Latest comments on the EA Forum

Some AI research projects that (afaik) haven't had much work done on them and would be pretty interesting:

• If the US were to co-build secure data centres in allied countries, would that be geopolitically stabilising or destabilising?
• What AI safety research agendas could be massively sped up by AI agents? What properties do they have (e.g. easily checkable, engineering > conceptual ...)?
• What will the first non-AI R&D uses of powerful and general AI systems be?
• Are there ways to leverage cheap (e.g. 100x lower than present-day cost) intelligence or manual labour to massively increase the US's electricity supply?
• What kinds of regulation might make it easier to navigate an intelligence explosion (e.g. establishing quick pathways to implement policy informed by AI experts, or establishing zones where compute facilities can be quickly built without navigating a load of red tape)? 

Deena's post only mentioned "of at least one large RCT underway, with results expected in a few years" without further reference, but on cursory googling it might be the CRADLE trial?

The Cement-based flooRs AnD chiLd hEalth trial is an individually randomised trial in Sirajganj and Tangail districts, Bangladesh. Households with a pregnant woman, a soil floor, walls that are not made of mud and no plan to relocate for 3 years will be eligible. We will randomise 800 households to intervention or control (1:1) within geographical blocks of 10 households to account for strong geographical clustering of enteric infection. Laboratory staff and data analysts will be blinded; participants will be unblinded. We will instal concrete floors when the birth cohort is in utero and measure outcomes at child ages 3, 6, 12, 18 and 24 months. 

The primary outcome is prevalence of any STH infection (Ascaris lumbricoides, Necator americanus or Trichuris trichiura) detected by quantitative PCR at 6, 12, 18 or 24 months follow-up in the birth cohort. Secondary outcomes include household floor and child hand contamination with Escherichia coli, extended-spectrum beta-lactamase producing E. coli and STH DNA; child diarrhoea, growth and cognitive development; and maternal stress and depression.

We will report findings on ClinicalTrials.gov, in peer-reviewed publications and in stakeholder workshops in Bangladesh.

While GiveWell doesn't seem to have looked into this specifically, this 2015 review of GiveDirectly mentioned that lack of cement floors was in one of GiveDirectly's two sets of eligibility criteria for its standard campaigns:

Thatched roofs: To date, GiveDirectly has used housing materials to select recipients in all of its standard campaigns, enrolling households who live in a house made of organic materials (thatched roof, mud walls, and a mud floor) and excluding households with iron roofs, cement walls, or cement floors.170 In GiveDirectly's campaigns in Kenya, about 35-45% of households have been eligible based on these criteria, while in Uganda about 80% of households have been found to be eligible.171 ... 

Happier Lives Institute's 2021 annual review did mention cement flooring among the "micro-interventions" they wanted to look into (alongside deworming, cataract surgery, digital mental health interventions, etc), but I haven't seen anything by them since on this, so I assume it didn't pass their internal review for further analysis. 

How can we unhypocritically expect AI superintelligence to respect "inferior" sentient beings like us when we do no such thing for other species? 

How can we expect AI to have "better" (more consistent, universal, compassionate, unbiased, etc) values than us and also always only do what we want it to do? 

What if extremely powerful, extremely corrigible AI falls into the hands of a racist? A sexist? A speciesist? 

Some things to think about if this post doesn't click for you

I think this "Human Alignment" your talking about is very important and neglected. You don't hear a many people call for an ethical transformation as a necessary adaptive step to the AGI era...  

No deadline? Can I register the day before? Or do you expect to potentially reach full capacity at some point before that?

No deadline? Can I register the day before? Or do you expect to potentially reach full capacity at some point before that?

Thanks for writing this up! Have you considered condensing this into a two-page policy brief? I am sure this could also be useful as a template for other countries. Feel free to dm me. 

Hi Vasco,

Thanks so much for your question and interest in our work!

Is there a way of supporting RP's wild animal welfare research without any unrestricted funds currently supporting it moving to other work (including research on farmed animals)?

Yes, absolutely. Since we do not currently have any unrestricted funds allocated to wild animal welfare, a restricted donation to this area would not cause funding shifts between departments, or animal welfare sub-causes. Instead, it would directly increase our capacity for wild animal welfare work. In fact, wild animal welfare is the least funded area of our animal welfare portfolio, despite its importance and potential for impact.

