On the macro-level issue of priorities, I’ve gathered some of my thoughts here.
Thanks for sharing, Bob! I would publish that as a post.
we both agree that we don’t think soil nematodes and some other soil invertebrates have enough likelihood of being sentient to be a high priority
Which are the soil invertebrates whose probability of sentience you think is too low for them to be a high priority? I assume soil nematodes, mites, and springtails, as Bob says “We have nothing like the evidence for sentience in Drosophila for mites, springtails, thrips, and the like”. Do you expect the probability of sentience to drop much more from black soldier fly (BSF) larvae to soil springtails, mites, and nematodes than from chickens to BSF larvae? I do not understand why one would expect this. Gemini guesses (random) soil nematodes, mites, and springtails have 300, 1.1 k, and 3.8 k neurons. Adult nematodes of the species C. elegans have 302 neurons. BSF larvae have 2.3 k to 20 k neurons depending on their stage of development (“L1: 2,300 [brain] cells [neurons]; L6: 20,000 cells”). So I believe BSF larvae have around 11.2 k neurons (= (2.3 + 20)/2*10^3), 37.3 (= 11.2*10^3/300) times those of soil nematodes, 10.2 (= 11.2*10^3/(1.1*10^3)) times those of soil mites, or 2.95 (= 11.2*10^3/(3.8*10^3)) times those of soil springtails. In contrast, chickens, as proxied by the red junglefowl, have 221 M neurons, 19.7 k (= 221*10^6/(11.2*10^3)) those of BSF larvae. In other words, the number of neurons decreases 4.29 (= log10(19.7*10^3)) orders of magnitude (OOMs) from chickens to BSF larvae, but only 0.470 (= log10(2.95)) OOMs from BSF larvae to soil springtails, and 1.10 (= log10(3.8*10^3/300)) OOMs from soil springtails to nematodes. “Number of neurons”^0.188 explains pretty well the welfare ranges in Bob’s book about comparing welfare across species across 6 OOMs of the number of neurons (RP estimated shrimps have 10^-6 as many neurons as humans; see Table 5), as illustrated below. This is many more OOMs than those from BSF larvae to soil springtails, mites, and nematodes.
I [Bob] doubt there’s a linear relationship between sentience and any such neurophysiological feature; at some point, I’m inclined to think that the probability falls off a cliff.
Bob, do you mean the probability of sentience (or absolute value of the expected individual welfare per animal-year) is proportional to “number of neurons”^a, where a < 1, for a sufficiently high number of neurons, but that such power law dramatically overestimates the probability of sentience of soil nematodes, mites, and springtails? If so, why would the probability of sentience fall off a cliff over the 0.470 OOMs of the number of neurons from BSF larvae to soil springtails, but not over the 6 OOMs from humans to shrimps? I understand there is much less evidence for the sentience of springtails than for that of shrimp or BSFs. However, one should account not only for the evidence available now, but also for what would be found out given further investigation. Some decades ago, someone could have produced the graph above without shrimps and BSFs, and guessed the probability of sentience would fall off a cliff somewhere between carps and shrimps or BSFs, whereas assuming the probability of sentience is a power law of the number of neurons (or other neurophysiological feature) would have predicted much better your results.
Moreover, one should be sensitive not only to differences in the probability of sentience, but also in population, even if one does not strongly endorse maximising expected welfare (as I do). The number of individuals increases astronomically from farmed BSF larvae to soil springtails, mites, and nematodes, but little from chickens to farmed BSF larvae. Iestimate there are 83.0 M (= 3.17*10^18/(3.82*10^10)), 166 M (= 6.33*10^18/(3.82*10^10)), and 12.8 billion (= 4.89*10^20/(3.82*10^10)) times as many soil springtails, mites, and nematodes as farmed BSF larvae and mealworms, but only 1.40 (= 3.82*10^10/(2.72*10^10)) times as many farmed BSF larvae and mealworms as chickens.
I’m sorry that I don’t have time to respond to all your questions, Vasco. The short version, though, is that I also want robustness in the case for sentience, so I’m much less inclined to make the kinds of extrapolations you’re suggesting here. I have the same view about our moral weight work: I put very little stock in any specific numbers, as I think that plausible moral weights will be defensible from several angles, each of which will suggest somewhat different estimates, with no obvious right way to aggregate them. (Again, there’s that skepticism about expected value!)
