I am a generalist quantitative researcher. I am open to volunteering and paid work. I welcome suggestions for posts. You can give me feedback here (anonymously or not).
Vasco Grilošø
Karma: 11,209
Got it. Consider these 2 pains:
A. Duration of 1 s, intensity of X, and probability of 10^-100.
B. Duration of 10^100 years, intensity of 99.99999999 % of X, and probability of 1.
Would you prefer averting A over B? In other words, would you prefer A) an infinitesimal chance of decreasing 1 s of very intense pain over B) certainly decreasing 10^100 years of an infinitesimaly less intense pain?
Thanks for the thoughts, Elif.
I agree annoying and excruciating pain have very different properties. However, it does not follow that an arbitrarily short time in excruciating pain is much worse than an arbitrarily long time in annoying pain? Liquid water and ice have different properties, but, for example, their mass and temperature can still be quantitatively compared. I do not think analogies with physics illustrate that some pain intensities cannot be quantitatively compared.
Would you prefer 10 years in annoying pain over a probability of 10^-100 of 0.1 s in excruciating pain? If so, what do you think about the questions I asked here?
Hi Elif. Thanks for the comment.
But one might simply reject these frameworks in favor of a person-affecting view, which I find far more intuitive.
Aggregation across existing individuals still matters in person-affecting views.
If no single observer in the universe experiences a ācatastrophe,ā can we truly say a catastrophe has occurred?
It depends on how catastrophe is defined. I think most people would consider a catastrophe all life on Earth dying painlessly, even though no one would experience anything in the process.
I donāt know about you but I would never take that gamble.
Would you take the gamble if it involved 1 min of excruciating pain? If not, would you want the painless death of people who have a probability of experiencing more than 1 min of excruciating pain in their real future higher than 1 in 1 trillion, 10^-12, to eliminate the risk of them experiencing excruciating pain? I think the probability of a random person experiencing more than 1 min of excruciating pain in their remaining life is much higher than 10^-12. The probability of dying in a road injury in high income countries in 2023 was 8.61*10^-5. So a probablity of 0.1 % of experiencing more than 1 min of excruciating pain in a road injury death would result in an annual probability of a random person in high income countries experiencing more than 1 min of excruciating pain of at least 8.61*10^-8 (= 8.61*10^-5*0.001), 86.1 k (= 8.61*10^-8/ā10^-12) times as large as 1 in 1 trillion. āAt leastā because there are other events besides road injury deaths which could lead to more than 1 min of excruciating pain.
Iām not sure what it means for a candidate to be āmore cost-effectiveā, but I assume it equates to ābetter at their jobā (assuming a fixed salary).
Yes. I meant more impactful per unit cost (accounting for the direct impact of the job, and its time and financial costs).
My impression, from spending a lot of time in EA orgs and working on several hiring rounds, is that the difference between ābest candidateā and āsecond-best candidateā is often well over 10% (though, to be fair, orgs may be wrong about which candidate is best).
The expected difference will tend to be smaller than the observed difference (the best candidates will tend to regress more towards the mean). I do not know how much this matters, and the extent to which organisations try to account for it, but I guess you are right that the most cost-effective candidate are often more than 10 % more cost-effective than the 2nd most cost-effective. Donating more 10 % of the gross salary to an organisation 10 to 100 times as cost-effective at the margin as A would still be 2 (= 0.1*10/ā0.5) to 20 (= 0.1*100/ā0.5) times as impactful as joining A for an alternative hire 50 % as impactful who would get the same salary.
Your model seems to assume that a donor will support the most cost-effective intervention in an area. But if less cost-effective interventions have a lot more jobs, doesnāt that imply they also have more funding, and thus that the average donor is more likely to support them?
Great point. I would just say ārandom fundingā instead of āaverage donorā. I see now that the vast majority of people considering joining A would not find the comparison I made just above relevant. If they thought there was another organisation 10 to 100 times as cost-effective at the margin as A, they would most likely only consider joining organisations significantly more cost-effective at the margin than A.
The implicit assumption Iām reading from your model is that someoneās impact within a job is worth exactly what they are paid.
Yes, I am making this assumption. I think it roughly applies to new jobs. My thinking is that funders should fund organisations until they are indifferent between funding them more or not. If a funder thought that an organisation spending 100 k$/āyear more on a new job would be worth 1 M$/āyear more to the funder, they would be leaving 900 k$/āyear (= 1*10^6 ā 100*10^3) of impact on the table, in the sense that giving 100 k$/āyear more to the organisation would be as impactful as the funder having 900 k$/āyear more to spend.
