I think this is interesting, but I don’t think we should infer too much from this relationship. This plot basically matches those we produced examining the relationship between cortical neuron count and perceived moral value of different animals (replicating SlateStarCodex and another’s surveys). As you can see below, we found extremely strong correlations. But we also found similarly strong correlations using EQ or total brain size rather than cortical neuron count, or using a crude 0-100 measure of how people ‘care’ about the animal in place of tradeoff ratios. Notably, when we replace the moral value measure with a simple ordinal ranking of the animals by neuron count (as in the second plot below), we find even stronger relationships.
My impression is therefore that the strong correlations more reflect the fact that we have a small number of datapoints with animals differing dramatically on a wide variety of predictor (or, in principle, outcome) variables which are all highly correlated, rather than indicating that neuron counts are distinctively predictive of any outcomes of interest. See Andrew Gelman’s similar discussion of our study.
I think to actually disentangle these we would need a larger sample of animals who diverge on the key dimensions (e.g., birds with high neuron density but small brains, or animals with higher neuron count but lower perceived similarity to humans).
My impression is therefore that the strong correlations more reflect the fact that we have a small number of datapoints with animals differing dramatically on a wide variety of predictor (or, in principle, outcome) variables which are all highly correlated, rather than indicating that neuron counts are distinctively predictive of any outcomes of interest. See Andrew Gelman’s similar discussion of our study.
I agree. Moreover, in allometry, “the study of the relationship of body size to shape,[1]anatomy, physiology and behaviour”, “The relationship between the two measured quantities is often expressed as a power law equation (allometric equation)”. So the logarithm of the individual number of neurons explaining well the logarithm of the welfare ranges means this will also be well explained by many other properties. If the welfare range is roughly proportional to “individual number of neurons”^”exponent 1“, and the individual number of neurons is roughly proportional to “property (e.g. individual brain mass)”^”exponent 2”, the welfare range will be roughly proportional to “property”^(“exponent 1“*”exponent 2”). This means the logarithm of the welfare range will be well explained by the logarithm of “property”. Relatedly, here is an illustration of why I think individual welfare per fully-healthy-animal-year could be proportional to “metabolic energy consumption per unit time at rest”^”exponent”.
Given the above, I do not think it matters much whether one estimates welfare per unit time based on the individual number of neurons, or another property which is a power law of it. I believe it matters much more than results are presented for many exponents of the power law determining the welfare per unit time. I did this in the post where I estimated the total welfare of animal populations assuming individual welfare per fully-healthy-animal-year is proportional to “individual number of neurons”^”exponent”, where I analysed exponents ranging from 0 to 2.
Relatedly, here is an illustration of why I think individual welfare per fully-healthy-animal-year could be proportional to “metabolic energy consumption per unit time at rest”^”exponent”.
I have now estimated the total welfare of animal populations, trees, and bacteria and archaea based on the assumption above. I had recommended research informing how to increase the welfare of soil animals, but I am now more pessimistic about this. I currently think it is better to focus on decreasing the uncertainty about how the welfare per unit time of different organisms and digital systems compares with that of humans.
I think this is interesting, but I don’t think we should infer too much from this relationship. This plot basically matches those we produced examining the relationship between cortical neuron count and perceived moral value of different animals (replicating SlateStarCodex and another’s surveys). As you can see below, we found extremely strong correlations. But we also found similarly strong correlations using EQ or total brain size rather than cortical neuron count, or using a crude 0-100 measure of how people ‘care’ about the animal in place of tradeoff ratios. Notably, when we replace the moral value measure with a simple ordinal ranking of the animals by neuron count (as in the second plot below), we find even stronger relationships.
My impression is therefore that the strong correlations more reflect the fact that we have a small number of datapoints with animals differing dramatically on a wide variety of predictor (or, in principle, outcome) variables which are all highly correlated, rather than indicating that neuron counts are distinctively predictive of any outcomes of interest. See Andrew Gelman’s similar discussion of our study.
I think to actually disentangle these we would need a larger sample of animals who diverge on the key dimensions (e.g., birds with high neuron density but small brains, or animals with higher neuron count but lower perceived similarity to humans).
Thanks for the comment, David!
I agree. Moreover, in allometry, “the study of the relationship of body size to shape,[1] anatomy, physiology and behaviour”, “The relationship between the two measured quantities is often expressed as a power law equation (allometric equation)”. So the logarithm of the individual number of neurons explaining well the logarithm of the welfare ranges means this will also be well explained by many other properties. If the welfare range is roughly proportional to “individual number of neurons”^”exponent 1“, and the individual number of neurons is roughly proportional to “property (e.g. individual brain mass)”^”exponent 2”, the welfare range will be roughly proportional to “property”^(“exponent 1“*”exponent 2”). This means the logarithm of the welfare range will be well explained by the logarithm of “property”. Relatedly, here is an illustration of why I think individual welfare per fully-healthy-animal-year could be proportional to “metabolic energy consumption per unit time at rest”^”exponent”.
Given the above, I do not think it matters much whether one estimates welfare per unit time based on the individual number of neurons, or another property which is a power law of it. I believe it matters much more than results are presented for many exponents of the power law determining the welfare per unit time. I did this in the post where I estimated the total welfare of animal populations assuming individual welfare per fully-healthy-animal-year is proportional to “individual number of neurons”^”exponent”, where I analysed exponents ranging from 0 to 2.
I have now estimated the total welfare of animal populations, trees, and bacteria and archaea based on the assumption above. I had recommended research informing how to increase the welfare of soil animals, but I am now more pessimistic about this. I currently think it is better to focus on decreasing the uncertainty about how the welfare per unit time of different organisms and digital systems compares with that of humans.