Our animal welfare work is primarily funded through restricted donations for specific projects or sub-causes, with most directed toward non-invertebrate and non-wild animal priorities. Only ~11% of RP's overall funding is unrestricted, and based on our current plans, donating to the Animal Welfare Department would not result in unrestricted funds being redirected elsewhere.

We take donor preferences very seriously. For larger donations, we’re happy to explicitly increase the budget for a department, sub-cause, or project by the exact amount of your contribution, eliminating any potential fungibility concerns entirely. For small donations (relative to project costs), there may be practical limitations if a project requires significantly more funding to proceed, but we’ll inform you if this is the case and will always work to honor donor preferences.

Please let me know if you have any other questions!

Interesting, I got the opposite impression from their about page ("4,000+ hospitals and health centers served, 51% fewer deaths from postpartum hemorrhaging in hospitals Zipline serves, 96% of providers report increased access to vaccinations in their area" which I assume means they're already targeting those hard-to-access areas), but of course they'd want to paint themselves in a good light and I'd be inclined to trust your in the field experience far more (plus general skepticism just being a sensible starting point).

Actually your point about a cheap bike being able to carry a lot more stuff makes obvious sense, and so me wonder how Zipline's modelling study in Ghana can claim that their cost per incremental fully immunised child was cheaper than "monthly immunization by mobile teams" which I assume includes dirt bikes. 

Is Rethink Priorities' (RP's) 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 RP's wild animal welfare research without any unrestricted funds currently supporting it moving to other work (including research on farmed animals)?

This is an interesting piece, Sam, thanks for writing it.

These are my almost entirely uninformed thoughts: based on a tiny bit of background knowledge of PLF and general observations of the animal movement.

It seems quite likely to me that PLF is coming to animal ag whether we like it or not. If this is the case, the important question isn't so much "should we promote PLF or do some other campaign?" but rather "how should we respond to PLF?"

At the end of your piece, you say "our role (if any) must be strictly defensive and containment-focused" - I can get behind most of this sentence, apart from the "(if any)". Surely, for many of the reasons you set out earlier in the article, to not engage with PLF at all would be borderline neglence on the part of our movement: the risks that this transforms the industry and locks in practices are just so high that we can't afford to ignore this development.

So then the question is what should we do about it? I think I would favour a broader approach than you suggest which places multiple bets.

It seems plausible to me that some organisations should be trying to contain/prevent this new technology. I think such a campaign could bring animal advocates, smaller farmers and the general public together in quite a powerful way that would be able to get decent media/political traction at least in NA and Europe.

However, it still seems like there is a big risk that such a campaign would overall fail. This might be because, for example, the lure of big profits from allowing such practices outweighs any political pushback, or even simply because other countries (e.g. China) do adopt these practices and are then simply able to provide imports much more cheaply, effectively 'offshoring' the cruelty by displacing domestic production.

For this reason, I would favour some organisations also taking a much more 'good cop' role, working behind the scenes with PLF developers and regulators in a much more cooperative way in addition to campaigns opposing PLF. If PLF does become widespread, there are potentially very large wellbeing gains to be had by influencing the development of the technology at an early stage and maybe even locking in some welfare considerations.

I don't think it is completely naive to think this is possible: For example:

1. the food industry doesn't just compete on price alone, so including welfare could be a product differentiator for PLF creators and/or users;
2. the combination of some outside public pressure might convince PLF creators to introduce some welfare considerations voluntarily in an effort to head off the risk of more onerous regulations if they were seen as making no welfare concessions.

So, while I'd agree that we should be pretty suspicious about PLF and not welcome it with open arms. I think that we could be making a serious strategic error by either ignoring it (this seems the worst possible option) or providing only implacable opposition across the board.

It's in the linked-to Google doc, but there is a proper podcast feed for this now:
https://podcasts.apple.com/us/podcast/rob-wiblins-top-recommended-econtalk-episodes-v0-2-feb-2020/id1538606917

This is great!