In my experience, EA orgs tend to describe themselves as much more limited by talent than funding.
I wonder whether this depends on the audience. Organisations have an incentive to highlight talent constraints in hiring efforts (to get more applicants), and funding constraints in fundraising (to get more funding).
I am not sure what organisations mean when they say they are limited by funding or talent. Organisations are always constrained by both funding and talent to some extent. Additional funding can be used to retain and acquire talent via higher salaries and greater spending on hiring, including on field-building efforts like fellowships. I believe it would be better for organisations to say how much they value the best candidates over the 2nd best candidates (for roles they are hiring for) in terms of additional funding instead of just saying they are funding or talent constrained.
I think someone who changes their career is more likely to start donating than vice-versa.
I agree. In addition, I think people who change to more impactful jobs will tend to donate to more cost-effective organisations.
If you think most people make far more impact by donating than working, and that itās much easier to convince people to donate than change careers, this factor wonāt change much; youād still prefer lots of new donors to a small number of new workers who also donate. But I still thought it was worth noting.
Makes sense. I do not know whether marginal funding should mostly go towards adocating for cost-effective donations, including via earning to give, or careers.
I appreciate the impetus to think about the question!
Likewise.
Here is for example an article on sauna that I have read:
https://āāpdf.sciencedirectassets.com/āā271341/āā1-s2.0-S0531556521X00095/āā1-s2.0-S05315565210029[ā¦]lY3QuY29t&ua=0a0b560605045c52&rr=9aad34daa873be38&cc=dk
For what I understand, is that there is no extremely hard evidence for benefits of sauna/āicebathing, but there is a lot of research that points in that direction.Someone I know sent me this. Here is my reply on 13 December 2025 covering the study, which I think is in the spirit of Sethās post.
Thanks for sharing, [name]! To clarify, I was mostly sceptical about the ice bathing (I looked into this a bit because my father does 10 ĀŗC bathing sometimes, and was claiming it has lots of benefits). I have now looked into the study above (āSauna use as a lifestyle practice to extend healthspanā). I think the evidence presented there is very weak. In terms of data from interventional studies, which involve comparing a group which receives a treatment (like more sauna) with a control group, and are the type of study which offers the strongest evidence, they mention the following (I searched the paragraphs with āintervā):
āFindings from a small intervention study in rodents demonstrated that local heat application during an immobilization period decreased muscle atrophy by 37% compared to a sham treatmentā. Very weak evidence because it is about rodents.
āA randomized controlled trial that examined the effects of sauna use in 24 patients with ischemic heart disease with chronic total coronary artery occlusion who had not responded to non-surgical procedures and had failed or were not candidates for percutaneous coronary intervention demonstrated that 15 waon sessions given over a 3-week period improved the patientsā vascular endothelial function as measured by flow-mediated dilation of the brachial artery. No significant improvements were observed in the control group that received standard medical care (Sobajima et al., 2013)ā. Very weak evidence because:
The sample size is tiny (24 people in the treatment group).
The treated people are very special people (so sauna being beneficial for them does not imply it has benefits for a random healthy person).
They did not measure mortality (I do not know the extent to which this correlates well with āflow-mediated dilation of the brachial arteryā).
āA small intervention study investigated the effects of repeat sauna use on endurance and other physiological effects in 6 male distance runners. The findings showed that one 30-minute sauna session twice a week for 3 weeks post-workout increased the time that it took for the study participants to run until exhaustion by 32% compared to their baselineā. Very weak evidence because:
The sample size really is super small (6 people)!
The treated people are not random people (they are distance runners).
They did not measure mortality (I do not know the extent to which this correlates with ātime that it took for the study participants to run until exhaustionā).
āAnother randomized controlled trial found that endurance training in a sauna suit led to improved performance and respiratory measures, including VO2max. The authors speculated that the improved performance time for the sauna suit group was due to improved VO2max and increased capacity for thermoregulation. For example, they noted that sweat rate during a heated 5 km time trial increased in the post-intervention group but not the control group (Van de Velde et al. 2017)ā.
I am confused. The summary of this stufy does not even mention āsaunaā.
In any case, I guess the same points about the study above apply. A āsauna suitā most likely implies a very small sample size, and they did not measure mortality.
