I am open to work.
Vasco Grilošø
Nice points, Holly! However, I think they only apply to small disagreements about AI timelines. I liked Epoch After Hoursā podcast episode Is it 3 Years, or 3 Decades Away? Disagreements on AGI Timelines by Ege Erdil and Matthew Barnett (linkpost). Ege has much longer timelines than the ones you seem to endorse (see text I bolded below), and is well informed. He is the 1st author of the paper about Epoch AIās compute-centric model of AI automation which was announced on 21 March 2025.
Ege
Yeah, I mean, I guess one way to try to quantify this is when you expect, I donāt know, we often talk about big acceleration, economic growth. One way to quantify is when do you expect, maybe US GDP growth, maybe global GDP growth to be faster than 5% per year for a couple of years in a row. Maybe thatās one way to think about it. And then you can think about what is your median timeline until that happens. I think if you think about like that, I would maybe say more than 30 years or something. Maybe a bit less than 40 years by this point. So 35. Yeah. And Iām not sure, but I think you [Matthew Barnett] might say like 15 or 20 years.
Relatedly, the median expert in 2023 thought the median date of full automation to be 2073.
I remain open to betting up to 10 k$ against short AI timelines. I understand this does not work for people who think doom or utopia are certain soon after AGI, but I would say this is a super extreme view. It also reminds me of religious unbettable or unfalsiable views. Banks may offer loans with better conditions, but, as long as my bet is beneficial, one should take the bank loans until they are marginally neutral, and then also take my bet.
Great post, Sam! I strongly upvoted it.
I think PLF is too broad for one to conclude that advocating for or against is a mistake. I would say it depends on which particular PLF systems are are advocated for, and what would be the counterfactual. I believe targeted advocacy for tracking welfare indicators beyond the ones needed to minimise costs would be useful.
More fundamentally, extending factory-farming may eventually be beneficial for animals (from their own perspective). I estimated broiler welfare and cage-free campaigns make the lives of chickens almost positive.
Is Rethink Prioritiesā (RPās) animal welfare department supported by any unrestricted funds? I have wondered whether making a restricted donation to the animal welfare department could result in some unrestricted funds moving from that to other departments, even if the movement in funds is smaller than the size of the donation. Is there a way of supporting RPās wild animal welfare research without any unrestricted funds currently supporting it moving to other work (including research on farmed animals)?
You are welcome, Saulius! Yes, it does. All the cost-effectiveness estimates of animal welfare interventions involved in the post are proportional to the welfare range of the species of the helped animals, which I set to āRPās median welfare range[s]ā (presented here; 0.332 for chickens, and 0.031 for shrimp).
Thanks for sharing your thoughts, Chris.
Thanks for elaborating, @gergo! I am tagging you because I have just expanded this comment.
Another consideration is a tradeoff between impact and cost-effectiveness.
Nitpick. Maximising cost-effectiveness and impact is equivalent holding spending constant. However, I understand you mean the cost-effectiveness will tend to decrease as the spending increases.
The best ones might be more cost-effective than Bluedot, but I doubt anyone looked at this very rigorously (this analysis is great, but doesnāt include a Bluedot-like program).
I agree that analysis is great. I have just asked them whether they have considered estimating their cost-effectiveness in QARYs (quality-adjusted research years) per $, as done in that analysis. You may want to nudge them too.
This could mean that the marginal impact of money donated to them is smaller, but funders still might prefer this to having to spend time on evaluating 10 AIS groups due to time costs)
Great point. Funders should maximise ācost-effectivenessā = āimpactā/āācostā = āimpactā/ā(āfinancial costā + ātime costā) = āimpactā/āāfinancial costā/ā(1 + ātime costā/āāfinancial costā) = ācost-effectiveness neglecting the time costā/ā(1 + āratio between the time and financial costā). Smaller grants have a higher ratio between the time and financial cost, so their cost-effectiveness neglecting the time cost has to be higher to clear a given cost-effectiveness bar.
The volunteer-run initiatives are likely more cost-effective kind of by default (though see next point).