One minor clarification (that I guess you are taking as "given" for this audience, but doesn't hurt to make explicit) is that the kind of "Within-Cause Prioritization" found within EA is very different from that found elsewhere, insofar as it is still done in service of the ultimate goal of "cross-cause prioritization". This jumped out at me when reading the following sentence:

A quick reading of EA history suggests that when the movement was born, it focused primarily on identifying the most cost-effective interventions within pre-existing cause-specific areas (e.g. the early work of GiveWell and Giving What We Can)

I think an important part of the story here is that early GiveWell (et al.) found that a lot of "standard" charitable cause areas (e.g. education) didn't look to be very promising given the available evidence. So they actually started with a kind of "cause prioritization", and simply very quickly settled on global poverty as the most promising area. This was maybe too quick, as later expansions into animal welfare and x-risk suggest. But it's still very different from the standard (non-EA) attitude of "different cause areas are incommensurable; just try to find the best charity within whatever area you happen to be personally passionate about, and don't care about how it would compare to competing cause areas."

That said, I agree with your general lesson that both broad cause prioritization and specific cross-cause prioritization plausibly still warrant more attention than they're currently getting!

This article is extremely well written and I really appreciated how well he supported his positions with facts.

However, this article seems to suggest that he doesn't quite understand the argument for making alignment the priority. This is understandable as it's rarely articulated clearly. The core limitation of differential tech development/d/acc/coceleration is that these kinds of imperfect defenses only buy time (this judgment can be justified with the articles he provides in his article). An aligned ASI, if it were possible, would be capable of a degree of perfection beyond that of human institutions. This would give us a stable long-term solution. Plans that involve less powerful AIs or a more limited degree of alignment mostly do not

Thanks for highlighting the relative lack of attention paid to cause prioritization and cross-cause prioritization. 

I have also written about how important it is to enable EAs to become familiar with existing cause prioritization findings. It's not just about how much research is done but also that EAs can take it into account and act on it.

Thanks for all your efforts. I think donating to ALLFED saves human lives roughly as cost-effectively as GiveWell's top charities^[1], so I would say it is a good opportunity for people supporting global health and development interventions^[2].

1. ^{^}

   I estimated policy advocacy to increase resilience to global catastrophic food shocks is 4.08 times as cost-effective as GiveWell's top charities.

2. ^{^}

   Although I believe the best animal welfare interventions are way more cost-effective, and I do not know whether saving human lives is beneficial or harmful accounting for effects on animals.

I love this post, I think this is a fundamental issue for intent-alignment. I don't think value-alignment or CEV are any better though, mostly because they seem irreversible to me, and I don't trust the wisdom of those implementing them (no person is up to that task).

I agree it would be good to I implement these recommendations, although I also think they might prove insufficient. As you say, this could be a reason to pause that might be easier to grasp by the public than misalignment. (I think currently, the reason some do not support a pause is perceived lack of capabilities though, not (mostly) perceived lack of misalignment).

I'm also worried about a coup, but I'm perhaps even more worried about the fate of everyone not represented by those who will have control over the intent-aligned takeover-level AI (IATLAI). If IATLAI is controlled by e.g. a tech CEO, this includes almost everyone. If controlled by government, even if there is no coup, this includes everyone outside that country. Since control over the world of IATLAI could be complete (way more intrusive than today) and permanent (for >billions of years), I think there's a serious risk that everyone outside the IATLAI country does not make it eventually. As a data point, we can see how much empathy we currently have for citizens from starving or war-torn countries. It should therefore be in the interest of everyone who is on the menu, rather than at the table, to prevent IATLAI from happening, if capabilities awareness would be present. This means at least the world minus the leading AI country.

The only IATLAI control that may be acceptable to me, could be UN-controlled. I'm quite surprised that every startup is now developing AGI, but not the UN. Perhaps they should.

I think there's an intersection between the PauseAI kind of stuff, and a great-powers reconciliation movement.

Most of my scenario-forecast likelihood-mass, where the scenarios feature near-term mass-death situations, exist in this intersection between great-power cold-wars, proxy-wars in the global-south, AI brinkmanship, and asymmetrical biowarfare.

Maybe combining PauseAI with a 🇺🇸/🇨🇳 reconciliation and collaboration movement, would be a more credible orientation.

Moral alignment of AI's is great.  But we need moral alignment of all intelligences.  Humans, literal whales, and AIs.  Confusion, trauma, misalignment, and/or extinction of some intelligences against others negatively affects the whole Jungian system.

We urgently need great power alignment, and prevention of the coming escalating proxy warfare. "AI-driven urgency for great-power reconciliation" actually ticks all the ITN framework boxes, IMHO.

