important practical role. Work on AI consciousness may help us to achieve consensus on reasonable precautionary measures and motivate future research directions with a more direct upshot. I don't think the results of this model can be directly plugged into any kind of BOTEC, and should be treated with care.
We favored a 1/6 prior for consciousness relative to every stance and we chose that fairly early in the process. To some extent, you can check the prior against what you update to on the basis of your evidence. Given an assignment of evidence strength and an opinion about what it should say about something that satisfies all of the indicators, you can backwards infer the prior needed to update to the right posterior. That prior is basically implicit in your choices about evidential strength. We didn't explicitly set our prior this way, but we would probably have reconsidered our choice of 1/6 if it was giving really implausible results for humans, chickens, and ELIZA across the board.
There is a tension here between producing probabilities we think are right and producing probabilities which could reasonably act as a consensus conclusion. I have my own favorite stance, and I think I have good reason for it, but I didn't try to convince anyone to give it more weight in our aggregation. Insofar as we're aiming in the direction of something that could achieve broad agreement, we don't want to give too much weight to our own views (even if we think we're right). Unfortunately,among people with significant expertise in this area, there is broad and fairly fundamental disagreement. We think that it is still valuable to shoot for consensus, even if that means everyone will think it is flawed (by giving too much weight to different stances.)
There seems to me to be a sensible view on which a simulacrum exists to the extent that computations relevant to making decisions on its behalf are carried out, regardless of what the token sampler chooses. This would suggest that there could conceivably be vast numbers of different simulacra instantiated even in a single forward pass.
One odd upshot of requiring the token sampler is that in contexts in which no tokens get sampled (prefill, training) you can get all of the same model computations but have no simulacra at all.