There could still be costs besides wages (such that the āfinancial costā above would not be 0), although the ātime costā of the grantmakers may well be the driver of the overall cost.
Mal said āAny project we do on screwworms will include the effects on the screwworms themselves as well as the effects on wild animalsā. I am glad they care about all animals.
Thanks for all your efforts, Habryka.
Jeff Sebo and Robert Long argue for extending moral consideration to AI systems if they have at least a 0.1 % chance of being conscious. By this criterion, moral consideration should also be extended to the 10^18 terrestrial arthropods, 10^20 marine arthropods, and 10^21 nematodes (numbers from Table S1 of Barn-On et al. (2018)), given Rethink Prioritiesā (RPās) estimates for the probability of sentience (see numbers at the bottom of the table below). Even for 10^3 sentient AI systems per human being, I would expect the vast majority of sentient beings to be arthropods and nematodes. There would be 8*10^12 (= 8*10^9*10^3) sentient AI sytems, 8.2*10^16 (= 0.082*10^18) and 8.2*10^18 (= 0.082*10^20) sentient terrestrial and marine arthropods in expectation, assuming a random arthropod is as likely to be sentient as a silkworm, and 6.8*10^19 (= 0.068*10^21) sentient nematodes in expectation. For most sentient beings to be digital in expectation, there would have to be way more than trillions of sentient digital minds.
I guess people are assuming digital sentience is a bigger deal than the naively suggested by the calculations above due to supposing the welfare range of sentient digital minds to be much wider than that of sentient arthropods or nematodes. This illustrates estimating the welfare range conditional on digital consciousness is important.
Great points, Abraham!
I also think that wild animal welfare just remains a problem for ~everyone, given that wild animal welfare impacts are downstream from most other interventions, so solving it should be a big priority. Insofar as people think that wild animal suffering is intractable because of uncertain impacts of your intervention on other wild animals, surely that would basically just apply to anything you do in the world that impacts wild animals (which is probably basically everything). If you buy the case for wild animals mattering morally, but think that downstream effects make it impossible to act on it, most charity seems to get stuck.
Agreed. Here are some calculations illustrating the effects of GiveWellās top charities on wild animals can easily be much larger than those on humans. I would say accounting for effects on farmed animals alone is enough to make it unclear whether extending human lives increases or decreases welfare.
The Insect Institute may be good too, but I am not familiar with their work.
I have now checked their publications. They seem to be trying to slow down the growth of the insect industry. I think this may be beneficial if farmed insects have negative lives, but harmful if they have positive lives. In any case, the direction of the effect also depends on how farmed insects replace other farmed animals. I believe the Arthropoda Foundation is more robustly beneficial. They focus on improving the conditions of insects, which is beneficial regardless of whether they have positive or negative lives, and how they replace other farmed animals.
Thanks for sharing, Ben! Lots of interesting resources.
I liked Epoch After Hoursā podcast episode Is it 3 Years, or 3 Decades Away? Disagreements on AGI Timelines by Ege Erdil and Matthew Barnett (linkpost). Ege has much longer timelines than the ones you seem to endorse (see text I bolded below), and is well informed. He is the 1st author of the paper about Epoch AIās compute-centric model of AI automation which was announced on 21 March 2025.
Ege
Yeah, I mean, I guess one way to try to quantify this is when you expect, I donāt know, we often talk about big acceleration, economic growth. One way to quantify is when do you expect, maybe US GDP growth, maybe global GDP growth to be faster than 5% per year for a couple of years in a row. Maybe thatās one way to think about it. And then you can think about what is your median timeline until that happens. I think if you think about like that, I would maybe say more than 30 years or something. Maybe a bit less than 40 years by this point. So 35. Yeah. And Iām not sure, but I think you [Matthew Barnett] might say like 15 or 20 years.
Relatedly, the median expert in 2023 thought the median date of full automation to be 2073.
I remain open to betting up to 10 k$ against short AI timelines. I understand this does not work for people who think doom or utopia are certain soon after AGI, but I would say this is a super extreme view. It also reminds me of religious unbettable or unfalsiable views. Banks may offer loans with better conditions, but, as long as my bet is beneficial, one should take the bank loans until they are marginally neutral, and then also take my bet.