I've been reading AI As Normal Technology by Arvind Narayanan and Sayash Kapoor: https://knightcolumbia.org/content/ai-as-normal-technology. You may know them as the people behind the AI Snake Oil blog.

I wanted to open up a discussion about their concept-cutting of AI as "normal" technology, because I think it's really interesting, but also gets a lot of stuff wrong.

I don't see the word "war" or "conflict" in this otherwise nice discussion...

I agree and wrote about it here: https://forum.effectivealtruism.org/posts/nnf2GsSq9fhdRCvZj/keynesian-altruism

The issue is not whether the AI understands human morality. The issue is whether it cares.

The arguments from the "alignment is hard" side that I was exposed to don't rely on the AI misinterpreting what the humans want. In fact, superhuman AI assumed to be better at humans at understanding human morality. It still could do things that go against human morality. Overall I get the impression you misunderstand what alignment is about (or maybe you just have a different association to words as "alignment" than me).

Whether a language model can play a nice character that would totally give back the dictatorial powers after takeover is barely any evidence whether the actual super-human AI system will step back from its position of world dictator after it has accomplished some tasks.

I think you make an important point that I'm inclined to agree with. 

Most of the discourse, theories, intuitions, and thought experiments about AI alignment was formed either before the popularization of deep learning (which started circa 2012) or before the people talking and writing about AI alignment started really caring about deep learning. 

In or around 2017, I had an exchange with Eliezer Yudkowsky in an EA-related or AI-related Facebook group where he said he didn't think deep learning would lead to AGI and thought symbolic AI would instead. Clearly, at some point since then, he changed his mind. 

For example, in his 2023 TED Talk, he said he thinks deep learning is on the cusp of producing AGI. (That wasn't the first time, but it was a notable instance and an instance where he was especially clear on what he thought.)

I haven't been able to find anywhere where Eliezer talks about changing his mind or explains why he did. It would probably be helpful if he did.

All the pre-deep learning (or pre-caring about deep learning) ideas about alignment have been carried into the ChatGPT era and I've seen a little bit of discourse about this, but only a little. It seems strange that ideas about AI itself would change so much over the last 13 years and ideas about alignment would apparently change so little. 

If there are good reasons why those older ideas about alignment should still apply to deep learning-based systems, I haven't seen much discussion about that, either. You would think there would be more discussion.

My hunch is that AI alignment theory could probably benefit from starting with a fresh sheet of paper. I suspect there is promise in the approach of starting from scratch in 2025 without trying to build on or continue from older ideas and without trying to be deferential toward older work.

I suspect there would also be benefit in getting out of the EA/Alignment Forum/LessWrong/rationalist bubble. 

I think 'biodiversity' generally implies a commitment to maintaining a very large number of species, over and above the identifiable value each one provides. It's not about protecting specifically identified valuable species.

It's worth saying that the fact that most arrows go up on the OWiD chart could just point to two independent trends, one of growth rising almost everywhere and another of happiness rising almost everywhere, for two completely independent reasons. Without cases where negative or zero growth persists for a long time, it's hard to rule this out. 

Mo Putera @ 2025-04-17T04:42 (+2)

It could in theory, but OWID's summary of the evidence mostly persuades me otherwise. Again I'm mostly thinking about how the Easterlin paradox would explain this OWID chart: 

I'm guessing Easterlin et al would probably counter that OWID didn't look at a long-enough timeframe (a decade is too short), and I can't immediately see what the timeframe is in this chart, so there's that.

Can confirm; Zipline is ridiculously cool. I saw their P1 Drones in action in Ghana and met some of their staff at EA conferences. Imo, Zipline is one of the most important organizations around for last-mile delivery infrastructure. They're a key partner for GAVI, and they save lives unbelievably cost-effectively by transporting commodities like snakebite antivenom within minutes to people who need it. 

Their staff and operations are among the most high-performing of any organization I've ever seen. Here are some pics from a visit I took to their bay area office in October 2024. I'd highly recommend this Mark Rober video, and checking out Zipline's website. If any software engineers are interested in high-impact work, I would encourage you to apply to Zipline!

Zipline have been around for about 10 years I think - boy do they have the cool factor. One big issue is that they can only carry as really tiny amount of stuff. Also the  places where they can potentially save money have to be super hard to access, because a dirt cheap motorcycle which can go 50km for a dollar of fuel can carry 50x as much weight.

