Associate researcher in animal welfare at Rethink Priorities. Also interested in reducing s-risks.

Background mostly in pure math, computer science and deep learning, but also some in statistics/econometrics and agricultural economics.

Suffering-focused, anti-speciesist, prioritarian, consequentialist. Also, I like math and ethics.

My shortform.

Wiki Contributions


Donating money, buying happiness: new meta-analyses comparing the cost-effectiveness of cash transfers and psychotherapy in terms of subjective well-being

Firstly, the recipient is plausibly not the only person impacted by a cash transfer. They can share it with their partner, children, and even friends or neighbours. Such sharing should benefit non-recipients' well-being. However, it’s also possible that any benefit that non-recipients receive could be offset by envy of their neighbour’s good fortune. There appears to be no evidence of significant negative within-village spillover effects, but there is some evidence for positive within-household and across-village spillover effects. We have not included these spillover effects in our main analysis because of the large uncertainty about the relative magnitude of spillovers across interventions and the slim evidence available to estimate the household spillover effects.


I may be misremembering, but doesn't GiveDirectly give to whole villages at a time, anyway, making negative spillover very unlikely? If that's the case, it seems like all of the spillover effects should be positive (in expectation).

Do you have any thoughts on how the spillover effects of these interventions might compare, and is there any interest in looking further into this? Mental health interventions may also improve productivity (and so increase income), and people's mental health can affect others (especially family, and parents' mental health on children in particular) in important ways. On the other hand, people build wealth (and other resources, including human capital) within their communities, and cash transfers/deworming could facilitate this, but this may happen over longer time scales.

I would guess the effects on SWB through increased income for the direct beneficiaries of StrongMinds are already included in the measurements of effects on SWB, assuming the research participants were similar demographically (including in income, importantly) as the beneficiaries of StrongMinds.

EDIT:  Saw this in your post:

That being said, even if we take the upper range of GiveDirectly’s total effect on the household of the recipient (8 SDs), psychotherapy is still around twice as cost-effective.

Donating money, buying happiness: new meta-analyses comparing the cost-effectiveness of cash transfers and psychotherapy in terms of subjective well-being

How would the possibility of scale norming with life satisfaction scores (using the scales differently across people or over time, in possibly predictable ways) affect these results? There's a recent paper on this, and also an attempt to correct for this here (video here). (I haven't read any of these myself; just the abstracts.)

Seeking feedback on new EA-aligned economics paper

I'm thinking in practice, it might just be better to explicitly consider different distributions, and do a sensitivity analysis for the expected value. You could maximize the minimum expected value over the alternative distributions (although maybe there are better alternatives?). This is especially helpful if there are specific parameters you are very concerned about and you can be honest with yourself about what you think a reasonable person could believe about their values, e.g. you can justify ranges for them.

Maybe it's good to do both, though, since considering other specific distributions could capture most of your known potential biases in cases where you suspect it could be large (and you don't think your risk of bias is as high in other ways than the ones covered), while the approach you describe can capture further unknown potential biases.

Cross-validation could help set   when your data follows relatively predictable trends (and is close to random otherwise), but it could be a problem for issues where there's little precedent, like transformative AI/AGI.

Seeking feedback on new EA-aligned economics paper

In the section on robustness in the second paper, does the constant parameter for the degree of bias, ψ, have a natural interpretation, and is there a good way to set its value?

Suffering-Focused Ethics (SFE) FAQ

I didn't vote on your comment, but I think you made some pretty important claims (critical of SFE-related work) with little explanation or substantiation:

The core writings I have read (e.g. much of Gloor & Mannino's or Vinding's stuff) tend to make normative claims but mostly support them using interpretations of reality that do not at all match mine.  I would be very happy if we found a way to avoid confusing personal best guesses with metaphysical truth. 

What interpretations are you referring to? When are personal best guesses and metaphysical truth confused?

very few to no decision-relevant cases of divergence between "practically SFE" people and others

Do you mean between "practically SFE" people and people who are neither "practically SFE" nor SFE?

Also, as a result of this deconfusion, I would expect there to  be very few to no decision-relevant cases of divergence between "practically SFE" people and others, if all of them subscribe to some form of longtermism or suspect that there's other life in the universe.

What do you mean? People working specifically to prevent suffering could be called "practically SFE" using the definition here. This includes people working in animal welfare pretty generally, and many of these do not hold principled SFE views. I think there are at least a few people working on s-risks who don't hold principled SFE views, e.g. some people working at or collaborating with the Center on Long-Term Risk (s-risk-focused AI safety) or Sentience Institute (s-risk-focused moral circle expansion; I think Jacy is a classical utilitarian).

Why is suspecting that there's other life in the universe relevant? And do you mean the accessible/observable universe?

(I've edited this comment a bunch for wording and clarity.)

Suffering-Focused Ethics (SFE) FAQ

I think prioritarianism and sufficientarianism are particularly likely to prioritize suffering, though, and being able to talk about this is useful, but maybe we should just say they are more suffering-focused than classical utilitarianism, not that they are suffering-focused or practical SFE.

A practical guide to long-term planning – and suggestions for longtermism

That being said, it's worth keeping in mind that EA has multiple such funds and could continue to start more, so it's far more likely that we'll still have at least one around for much longer.

A practical guide to long-term planning – and suggestions for longtermism

A discount rate of 0.7% would suggest a fund could be passed forward about 300 years (details in Appendix). 

