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 Grilođ¸
Karma: 11,895
What Will AI Do To Global InequalÂity?
Hello ion. Welcome to the EA Forum. On 1, I am open to bets against short timelines for transformative AI (TAI), or what they supposedly imply, up to 10 k$. On 2, the environmental and survival pressures are to a significant extent established by humans, who are the ones setting up the training of AIs?
Hi Jim. There is no particular reason. I have just been sharing content I find interesting. My views about AI risk have not changed much recently. I guess the risk of human extinction over the next 10 years is something like 10^-7.
How Fast Is Post-AGI Growth?
Hi Jienna. Thanks for the update.
Corrected days-of-suffering formula. The old formula multiplied indirect impacts (like feed fish and pre-slaughter mortality) by the primary animalâs lifespan. A wild-caught anchovy used as feed suffers for one day regardless of what eats them. The corrected formula treats indirect impacts separately.
Have Faunalytics and Bryant Research (@Chris Bryant đ¸ or @Richie?) considered excluding effects on wild animals (such as wild fishes used for feed)? I am very uncertain about whether expanding agricultural land increases or decreases the welfare/âsuffering/âhappiness of soil invertebrates, and I believe effects on these may be much larger than the effects on farmed animals and wild animals used for feed.
I also do not know whether decreasing the catch of wild fishes for feed increases or decreases animal welfare. Michael St. Jules concluded âdecreasing the catch of wild aquatic animals for feed is likely to decrease aquaculture and increase insect farming, but has unclear effects on brine shrimp nauplii and other live feedâ.
Hello Kearney and Melanie. Thanks for the update.
Weighting each grantâs modeled social return on investment [SROI] by its funding amount gives a mean of roughly 5,500x â though these are early, uncertain estimates.
How do you estimate the benefits of a career transition in CG$ (1 CG$ is the value of giving 1 $ to someone earning 50 k$/âyear)? Have you considered estimating this from the product between the marginal SROI of the organisation benefiting from the transition, and their willingness to pay for the transition?
The most common grants were to EA-affiliated career advising programs and high-impact fellowship programs in Sweden, France, Israel, Norway, Estonia, and Poland. These were explicitly evaluated by how many high-impact career transitions could be facilitated per unit of budget cost, and measured against our bar.
How did you evaluate the other organisations?
Based on our cost-effectiveness estimates for about 85% of the recommended grants, weighted by funding, every dollar invested is expected to create about $5,500 of social benefit.
Which % of funding is covered by the 85 % of grants for which you estimates the cost-effectiveness?
Of the 28 grants, 13 are cause-neutral, supporting high-impact transitions across our priority areas. The other 15 overlap strongly with one or two specific cause areas, with global health and development the heaviest concentration. We set out to focus on mid-career and senior professionals (much existing infrastructure skews toward students), and on LMIC talent pipelines, where many high-impact organizations operate but career infrastructure remains thin. The bigger surprise was about the ecosystem itself. We estimated a 25% chance weâd receive more than 40 applications; we got almost 500. We also came in assuming weâd fund far fewer proposals and give away considerably less money.
That is an impressive number of applications. I like that you shared a forecast you made for the number of applications despite it being far off.
That makes sense. I do not know what the ideal giving multiplier on total spending is, but I think the marginal giving multiplier should be close to 1 (assuming it accounts for all effects). Spending should increase if it is above 1 because this would mean spending 1 $ more would lead to more than 1 $ of benefits.
Hi Fin. Great post. I have also become persuaded by illusionism since I started learning about it in June 2026 (after reading this comment from @Michael St Jules đ¸). You may be interested in chapter 9, âLosing Your Religion: Ethics After Illusionismâ, of François Kammererâs book âHouse of Mirrors: The Illusion of Phenomenal Consciousnessâ.
Thanks for the helpful clarifications.
Thanks for the question and sorry for the slow reply.
No worries. I would say 5 days is a short time. I also do not think it makes sense for me to believe people should reply faster without knowing what they are prioritising instead, although I sometimes send reminders to ensure my messages were seen, or not forgotten.
So, we assumed that the interventions THL carried out each affected 6 to 15x as many hens than reported on average through speeding up the timelines to hens becoming cage-free.