āA small intervention study in humans found that daily heat treatments applied locally to muscle during 10 days of immobilization prevented the loss of mitochondrial function, increased HSP levels, and attenuated skeletal muscle atrophy by 37% compared to a sham treatment group (Hafen et al., 2019)ā. Very weak evidence because:
By small, they mean ā23 healthy volunteersā.
They did not study realistic conditions. āThis study investigated the effects of daily heat therapy on human skeletal muscle subjected to 10 days of immobilizationā. The question is whether sauna helps random healthy people longterm, not whether they temporarily benefit people who were spending 10 days without moving.
This is not about sauna. It is about local heating. ādaily 2-h heat treatment using pulsed shortwave diathermyā.
They did not measure mortality.
Hi James.
NPS of 63 (mean recommendation: 8.86/ā10).
In case anyone is wondering, NPS stands for net promoter score, and is equal āfraction of people giving a score of 9 or 10 (promoters)āāāfraction of people giving a score of 0 to 6 (detractors)ā.
Hi Ozy.
By looking at this picture, you can see that weāre pretty certain that humans are conscious and very certain that ELIZA is not conscious. Weāre uncertain about whether chickens and 2024 LLMs are conscious.
These conclusions only hold for a prior probability of consciousness of 1ā6. I think this prior is very arbitrary. So I believe the results for the (posterior) probability of consciousness are also very arbitrary.
We are pretty certain about how uncertain we are about whether 2024 LLMs are conscious (we think thereās about a 10% chance theyāre conscious). But weāre not only uncertain about whether chickens are conscious, weāre very uncertain about how uncertain we are about whether chickens are conscious.
The model underestimates the uncertainty of the distributions of the probability of consciousness. The weights of the stances are set to point estimates. However, Figure 7 of the report shows the 13 experts surveyed were very uncertain about the weights. I would set the weights of the stances to very wide distributions to represent the very high model uncertainty.
Hi Aaron. Nice post.
In the golden age of public giving, the EA community spent most of its time thinking about where to give. Now, itās more about finding the right job. Thatās appropriate; the average person can make a much bigger impact by working than giving.
Do you mean random people in the whole population can have a much greater impact through work than giving? I find this hard to believe. Benjamin Todd thinks āitās defensible to say that the best of all interventions in an area are about 10 times more effective than [as effective as] the mean, and perhaps as much as 100 timesā, which is in agreement with the cost-effectiveness of interventions following a heavy-tailed distribution. If so, jobs were uniformly distributed across interventions, and a person in a random job within an area were 10 % more cost-effective than the 2nd best candidate for their job, them donating 10 % more of their gross salary to the best interventions in the area could have 10 (= 0.1*10/ā0.1) to 100 (= 0.1*100/ā0.1) times as much impact through donations as through work. In reality, I assume there are more jobs in less cost-effective interventions, as the best interventions only account for a small fraction of the overall funding. Based on Benās numbers, if there are 10 times as many people in jobs as cost-effective as a random one as in the most-effective jobs, a person in a random job within an area who is 10 % more cost-effective than the 2nd best candidate for their job, and donates 10 % more of their gross salary to the best interventions in the area is 100 (= 10*10) to 1.00 k (= 100*10) times as impactful as a person in the same job not donating.
Hereās a list of examples.
Nitpick. The link links to your post instead of the appendix.
Hi titotal. A āspeck of dust in the eyeā is supposed to represent something which decreases welfare very little (considering all effects, including decreasing boredom), thus being very slightly bad, and worth avoiding (all else equal). So one can interpret premise 1 (āmild pain is badā) as āmildly decreasing welfare is badā. I believe the arguments in the post work for an arbitrarily small decrease in welfare. Do you agree?
InĀfinite Dust Specks Are Worse Than One Torture
I meant my future decisions would be the same in reality if I could not gather additional evidence regardless of whether the mass of the 2 identical objects was exactly the same or differed by 10^-6 kg.
Do you think annual human welfare per human-year has increased since 1900? Child mortality decreased 37.3 pp (= 0.41 ā 0.037) since then until 2023. If you agree annual human welfare per human-year has increased since 1900, are you confident that similar progress cannot be extented to non-humans? Would you have argued 200 years ago that we are all clueless about how to increase human welfare? I agree research can backfire. However, at least historically, doing research on the sentience of animals, and on how to increase their welfare has mostly been beneficial for the target animals?
For example, the articles make the claim that brains make no clear separation between hardware and software. Okay, that seems to be true. But so what? Why should I believe that a lack of hardware/āsoftware distinction is a necessary property for consciousness to arise?