My lukewarm take is that hey have done well, but as with most things haven't quite lived up to their initial hype.

I think you make an important point that I'm inclined to agree with. 

Most of the discourse, theories, intuitions, and thought experiments about AI alignment was formed either before the popularization of deep learning (which started circa 2012) or before the people talking and writing about AI alignment started really caring about deep learning. 

In or around 2017, I had an exchange with Eliezer Yudkowsky in an EA-related or AI-related Facebook group where he said he didn't think deep learning would lead to AGI and thought symbolic AI would instead. Clearly, at some point since then, he changed his mind. 

For example, in his 2023 TED Talk, he said he thinks deep learning is on the cusp of producing AGI. (That wasn't the first time, but it was a notable instance and an instance where he was especially clear on what he thought.)

I haven't been able to find anywhere where Eliezer talks about changing his mind or explains why he did. It would probably be helpful if he did.

All the pre-deep learning (or pre-caring about deep learning) ideas about alignment have been carried into the ChatGPT era and I've seen a little bit of discourse about this, but only a little. It seems strange that ideas about AI itself would change so much over the last 13 years and ideas about alignment would apparently change so little. 

If there are good reasons why those older ideas about alignment should still apply to deep learning-based systems, I haven't seen much discussion about that, either. You would think there would be more discussion.

My hunch is that AI alignment theory could probably benefit from starting with a fresh sheet of paper. I suspect there is promise in the approach of starting from scratch in 2025 without trying to build on or continue from older ideas and without trying to be deferential toward older work.

I suspect there would also be benefit in getting out of the EA/Alignment Forum/LessWrong/rationalist bubble. 

Nice! I've been enjoying your quick takes / analyses, and find your writing style clear/easy to follow. Thanks Mo! (I think this could have been a great top level post FWIW, but to each their own :) )

I'm not seeing where Deena wrote that biodiversity in general was important? 

Both studies suggest that protecting certain animal populations might have large, direct effects on human health that we’re overlooking. But there are good reasons to be cautious. These are outlier results; there isn’t much else in the way of evidence for estimates of this magnitude for the impact of biodiversity loss on human mortality. There’s also the possibility of publication bias. In particular, since both papers come from the same author, this may be driven by a file drawer effect, where a researcher looks at many potential similar cases but the null findings are less likely to see the light of day.

Still, if these effects are real, they could change how we think about conservation. Saving vultures or bats wouldn’t just be about biodiversity—it could also be a form of public health policy.

Was sent a resource in response to this quick take on effectively opposing Trump that at a glance seems promising enough to share on its own: 

From A short to-do list by the Substack Make Trump Lose Again:

1. Friends in CA, AZ, or NM: Ask your governor to activate the national guard (...)
2. Friends in NC: Check to see if your vote in the NC Supreme Court race is being challenged (...)
3. Friends everywhere: Call your senators and tell them to vote no on HR 22 (...)
4. Friends everywhere: If you’d like to receive personalized guidance on what opportunities are best suited to your skills or geographic area, we are excited to announce that our personalized recommendation form has been reopened! Fill it in here!

Bolding is mine to highlight the 80k-like opportunity. I'm abusing the block quote a bit by taking out most of the text, so check out the actual post if interested! 

There's also a volunteering opportunities page advertising "A short list of high-impact election opportunities, continuously updated" which links to a notion page that's currently down.

Thanks for this example, Geoffrey. Hm, that's interesting! This has gotten a bit more complicated than I thought.

It seems: 

1. Surprisingly, scale stretching could lead to attenuation or magnification depending on the underlying relationship (which is unobserved)

Let h be latent happiness; let LS be reported happiness. 

Your example:

$P\left(h\right)=0.40-h/100$

$t=1,h\equiv LS=>dP/dLS=-1/100$

$t=2,h=2LS=>dP/dh\ast dh/dLS=2\ast -1/100=-1/50$

So yes, the gradient gets steeper. 

Consider another function. (This is also decreasing in h)

$P\left(h\right)=1/h$ 

$t=1,h=LS=>dP/dLS=dh/dLS=dP/dh\ast dh/dLS=-1/h^2\ast 1=-1/\left(L{S}^2\right)$

$t=2,h=2LS=>dp/dLS=dP/dh\ast dh/dLS=-1/h^2\ast 2=-2/\left(4L{S}^2\right)=-1/\left(2L{S}^2\right)$

i.e., the gradient gets flatter. 