Note the benefits of accelerating the end of the cage-free implementation by N years while not accelerating the start of it are 50 % as large as the benefits from accelerating both the start and end of the cage-free implementation by N years.
Then, since the marginal cost-effectiveness in 2026 is probably much less than the average cost-effectiveness from 2015-2024, we discounted the resulting THL self-reported estimate by 70%. This is a rough guess.
The marginal cost-effectiveness in 2026 is probably lower than the average cost-effectiveness in 2025. So ACEâs estimate for the average cost-effectiveness in 2025 should also be discounted (by less than 70 %)?
If you know of any good marginal cost-effectiveness estimates, or other cost-effectiveness estimates, please do link them below!
Here are estimates for the cost-effectiveness of Animaâs work targeting McDonalds in Norway from 2026 on to increase the welfare of broilers. They are based on guesses from Niklas Fjeldberg and Toby Schiønning.
Ethics withÂout senÂtience. FacÂing up to the probÂaÂble inÂsignifiÂcance of pheÂnomÂeÂnal consciousness
Hi Anthony.
Yeah I still disagree for this 4th example as well, for the same reasons as the 1st.
Thanks for confirming. I would be surprised if there is any counterexample you find persuading.
Hi Ben. That makes sense.
I think that is what Richard Chappell is referring to when he mentions âradical skepticsâ here. But I would be interested in a version of his post which defends the claim that EAs are reasonably justified in making the trade-offs we currently make (even if a radical skeptic would not be convinced of this defense).
Likewise.
Great points.
Sometimes, this is not an option when making a decision. E.g., when advising a friend who received two job offers in animal welfare, neither of which decrease uncertainty; or when recommending grants for a fund (where none of the grants decrease uncertainty).
Even among roles whose major focus is not decreasing one of your major uncertainties, there could be some which decrease them significantly more than others. Roles which allow moving more funds (via personal donations or grants) or people (via career advice and recruitment) to projects decreasing the major uncertainties. Roles at organisations helping considerably different species instead of similar species. For example, a researcher at Animal Charity Evaluators (ACE) has to prioritise between helping cows, pigs, chickens, fish, and invertebrates, at least implicitly (and explicitly in the context of ACEâs charity evaluations). In contrast, a researcher at the Fish Welfare Initiative (FWI) overwhelmingly focuses on fish.
I think moral weights and pain trade-offs might be so seriously underdetermined that I expect that most research I could do or support right now would not reduce uncertainty about them. I expect youâll have a counter in that, though, and I expect my views to evolve.
Some advocacy and research could build capacity for decreasing the major uncertainties even if they do not directly decrease them.
In comparison, assessing 2nd-order effects of certain interventions (such as economic effects, and land use effects) is tractable, although difficult.
I agree.
I kind of rule out âlong-horizon thinkingâ in my decision procedure on principle. Hence, if I expected pain tradeoff ratios to become a resolvable issue after 30 years of empirical research necessiating $50M of investment per year, I donât think Iâd want to seed it, because I think that the chance that this works is negligible (youâll likely object to this on decision-theoretic grounds, I assume).
I think the expected impact increases gradually with spending, and I would focus on the expected impact per $ instead of the increase in the probability of at least some level of success per $. Say spending 1.5 billion $ (= 50*10^6*30) over 30 years would make comparisons between pain intensities 100 % resolvable. I would expect cost-effectiveness to decreases with spending. So, for example, I would expect seed funding of 15 k$ to make comparisons between pain intensities much more than 0.001 % resolvable (= 15*10^3/â(1.5*10^9)).
This is more uncomfortable to admit (and Iâm particularly uncertain on that point), but if I were to focus almost entirely on research that decreases uncertainty, appearing as someone who recommends no object-level actions may be costly to me in a way that could reduce the change that my research is successful in decreasing uncertainty, because it could receive less support.
I agree fundraising for the research decreasing the major uncertainties will be easier if it informs funding decisions related to non-research interventions. At the same time, I expect some degree of specialisation to be useful. Some people could focus on the research decreasing the major uncertainties (for instance, Rethink Prioritiesâ (RPâs) worldview investigations team (WIT)), and others on funding decisions related to non-research interventions (like grantmakers).