This is an interesting point. Naively compounding using expected market returns-discount rates would tell you the fund will grow forever, but what we really should expect is that the fund will eventually fail, and in the unlikely event that it's still running properly past 300 years (based on your discount rate), it'll be massive, and we might not even be able to use most of it very usefully, with decreasing marginal altruistic returns to resources.

Why does (any particular) AI safety work reduce s-risks more than it increases them?

Perhaps the more substantive disagreement is what fraction of the work is in which category. I see most but not all ongoing technical work as being in the first category, and I think you see almost all ongoing technical work as being in the second category. (I think you agreed that "publishing an analysis about what happens if a cosmic ray flips a bit" goes in the first category.)

Ya, I think this is the crux. Also, considerations like the cosmic ray flips a bit tend to force a lot of things into the second category when they otherwise wouldn't have been, although I'm not specifically worried about cosmic ray bit flips, since they seems sufficiently unlikely and easy to avoid.

(Luke says "AI-related" but my impression is that he mostly works on AGI governance not technical, and the link is definitely about governance not technical. I would not be at all surprised if proposed governance-related projects were much more heavily weighted towards the second category, and am only saying that technical safety research is mostly first-category.)


The "cluelessness" intuition gets its force from having a strong and compelling upside story weighed against a strong and compelling downside story, I think.

This is actually what I'm thinking is happening, though (not like the firefighter example), but we aren't really talking much about the specifics. There might indeed be specific cases where I agree that we shouldn't be clueless if we worked through them, but I think there are important potential tradeoffs between incidental and agential s-risks, between s-risks and other existential risks, even between the same kinds of s-risks, etc., and there is a ton of uncertainty in the expected harm from these risks, so much that it's inappropriate to use a single distribution (without sensitivity analysis to "reasonable" distributions, and with this sensitivity analysis, things look ambiguous), similar to this example, and we're talking about "sweetening" one side or the other i, but that's totally swamped by our uncertainty.

If the first-order effect of a project is "directly mitigating an important known s-risk", and the second-order effects of the same project are "I dunno, it's a complicated world, anything could happen", then I say we should absolutely do that project.

What I have in mind is more symmetric in upsides and downsides (or at least, I'm interested in hearing why people think it isn't in practice), and I don't really distinguish between effects by order*. My post points out potential reasons that I actually think could dominate. The standard I'm aiming for is "Could a reasonable person disagree?", and I default to believing a reasonable person could disagree when I point out such tradeoffs until we actually carefully work through them in detail and it turns out it's pretty unreasonable to disagree.

*Although thinking more about it now, I suppose longer chains are more fragile and likely to have unaccounted for effects going in the opposite direction, so maybe we ought to give them less weight, and maybe this solves the issue if we did this formally? I think ignoring higher-order effects is formally irrational using vNM rationality or stochastic dominance, although maybe fine in practice, if what we're actually doing is just an approximation of giving them far less weight with a skeptical prior and then they actually just get dominated completely by more direct effects.

Why might one value animals far less than humans?

With a fair amount of credence on the view that the value is not almost all destroyed, the expected value of big minds is enormously greater than that of small minds.


I think you may need to have pretty overwhelming credence in such a views, though. EDIT: Or give enough weight to sufficiently superlinear views.

From this Vox article, we have about 1.5 to 2.5 million mites on our bodies. If we straightforwardly consider the oversimplified view that moral weight scales proportionally with the square root of total neuron count in an animal, then these mites (assuming at least 300 neurons each, as many as C elegans) would have ~100x as much moral weight as the human they're on. Assigning just a 1% credence to such a view, assuming we're fixing the moral weight of humans and taking expected values puts the mites on our own bodies ahead of us (although FWIW, I don't think this is the right way to treat this kind of moral uncertainty; there's no compelling reason for humans to be the reference point and I think doing it this way actually favours nonhumans).

(86 billion neurons in the human brain)^0.5 = ~300,000.

(300^0.5) *2 million = ~35 million.

The square root may seem kind of extreme starting from your thought experiment, but I think there are plausible ways moral weight scales much more slowly than linearly in brain size in practice, and considering them together, something similar to the square root doesn't seem too improbable as an approximation:

  1. There isn't any reasonably decisive argument for why moral weight should scale linearly with network depth when we compare different animals (your thought experiments don't apply; I think network depth is more relevant insofar as we want to ask whether the network does a certain thing at all, but making honey bee networks deeper without changing their functions doesn't obviously contribute to increasing moral weight). Since brains are 3 dimensional, this can get us to  as a first approximation.
  2. For a larger brain to act the same way on the same inputs, it needs to be less than proportionally sensitive or responsive or both. This is a reason to expect smaller brains to be disproportionately sensitive. This can also result in a power of neuron count lower than 1; see here.
  3. There may disproportionately more redundancy in larger brains that doesn't contribute much to moral weight, including more inhibitory neurons and lower average firing rates (this is related to 2, so we don't want to double count this argument).
  4. There may be disproportionately more going on in larger brains that isn't relevant to moral weight.

I think we can actually check claims 2-4 in practice.

Also note that, based on this table:

  1. cats and house mice have ~13,000-14,000 synapses per neuron
  2. brown rats have ~2,500
  3. humans have ~1,750
  4. honey bees have ~1,000
  5. sea squirts and C elegans have ~20-40.

I'm not sure how consistently they're counting in whole brain vs central nervous system vs whole body.

Load More