I don’t think the case for Vasco’s argument depends really on sentience in non-arthropods. There are like a billion soil arthropods for every person, so funding research on soil animals looks similarly important. And a lot of these are ants who are more likely to be sentient than black soldier flies.
I do find the comment “I also want robustness in the case for sentience,” a bit puzzling in context. As I understood it, Vasco’s argument was that it’s not very unlikely that animals even simpler than arthropods are sentient (mites, springtails, etc). That there’s not a robust case that they are analogous isn’t a strong argument for them not being analogous (and will, in fact, be a reason for uncertainty and research).
Broadly agree with a lot of the document though, especially the funding stuff! I think funding Arthropoda is great!
I also want robustness in the case for sentience, so I’m much less inclined to make the kinds of extrapolations you’re suggesting here
I guess you are saying you prefer narrower distributions of the probability of sentience holding the expected probability of sentience constant. This contributes towards you prioritising BSF larvae over soil springtails, mites, and nematodes (all else equal) because there is more uncertainty about the probability of sentience of the latter. What matters for me is expected welfare, which only depends on the expected probability of sentience. However, I think one would have to value narrower distributions of the probability of sentience to an unreasonable amount to prioritise BSF larvae and mealworms over soil springtails, mites, and nematodes if one prioritises farmed BSF larvae and mealworms over chickens. The distribution of the probability of sentience of chickens is much narrower than that of farmed BSF larvae and mealworms, and I estimate the population of farmed BSF larvae and mealworms is just 1.40 times that of chickens (see my last comment). In contrast, I believe the distribution of the probability of sentience of farmed BSF larvae and mealworms is not that much narrower than that of soil springtails, mites, and nematodes, and I estimate the population of soil springtails, mites, and nematodes is 83.0 M, 166 M, and 12.8 billion times that of farmed BSF larvae and mealworms (see my last comment).
Furthermore, Iestimate the population of soil ants and termites is 1.31 M (= 5*10^16/(3.82*10^10)) and 2.62 M (= 1*10^17/(3.82*10^10)) times that of farmed BSF larvae and mealworms, and I suspect the probability of sentience of soil ants and termites is both higher in expectation and narrower, as I calculate soil ants and termites have 22.3 (= 250*10^3/(11.2*10^3)) and 8.93 (= 100*10^3/(11.2*10^3)) times as many neurons as BSF larvae and mealworms. In any case, I would prioritise soil ants and termites even if I thought the distribution of the probability of their sentience was much wider than that of farmed BSF larvae and mealworms. I would just focus on research on the sentience of soil ants and termites instead of ways of increasing their welfare.
I have the same view about our moral weight work: I put very little stock in any specific numbers
Me too. Yet, I do not think the points I am making depend on a specific way of comparing welfare across species. I estimate the absolute value of the total welfare of soil animals is many OOMs larger than that of farmed BSF larvae and mealworms for indidividual welfare per animal-year proportional to “number of neurons”^”exponent of the number of neurons”, and this exponent ranging from 0 to 2 (see the 1st graph below), which covers dramatically different welfare ranges (see the 2nd graph below).
Thanks, all. Let me add something that may help clarify why we’re always at loggerheads. I’m not actually thinking about these questions in probabilistic terms at all. In my view, the evidential situation for most arthropods is so sparse that I don’t actually believe we’re in a position to assign meaningful probabilities of sentience—even extremely rough ones. We’re squarely in the domain of the precautionary, not the probabilistic. When the evidence is this patchy and the mechanisms this poorly understood, numerical probability assignments feel more like artifacts of modeling choices than reflections of the world. So, when I talk about “robustness,” I’m not covertly appealing to narrower or wider probability distributions; I’m saying that the entire framework of attaching numbers to these uncertainties feels inappropriate.
This is one of several reasons why focusing on well-studied insects makes sense to me. It’s not that I think BSF larvae are 10× or 100× more likely to be sentient than springtails. It’s that we have a type of evidence for some insects—convergent behavioral, physiological, and neuroanatomical findings—that simply doesn’t exist at all for mites, springtails, and nematodes. And without that evidential base, I’m wary of using a first-pass model to set priorities. Expected value becomes extremely fragile under those conditions, as the inputs aren’t grounded: they’re guesses stacked on guesses.