Relatedly, I liked the article The Abstraction Fallacy: Why AI Can Simulate But Not Instantiate Consciousness by Alexander Lerchner, and the reply to it by Shelly Albaum.
Hi Bob and Geoffrey.
Yes, what we do is very unusual outside of academiaāand inside it too. Re: other groups that do global priorities research, the most prominent ones are GPI, PWI, and the cause prio teams at OP.
GPI refers to the Global Priorities Institute, which has meanwhile closed down, PWI to the Population Wellbeing Initiative, and OP to Open Philanthropy, which has meanwhile rebranded to Coefficient Giving.
Thanks for the post, Huw.
He provides some evidence that satisfaction doesnāt correlate with momentary wellbeing. However, the WHRās own data finds a highly statistically significant correlation between positive affect and life satisfaction on the national level (Tables 8ā10; FWIW itās likely that they donāt find the same for negative affect because linear models donāt find additional explanatory power in intercorrelated variables).
I think Yasha is particularly concerned about comparisons across countries.
But perhaps the biggest problem with the World Happiness Report is that metrics of self-reported life satisfaction donāt seem to correlate particularly well with other kinds of things we clearly care about when we talk about happiness. At a minimum, you would expect the happiest countries in the world to have some of the lowest incidences of adverse mental health outcomes. But it turns out that the residents of the same Scandinavian countries that the press dutifully celebrates for their supposed happiness are especially likely to take antidepressants or even to commit suicide. While Finland and Sweden consistently rank at the top of the happiness league table, for example, both countries have also persistently experienced some of the highest suicide rates in the European Union, ranking in the top five EU countries according to one recent statistic.
Southern European countries have the lowest age-standardised disease burden per capita from self-harm in the Europe, but not the highest life satisfaction.
There correlation between life satisfaction and happiness across countries is positive, but very weak.
Great to know.
The CEAs of the animal welfare interventions looked very thorough. I left some comments.
Thanks for the comment, Michael. I read the post and Sethās original essay, and listened to the episode of The 80,000 Hours Podcast with Seth. I would agree the title of the post is a bit of a misnomer. I think one may update towards a lower chance of digital systems being conscious as a result of Sethās arguments, but they are far from conclusive. I only know I am conscious right now (and I am very confident I was conscious moments ago). So I think a system which is more similar to me at a fundamental physical level should have a higher chance of being conscious. However, I have no idea about what this implies in terms of concrete probabilities of consciousness. As far as I can tell, the available evidence is compatible with frontier large language models (LLMs) having a probability of consciousness of 10^-6, but also 99.999 %.
Hi Daniel and titotal. Thanks for the discussion.
I only know I am conscious right now (and I am very confident I was conscious moments ago). So I think a system which is more similar to me at a fundamental physical level should have a higher chance of being conscious. I have no idea about what this implies in terms of concrete probabilities of consciousness. As far as I can tell, the available evidence is compatible with frontier large language models (LLMs) having a probability of consciousness of 10^-6, but also 99.999 %.
As a side note, I would take for granted that all animals and digital systems are sentient, and focus on assessing the distribution of the intensity of subjective experiences. I think asking about the probability of sentience of an animal or digital system shares some of the issues of asking about the probability that an object is hot. People have different concepts about what āhotā means, and they do not depend just on temperature (for example, the minimum temperature for hot wood is higher than the minimum temperature for hot metal because this transfers heat more efficiently). I understand sentience as having subjective experiences whose intensity is not exactly 0. However, I suspect some people understand it as having subjective experiences which are sufficiently intense. Different bars for this will lead to different probabilities. Asking about the distribution of the intensity of subjective experiences mitigates this. For example, one could ask about the probability of the mean intensity of what an LLM experienced writing a message exceeding the mean intensity of human experiences. It still seems super hard to get numbers for this, but what they refer to may be more concrete than a vague concept like sentience.
I do not see how philosophical zombies (p-zombies) could be physically possible. If they were just like humans at a fundamental physical level, they would in fact be humans. So they would be as conscious as humans, which I assume are conscious (because I am a conscious human right now, and other humans do not seem relevantly different).
Mal and Bob, what would you fund with 50 k$ of unrestricted funding under expectational total hedonistic utilitarianism? How about under your own moral views? Why? I understand you are open to funding research on soil animals, but I wonder whether you would prefer funding more research on farmed invertebrates.