I have a question about the campaigner role. Are any of the 4 stages before the interview (5th stage) eliminatory?
Timeline
1 April ā 4 May 2025 ā Apply ā Time for you to apply.
10ā19 May 2025 ā Join us on a video call to answer a few brief questions ā It will be scheduled at a time convenient for you and take 15-30 minutes.
24 May ā 1 June 2025 ā Do role-specific tasks and answer cultural fit questions ā We will send you some tasks to complete in a questionnaire format.
Reference check ā We will ask you for contact information for a few professional references, such as former employers, supervisors, to talk about your skills and traits.
7ā15 June 2025 ā Join an interview ā If you successfully pass the previous stages, we will invite you for an interview. You will be compensated for your time.
25 June ā 25 July 2025 ā Do a short work trial with us ā If you successfully pass the interview stage, we will invite you to join us for a two-day work trial. You will be compensated for this.
26ā30 July ā Learn whether you are accepted ā We will contact you to let you know about the final decision regarding your application.
August 2025 ā Get feedback from us ā If you are interested, we will do our best to provide feedback to help you understand what we rated positively in your application and what we didnāt. We hope this will help you in getting another impactful position. This stage is heavily dependent on the amount of participants due to constraints in our capacity.
Please note: the timeline may slightly change due to unforeseen circumstances. We will do our best to make sure it wonāt.
Although I was wondering how sure we are that a death caused by the screwworm is worse than the average death in nature for those animals.
Thanks for noting this, Keyvan! I also worry about that.
I am fine with neglecting indirect effects on other wild animals besides the infected animals and screwworms, but I think these are the most directly affected (the goal of the intervention is their eradication), so they should be considered. Do you know whether the increase in welfare of the no longer infected wild animals would be larger than the decrease in welfare of the eradicated screwworms assuming these have positive lives? If it takes 100 worm-years, like 100 worms for 1 year, to kill a host animal, and each lethal infection is worse than the counterfactual death by 0.5 host-years of fully healthy life, the welfare per worm-year would only have to be more than 0.5 % (= 0.5/ā100) of the welfare per host-year of fully healthy life for the intervention to be harmful[1]. This seems possible considering that Rethink Prioritiesā median welfare range of silkworms is 0.388 % (= 0.002/ā0.515) of that of pigs. I also think the worms may have positive lives because they basically live inside the food they eat. I suspect there is a natural tendency to neglect the effect on worms because they are disguting (at least to me[2]), but this is not a good reason to disregard their welfare.
Thanks for the post, Matthew! Strongly upvoted.
I think it is very unclear whether wild insects have positive or negative lives. So I would focus on understanding their experiences, and how to improve them, instead of decreasing or increasing the number of insects. In any case, I like your recommendation to donate to the Shrimp Welfare Project (SWP) and Wild Animal Initiative (WAI). The Insect Institute may be good too, but I am not familiar with their work.
Thanks for sharing! The job ads are super comprehensive.
Thanks for the great clarifications, Lauren! Strongly upvoted.
Another specific i found out yesterday, someone was able to pass something through their local gov that led to 400 million animals being spared that wasnāt even on the radar before they entered. It seems extremely unlikely that this kind of leverage and counterfactual would be the case for the best vs. next best candidate in an NGO.
Interesting example! I would be interested to know more, but I understand it may be sensible information to share publicly. I think one can help 400 M shrimp donating 26.7 k$ (= 400*10^6/ā(15*10^3)) to the Shrimp Welfare Project (SWP). So, if your example was representative of the impact of a career in policy inside the system, and the impact per animal helped in your example matched that of SWP (which I estimated to be 0.0426 DALYs averted), maximising donations could still be better. For a career of 40 years, one would only need to donate 668 $ (= 26.7*10^3/ā40) more to SWP per year relative to the career in policy inside the system.