2. Scale shifting should always lead to attenuation (if the underlying relationship is negative and convex, as stated in the piece)

Your linear probability function doesn't satisfy convexity. But, this seems more realistic, given the plots from Oswald/Kaiser look less than-linear, and probabilities are bounded (whilst happiness is not).

Again consider:

$P\left(h\right)=1/h$

T=1: LS = h => P(h) =1/LS

T=2: LS = h-5 <=> h = LS+5 => P(h) = 1/(LS+5)

Overall, I think the fact that the relationship stays the same is some weak evidence against shifting – not stretching. FWIW, in the quality-of-life literature, shifting occurs but little stretching. 

Sorry to hear man. I tried to reach out to someone at OP a few months ago when I heard about your funding difficulties but I got ignored :(. Anyways, donated $100 and made a twitter thread here

Yesterday, I watched this talk by François Chollet, which provides support for a few of the assertions I made in this post.

Sorry, I agree this message is somewhat misleading - I'll ask our ops team to review this.

Hey! Thanks for the comments. I’d be super happy to hear your personal takes, Caleb!

Ok, but the message I received was specifically saying you can’t fund for-profits and that we can re-apply as a non-profit:

"We rejected this on the grounds that we can't fund for-profits. If you reorganize as a non-profit, you can reapply to the LTFF in an future funding round, as this would change the application too significantly for us to evaluate it in this funding round.
Generally, we think it's good when people run for-profits, and other grant makers can fund them."

We will reconsider going the for-profit route in the future (something we’ve thought a lot about), but for now have gotten funding elsewhere as a non-profit to survive for the next 6 months.

Given that they've made a public Manifund application, it seems fine to share that there has been quite a lot of discussion about this project on the LTFF internally. I don't think we are in a great place to share our impressions right now, but if Connor would like me to, I'd be happy to share some of my takes in a personal capacity.

Hi Deena, thanks for sharing this! As an occupational health epidemiologist, the point about environmental noise exposure particularly resonated with me.

In occupational settings, we take noise seriously: we monitor exposures, set enforceable thresholds, and implement controls. But in communities, chronic environmental noise often goes unmeasured and unaddressed – despite clear links to the health issues you mentioned.

There’s a lot we could borrow from occupational health to protect the public more effectively. A few examples:

1. Community noise mapping and thresholds: Just like exposure assessments at work, cities could monitor residential noise levels over time – especially at night – and act when WHO-recommended thresholds (e.g., 55 dB Lnight) are exceeded.

2. Zoning and built environment controls: Like engineering controls in workplaces, urban planning could prioritise noise buffers like green spaces, sound-insulating materials in construction, or rerouting traffic away from dense housing.

3. Noise fatigue tracking in high-risk populations: In occupational health, we monitor fatigue and hearing loss over time. A similar approach could be piloted in schools, elder care, or high-exposure neighbourhoods using wearable tech or longitudinal surveys.

Noise might be “invisible,” but it’s a modifiable risk factor. We just need to start treating it that way in public health.

Thanks. We should probably try to display this on our website properly. We have been able to fund for-profits in the past, but it is pretty difficult. I don't think the only reason we passed on your application was that it's for-profit, but that did make our bar much higher (this is a consequence of US/UK charity law and isn't a reflection on the impact of non-profits/for-profits).

By the way, I personally think that your project should probably be a for-profit, as it will be easier to raise funding, users will hold you to higher standards, and your team seems quite value-aligned.

In case this is useful to anyone in the future: LTFF does not provide funding for for-profit organizations. I wasn't able to find mentions of this online, so I figured I should share.

I was made aware of this after being rejected today for applying to LTFF as a for-profit. We updated them 2 weeks ago on our transition into a non-profit, but it was unfortunately too late, and we'll need to send a new non-profit application in the next funding round.

Thanks for highlighting the relative lack of attention paid to cause prioritization and cross-cause prioritization. 

I have also written about how important it is to enable EAs to become familiar with existing cause prioritization findings. It's not just about how much research is done but also that EAs can take it into account and act on it.