So the way I think about prioritization has less to do with estimated probabilities and more to do with where precautionary reasoning can actually get traction. Work on farmed and research arthropods produces immediate welfare improvements, helps develop welfare indicators, and builds the scientific ecosystem we’ll need if we ever hope to understand smaller arthropods. That’s a much more stable basis for action than trying to set priorities via BOTECs.
Anyway, we’ll just have to agree to disagree, as we just keep running up against the same issues over and over!
Thanks for the clarifications, Bob! @Bentham’s Bulldog, you may be interested in the 2 comments above from Bob.
In my view, the evidential situation for most arthropods is so sparse that I don’t actually believe we’re in a position to assign meaningful probabilities of sentience—even extremely rough ones. We’re squarely in the domain of the precautionary, not the probabilistic. [...]
This is one of several reasons why focusing on well-studied insects makes sense to me. It’s not that I think BSF larvae are 10× or 100× more likely to be sentient than springtails. It’s that we have a type of evidence for some insects—convergent behavioral, physiological, and neuroanatomical findings—that simply doesn’t exist at all for mites, springtails, and nematodes. And without that evidential base, I’m wary of using a first-pass model to set priorities. Expected value becomes extremely fragile under those conditions, as the inputs aren’t grounded: they’re guesses stacked on guesses.
In this case, I feel like it would also be reasonable to argue that the evidential situation with respect to comparing the individual welfare per animal-year (not probability of sentience) of different species is so sparse that one should just focus on increasing the welfare of vertebrates. At least from my perspective, any comparison of the welfare (not probability of sentience) of shrimps with that of humans involves “guesses stacked on guesses”.
In addition, I see the lack of robust evidence for the sentience of soil springtails, mites, and nematodes as a case for further research on their sentience (although I would be surprised if it updated me towards thinking their expected individual welfare per animal-year is much lower than suggested by “number of neurons”^”exponent”). At some point, there was not robust evidence for the sentience of BSF larvae.
In any case, I assume the points about the robustness of evidence do not apply to soil ants and termites.
So the way I think about prioritization has less to do with estimated probabilities and more to do with where precautionary reasoning can actually get traction. Work on farmed and research arthropods produces immediate welfare improvements, helps develop welfare indicators, and builds the scientific ecosystem we’ll need if we ever hope to understand smaller arthropods.
This makes sense to me. I ranked Arthropoda 1st in the Donation Election on that basis. At the same time, I suspect the optimal spending on research on soil animals is not 0. I got no results for “ants”, “termites”, “springtails”, “mites”, or “nematodes” on WAI’s grantees page.
That’s a much more stable basis for action than trying to set priorities via BOTECs.
I wonder whether there are some calculations one could do to compare the cost-effectiveness of building capacity for research on soil animals via doing this directly, or indirectly through research on farmed invertebrates.
Lastly, this article is good. The possibility the they’re right is one of the things that makes me inclined to see insects as the limit case.
Strongly agree about “the evidential situation with respect to comparing the individual welfare per animal-year”! I’ve always taken the numbers from the MWP much less seriously than others. I see that work as one part of a large picture, depending heavily on other arguments.
Thanks for sharing, Bob! I would publish that as a post.
Which are the soil invertebrates whose probability of sentience you think is too low for them to be a high priority? I assume soil nematodes, mites, and springtails, as Bob says “We have nothing like the evidence for sentience in Drosophila for mites, springtails, thrips, and the like”. Do you expect the probability of sentience to drop much more from black soldier fly (BSF) larvae to soil springtails, mites, and nematodes than from chickens to BSF larvae? I do not understand why one would expect this. Gemini guesses (random) soil nematodes, mites, and springtails have 300, 1.1 k, and 3.8 k neurons. Adult nematodes of the species C. elegans have 302 neurons. BSF larvae have 2.3 k to 20 k neurons depending on their stage of development (“L1: 2,300 [brain] cells [neurons]; L6: 20,000 cells”). So I believe BSF larvae have around 11.2 k neurons (= (2.3 + 20)/2*10^3), 37.3 (= 11.2*10^3/300) times those of soil nematodes, 10.2 (= 11.2*10^3/(1.1*10^3)) times those of soil mites, or 2.95 (= 11.2*10^3/(3.8*10^3)) times those of soil springtails. In contrast, chickens, as proxied by the red junglefowl, have 221 M neurons, 19.7 k (= 221*10^6/(11.2*10^3)) those of BSF larvae. In other words, the number of neurons decreases 4.29 (= log10(19.7*10^3)) orders of magnitude (OOMs) from chickens to BSF larvae, but only 0.470 (= log10(2.95)) OOMs from BSF larvae to soil springtails, and 1.10 (= log10(3.8*10^3/300)) OOMs from soil springtails to nematodes. “Number of neurons”^0.188 explains pretty well the welfare ranges in Bob’s book about comparing welfare across species across 6 OOMs of the number of neurons (RP estimated shrimps have 10^-6 as many neurons as humans; see Table 5), as illustrated below. This is many more OOMs than those from BSF larvae to soil springtails, mites, and nematodes.