Thanks for the comment, Sjlver!
many downward adjustments (and lack of upward adjustments)
My cost-effectiveness estimate is supposed to be unbiased in the sense of not being too low or high in expectation.
During Veganuary 2020, my wife and I made the decision to become vegan. We had been vegetarians before, and found out during Veganuary that a fully vegan lifestyle was easier than expected. Since then, one of our flatmates transitioned from omnivore to vegan. Another flatmate stayed omnivore but ate mostly vegan during the year that she lived with us. This is an extreme example, but it shows that the 31 emails can affect more than just one person, and for a duration longer than 6 months.
To be clear, I think one single email or video can turn someone from omnivoure to vegan. However, I believe that is super far from the expected effect.
Overall, there seems to be a clear trend in Germany toward more vegan products.
The supply per capita of poultry meat in Germany has not had a clear downwards trend, although it does seem like it has already peaked.
Likewise for the supply per capita of fish and other seafood in Germany.
However, this is very weak evidence of the impact of Veganuary. There are many factors which affect meat consumption in Germany besides Veganuary, and that may well be the country which Veganuary targets with the most positive trends. In the UK, the consumption per capita of poultry meat has been increasing, although that on fish and other seafood has recently been decreasing.
Oat milk shelves are larger than cow milk shelves in many retailers nowadays
Nitpick. Dairy accounts for a very small fraction of animal suffering. I think decreases in its consumption only matter to the extent they predict decreases in the consumption of eggs, poultry birds, fish, or other seafood.
Thanks for the comment, Sjlver!
many downward adjustments (and lack of upward adjustments)
My cost-effectiveness estimate is supposed to be unbiased in the sense of not being too low or high in expectation.
During Veganuary 2020, my wife and I made the decision to become vegan. We had been vegetarians before, and found out during Veganuary that a fully vegan lifestyle was easier than expected. Since then, one of our flatmates transitioned from omnivore to vegan. Another flatmate stayed omnivore but ate mostly vegan during the year that she lived with us. This is an extreme example, but it shows that the 31 emails can affect more than just one person, and for a duration longer than 6 months.
To be clear, I think one single email or video can turn someone from omnivoure to vegan. However, I believe that is super far from the expected effect.
Overall, there seems to be a clear trend in Germany toward more vegan products.
The supply per capita of poultry meat in Germany has not had a clear downwards trend, although it does seem like it has already peaked.
Likewise for the supply per capita of fish and other seafood in Germany.
However, this is very weak evidence of the impact of Veganuary. There are many factors which affect meat consumption in Germany besides Veganuary, and that may well be the country which Veganuary targets with the most positive trends. In the UK, the consumption per capita of poultry meat has been increasing, although that on fish and other seafood has recently been decreasing.
Oat milk shelves are larger than cow milk shelves in many retailers nowadays
Nitpick. Dairy accounts for a very small fraction of animal suffering. I think decreases in its consumption only matter to the extent they predict decreases in the consumption of eggs, poultry birds, fish, or other seafood.
Hi Fabienne,
Have you considered running a similar study for long-chain omega-3 supplementation @Paul_Christiano may be interested too. Witte et al. (2013) ran a randomised, double-bind, placebo-controlled trial involving 65 healthy people which found it āenhanced executive functions by 26%, whereas performance remained constant after placebo (paired t-test, t(31) = 3, P = 0.005; Fig. 3A)ā. Based on Fig. 3A, they are saying the mean of the following composite score for executive function among participants in the treatment group increased by 0.26. ā[z phonemic fluency + z semantic fluencyāz TMT [trail making test] (part Bāpart A)/āpartAāz STROOP [Stroop Color-Word test] (part 3 - (part 1 + part 2))/ā2]/ā4ā, where z refers to the z-score of a person on a test. āRegarding the composite score for memory, both groups showed a similar retest effect at follow-up, with no significant effect of group (ANOVARM, group X time: P = 0.6; time: F1,63 = 19.8, P < 0.001)ā. āFor sensorimotor speed, a global retest effect was noticed that did not differ between groups (composite score, ANOVARM, time: F1,63 = 14, P < 0.001)ā.