Here are relevant results presented in the book What We Owe to the Future (WWOF).
Hi Jeff. I appreciate your transparency, and how you are thinking about this.
Thanks for the discussion which prompted me to look into it.
Hi Jim. I agree trying to come up with counterexamples is useful. Below are 3 potential counterexamples I have given. @Anthony DiGiovanni đ¸ does not consider them counterexamples (see Anthonyâs replies for details).
1st example, which Elliot already quoted in this thread.
Consider these 2 options for what I could do tomorrow:
Torturing my family, and friends, and then killing myself. I would never do this.
Donating 100 $ to the Shrimp Welfare Project (SWP), which I estimate would be as good as averting 6.39 k (= 639/â10*100) human-years of disabling pain.
My understanding is that you think it is âirreducibly indeterminateâ which of the above is better to increase expected impartial welfare, whereas I believe the 2nd option is clearly better.
2nd example.
given any 2 objects, I believe my best guess should be that the expected mass of one is smaller, equal, or larger than that of the other.
3rd example.
Hi Anthony. Do you think the expected welfare of 2 states of the world which only differ infinitesimally can be incomparable? I do not see how this could be possible. For example, it feels super counterintuitive to me that, given 2 identical states, moving an electron by 10^-100 m in one of the states would make their expected welfare incomparable. I guess one can get from any state of the universe to another in an astronomical number of infinitesimal steps, and I believe any 2 states which only differ infinitesimally are comparable. So I conclude any 2 states are comparable too, even if it is very hard to compare them, to the point that I do not know if electrically stunning shrimps increases or decreases welfare in expectation.
Here is a 4th example. Consider these 2 actions:
Killing the 100 people who are expected to decrease the most the uncertainty about how to compare the expected value of different actions. For example, Bob Fischer who has worked on decreasing uncertainty about comparing welfare across species.
Grating 10 M$ to the 100 people above (100 k$ per person, but the grant size could vary). The 10 M$ would otherwise be spent torturing people as much as possible.
I think we are justifyed in c-preferring the 2nd action. I believe Anthony disagrees
Hi Rocky and Tom. Great post. I share your concerns.
Our worry is not that welfare tech is a bad idea (some of it is likely excellent). Rather, the worry is that it has several traits that make it vulnerable to becoming a bubble. It has a clean theory of change, no obvious need for mass public persuasion, compatibility with short AI timelines, appeal to tech-oriented funders and builders, and the ability to be framed as a silver bullet: build the tool, deploy the tool, reduce suffering at scale.
I very much agree.
Some considered Shrimp Welfare Projectâs [SWPâs] stunning work as perhaps the most impactful program for farmed animals to date (even by many orders of magnitude). However, this has recently turned to disillusionment for many, as the June 2026 post âanimal welfare has an evidence problemâ outlined some of the key uncertainties around current stunning equipmentâs efficacy.
I estimated SWPâs stunning work is 383 and 139 times as cost-effective as broiler welfare and layer cage-free corporate campaigns (neglecting effects on non-target animals). However, I think the estimates depend way more on welfare comparisons across species than the probability of the stunning work benefiting and harming shrimps.
Say the cost-effectiveness of SWPâs stunning work is proportional to the difference between the probability of benefiting and harming shrimps, and that shrimps are either benefited or harmed. Suppose the probability of benefiting shrimps decreases from 75 % to 60 %, which means the probability of harming shrimps increases from 25 % to 40 %. The cost-effectiveness would become 40 % (= (0.6 â 0.4)/â(0.75 â 0.25)) as large.
My estimate for the cost-effectiveness of SWPâs stunning work is proportional to the value of 1 shrimp-QALY. I assumed this to be 9.34 % (= 0.031/â0.332) of the value of 1 chicken-QALY. Suppose the value of 1 anima-QALY is instead proportional to the individual number of neurons, which I consider plausible. I calculate shrimps have 0.0389 % as many neurons as chickens. So the cost-effectiveness would become 0.416 % (= 3.89*10^-4/â0.0934) as large.
I would be happy to review your estimates of the SROI of the grants for free if you think it may be useful.