Bob, do you mean the probability of sentience (or absolute value of the expected individual welfare per animal-year) is proportional to “number of neurons”^a, where a < 1, for a sufficiently high number of neurons, but that such power law dramatically overestimates the probability of sentience of soil nematodes, mites, and springtails? If so, why would the probability of sentience fall off a cliff over the 0.470 OOMs of the number of neurons from BSF larvae to soil springtails, but not over the 6 OOMs from humans to shrimps? I understand there is much less evidence for the sentience of springtails than for that of shrimp or BSFs. However, one should account not only for the evidence available now, but also for what would be found out given further investigation. Some decades ago, someone could have produced the graph above without shrimps and BSFs, and guessed the probability of sentience would fall off a cliff somewhere between carps and shrimps or BSFs, whereas assuming the probability of sentience is a power law of the number of neurons (or other neurophysiological feature) would have predicted much better your results.
Moreover, one should be sensitive not only to differences in the probability of sentience, but also in population, even if one does not strongly endorse maximising expected welfare (as I do). The number of individuals increases astronomically from farmed BSF larvae to soil springtails, mites, and nematodes, but little from chickens to farmed BSF larvae. I estimate there are 83.0 M (= 3.17*10^18/(3.82*10^10)), 166 M (= 6.33*10^18/(3.82*10^10)), and 12.8 billion (= 4.89*10^20/(3.82*10^10)) times as many soil springtails, mites, and nematodes as farmed BSF larvae and mealworms, but only 1.40 (= 3.82*10^10/(2.72*10^10)) times as many farmed BSF larvae and mealworms as chickens.
I’m sorry that I don’t have time to respond to all your questions, Vasco. The short version, though, is that I also want robustness in the case for sentience, so I’m much less inclined to make the kinds of extrapolations you’re suggesting here. I have the same view about our moral weight work: I put very little stock in any specific numbers, as I think that plausible moral weights will be defensible from several angles, each of which will suggest somewhat different estimates, with no obvious right way to aggregate them. (Again, there’s that skepticism about expected value!)
I don’t think the case for Vasco’s argument depends really on sentience in non-arthropods. There are like a billion soil arthropods for every person, so funding research on soil animals looks similarly important. And a lot of these are ants who are more likely to be sentient than black soldier flies.
I do find the comment “I also want robustness in the case for sentience,” a bit puzzling in context. As I understood it, Vasco’s argument was that it’s not very unlikely that animals even simpler than arthropods are sentient (mites, springtails, etc). That there’s not a robust case that they are analogous isn’t a strong argument for them not being analogous (and will, in fact, be a reason for uncertainty and research).
Broadly agree with a lot of the document though, especially the funding stuff! I think funding Arthropoda is great!
I guess you are saying you prefer narrower distributions of the probability of sentience holding the expected probability of sentience constant. This contributes towards you prioritising BSF larvae over soil springtails, mites, and nematodes (all else equal) because there is more uncertainty about the probability of sentience of the latter. What matters for me is expected welfare, which only depends on the expected probability of sentience. However, I think one would have to value narrower distributions of the probability of sentience to an unreasonable amount to prioritise BSF larvae and mealworms over soil springtails, mites, and nematodes if one prioritises farmed BSF larvae and mealworms over chickens. The distribution of the probability of sentience of chickens is much narrower than that of farmed BSF larvae and mealworms, and I estimate the population of farmed BSF larvae and mealworms is just 1.40 times that of chickens (see my last comment). In contrast, I believe the distribution of the probability of sentience of farmed BSF larvae and mealworms is not that much narrower than that of soil springtails, mites, and nematodes, and I estimate the population of soil springtails, mites, and nematodes is 83.0 M, 166 M, and 12.8 billion times that of farmed BSF larvae and mealworms (see my last comment).