I don't judge people for having a different eating pattern than me (i eat like 90% plenny shake 😅), but I do judge people who aren't vegan. That question tripped me up a bit, I think I answered somewhat agree but in the spirit of it I probably should've answered strongly agree

Sorry to hear man. I tried to reach out to someone at OP a few months ago when I heard about your funding difficulties but I got ignored :(. Anyways, donated $100 and made a twitter thread here

Some AI research projects that (afaik) haven't had much work done on them and would be pretty interesting:

• If the US were to co-build secure data centres in allied countries, would that be geopolitically stabilising or destabilising?
• What AI safety research agendas could be massively sped up by AI agents? What properties do they have (e.g. easily checkable, engineering > conceptual ...)?
• What will the first non-AI R&D uses of powerful and general AI systems be?
• Are there ways to leverage cheap (e.g. 100x lower than present-day cost) intelligence or manual labour to massively increase the US's electricity supply?
• What kinds of regulation might make it easier to navigate an intelligence explosion (e.g. establishing quick pathways to implement policy informed by AI experts, or establishing zones where compute facilities can be quickly built without navigating a load of red tape)? 

Great post!

As a senior professional who went through the hiring process for EA groups, but also as a senior professional who has hired people (and hires people) both for traditional (profit-driven) organisations and for impact/mission-driven organisations, my only comment would be that this is great advice for any role. 

As hiring managers, we love people who are passionate and curious, and it just feels weird for someone to claim to be passionate about something but not have read up about it or followed what's happening in their field. 

In terms of the job-search within EA, the only detail I would add is that there are a huge number of really nice, friendly, supportive people who give great feedback if you ask. One of my first interviewers did a 1-hour interview, after which he (rightly) did not continue the process. He explained very clearly why and what skills I was missing. He also set up an additional call where he talked through how my skill-set might be most valuable within an impactful role, and some ideas. He gave me lots of connections to people he knew. And so on. And he offered to help if I needed help. 

Within EA, this is the norm. People really respect that someone more senior wants to help make the world a bit better, they want to help. 

 

It’s great that CEA will be prioritizing growing the EA community. IMO this is a long time coming.

Here are some of the things I’ll be looking for which would give me more confidence that this emphasis on growth will go well: 

• Prioritizing high-value community assets. Effectivealruism.org is the de facto landing page for anyone who googles “effective altruism”. Similarly, the EA newsletter is essentially the a mailing list that newbies can join. Historically, I think both these assets have been dramatically underutilized. CEA has acknowledged under-prioritizing effectivealtruism.org (“for several years promoting the website, including through search engine optimization, was not a priority for us”) and the staffmember responsible for the newsletter has also acknowledged that this hasn’t been a priority ( “the monthly EA Newsletter seems quite valuable, and I had many ideas for how to improve it that I wanted to investigate or test… [But due to competing priorities] I never prioritized doing a serious Newsletter-improvement project. (And by the time I was actually putting it together every month, I’d have very little time or brain space to experiment.”) Both assets have the potential to be enormously valuable for many different parts of the EA community.
• Creation of good, public growth dashboards. I sincerely hope that CEA will prioritize creating and sharing new and improved dashboards measuring community growth, the absence of which the community has been questioning for nearly a decade. CEA’s existing dashboard provides some useful information, but it has not always been kept up to date (a recent update helped with this, but important information like traffic to effectivealtruism.org and Virtual Program attendance are still quite stale). And even if all the information were fresh, the dashboard in its current state does not really measure the key question (“how fast is the community growing?”) nor does it provide context on growth (“how fast is the community growing relative to how fast we want it to grow?”) Measuring growth is a standard activity for businesses, non-profits, and communities; EA has traditionally underinvested in such measurement and I hope that will be changing under Zach’s leadership. If growth is “at the core of [CEA’s] mission”, CEA is the logical home for producing a community-wide dashboard and enabling the entire community to benefit from it.
• Thoughtful reflection on growth measurement. CEA’s last public effort at measuring growth was an October 2023 memo for the Meta Coordination Forum. This project estimated that 2023 vs. 2022 growth was 30% for early funnel projects, 68% for mid funnel projects, and 8% for late funnel project. With the benefit of an additional 18 months of metric data and anecdata, these numbers seem highly overoptimistic. Forum usage metrics have been on a steady decline since FTX’s collapse in late 2022, EAG and EAGx attendance and connections have all decreased in 2023 vs. 2022 and 2024 vs. 2023, the number of EA Funds donors continues to decline on a year over year basis as has been the case since FTX’s collapse, Virtual Program attendance is on a multi-year downward trend, etc. There are a lot of tricky methodological issues to sort out in the process of coming up with a meaningful dashboard and I think the MCF memo generally took reasonable first stabs at addressing them; however, future efforts should be informed by shortcomings that we can now observe in the MCF memo approach.
• Transparency about growth strategy and targets. I think CEA should publicly communicate its growth strategy and targets to promote transparency and accountability. This post is a good start, though as Zach writes it is “not a detailed action plan. The devil will of course be in those details.” To be clear, I think it’s important that Zach (who is relatively new in his role) be given a long runway to implement his chosen growth strategy. The “accountability” I’d like to see isn’t about e.g. community complaints if CEA fails to hit monthly or quarterly growth targets on certain metrics. It’s about honest communication from CEA about their long-term growth plan and regularly public check-ins about whether empirical data suggests the plan is going well or not. (FWIW, I think CEA has a lot of room for improvement in this area… For instance, I’ve probably read CEA’s public communications much more thoroughly than almost anyone, and I was extremely surprised to see the claim in the OP that “Growth has long been at the core of our mission.”)