Furthermore, I estimate the population of soil ants and termites is 1.31 M (= 5*10^16/(3.82*10^10)) and 2.62 M (= 1*10^17/(3.82*10^10)) times that of farmed BSF larvae and mealworms, and I suspect the probability of sentience of soil ants and termites is both higher in expectation and narrower, as I calculate soil ants and termites have 22.3 (= 250*10^3/(11.2*10^3)) and 8.93 (= 100*10^3/(11.2*10^3)) times as many neurons as BSF larvae and mealworms. In any case, I would prioritise soil ants and termites even if I thought the distribution of the probability of their sentience was much wider than that of farmed BSF larvae and mealworms. I would just focus on research on the sentience of soil ants and termites instead of ways of increasing their welfare.
Me too. Yet, I do not think the points I am making depend on a specific way of comparing welfare across species. I estimate the absolute value of the total welfare of soil animals is many OOMs larger than that of farmed BSF larvae and mealworms for indidividual welfare per animal-year proportional to “number of neurons”^”exponent of the number of neurons”, and this exponent ranging from 0 to 2 (see the 1st graph below), which covers dramatically different welfare ranges (see the 2nd graph below).
Thanks, all. Let me add something that may help clarify why we’re always at loggerheads. I’m not actually thinking about these questions in probabilistic terms at all. In my view, the evidential situation for most arthropods is so sparse that I don’t actually believe we’re in a position to assign meaningful probabilities of sentience—even extremely rough ones. We’re squarely in the domain of the precautionary, not the probabilistic. When the evidence is this patchy and the mechanisms this poorly understood, numerical probability assignments feel more like artifacts of modeling choices than reflections of the world. So, when I talk about “robustness,” I’m not covertly appealing to narrower or wider probability distributions; I’m saying that the entire framework of attaching numbers to these uncertainties feels inappropriate.
This is one of several reasons why focusing on well-studied insects makes sense to me. It’s not that I think BSF larvae are 10× or 100× more likely to be sentient than springtails. It’s that we have a type of evidence for some insects—convergent behavioral, physiological, and neuroanatomical findings—that simply doesn’t exist at all for mites, springtails, and nematodes. And without that evidential base, I’m wary of using a first-pass model to set priorities. Expected value becomes extremely fragile under those conditions, as the inputs aren’t grounded: they’re guesses stacked on guesses.
So the way I think about prioritization has less to do with estimated probabilities and more to do with where precautionary reasoning can actually get traction. Work on farmed and research arthropods produces immediate welfare improvements, helps develop welfare indicators, and builds the scientific ecosystem we’ll need if we ever hope to understand smaller arthropods. That’s a much more stable basis for action than trying to set priorities via BOTECs.
Anyway, we’ll just have to agree to disagree, as we just keep running up against the same issues over and over!
Lastly, this article is good. The possibility the they’re right is one of the things that makes me inclined to see insects as the limit case.
Thanks for the clarifications, Bob! @Bentham’s Bulldog, you may be interested in the 2 comments above from Bob.
In this case, I feel like it would also be reasonable to argue that the evidential situation with respect to comparing the individual welfare per animal-year (not probability of sentience) of different species is so sparse that one should just focus on increasing the welfare of vertebrates. At least from my perspective, any comparison of the welfare (not probability of sentience) of shrimps with that of humans involves “guesses stacked on guesses”.
In addition, I see the lack of robust evidence for the sentience of soil springtails, mites, and nematodes as a case for further research on their sentience (although I would be surprised if it updated me towards thinking their expected individual welfare per animal-year is much lower than suggested by “number of neurons”^”exponent”). At some point, there was not robust evidence for the sentience of BSF larvae.
In any case, I assume the points about the robustness of evidence do not apply to soil ants and termites.
This makes sense to me. I ranked Arthropoda 1st in the Donation Election on that basis. At the same time, I suspect the optimal spending on research on soil animals is not 0. I got no results for “ants”, “termites”, “springtails”, “mites”, or “nematodes” on WAI’s grantees page.
I wonder whether there are some calculations one could do to compare the cost-effectiveness of building capacity for research on soil animals via doing this directly, or indirectly through research on farmed invertebrates.
Thanks for sharing! I will have a look.
Strongly agree about “the evidential situation with respect to comparing the individual welfare per animal-year”! I’ve always taken the numbers from the MWP much less seriously than others. I see that work as one part of a large picture, depending heavily on other arguments.
And thank you for voting for Arthropoda!