Have you applied to LTFF? Seems like the sort of thing they would/should fund. @Linch @calebp if you have actually already evaluated this project I would be interested in your thoughts as would others I imagine! (Of course, if you decided not to fund it, I'm not saying the rest of us should defer to you, but it would be interesting to know and take into account.)

Hello Charlie. This looks like an interesting research question. However, I do have a few comments on the interpretation and statistical modelling. Both statistical comments are on subtle issues, which I would not expect an undergraduate to be aware of. Many PhD students won't be aware of them either!

On interpretation with respect to the Easterlin paradox: your working model, as far as I can tell, appears to assume that quit rates decrease in a person's latent happiness, but not in their reported happiness. However, if shifts in reporting are caused by social comparisons (i.e., all the other jobs or relationships you see around you improving) then from that direction rescaling no longer implies a flatter relationship, as the quality of the jobs or relationships available upon quitting have increased. However, your results are indicative that other forms of rescaling are not occurring e.g., changes in culture. I think this distinction is important for interpretation. 

The first of the statistical comments that these changes in probabilities of making a change could also be explained by a general increase in the ease of or tendency to get a hospital appointment/new job.  This stems from the non-linearity of the logistic function. Logistic regression models a latent variable that determines the someone's tendency to quit and converts this into a probability. At low probabilities, an increase in this latent variable has little effect on the probability of separation due to the flatness of the curve mapping the latent variable into probabilities. However, the same increase at values of the variable that give probabilities closer to 0.5 will have a big effect on the probability of separation as the curve is steep in this region. As your research question is conceptual (you're interested in whether life satisfaction scales map to an individual's underlying willingness to quit in the same way over time), rather than predicting the probability of separations, the regression coefficients on time interacted with life satisfaction should be the parameter of interest rather than the probabilities. These effects can often go in different directions. A better explanation of this issue with examples is available here: https://datacolada.org/57 I also don't know whether results will be sensitive to assuming a logit functional form relative to other reasonable distributions, such as a probit.

Another more minor comment, is that you need to be careful when adding individual fixed effects to models like this, which you mentioned you did as a robustness check. In non-linear models, such as a logit, doing this often creates an incidental parameters problem that make your regression inconsistent. In this case, you would also be dealing with the issue that it is impossible to separately identify age, time, and cohort effects. Holding the individual constant, the coefficient of a change in time with life satisfaction would include both the time effect you are interested in and an effect of ageing that you are not.

I'd be happy to discuss any of these issues with you in more detail.

Another contributing factor might be that EAs tend to get especially worried when pain stops them from being able to do their work. That would certainly help explain the abnormally high prevalence of wrist pain from typing among EAs.

(NB this wrist pain happened to me years ago and I did get very worried.)

I think it would be good if you could highlight what is new here, vs re-hashing one half of standard arguments (and not covering why people disagree).

This is great!

One minor clarification (that I guess you are taking as "given" for this audience, but doesn't hurt to make explicit) is that the kind of "Within-Cause Prioritization" found within EA is very different from that found elsewhere, insofar as it is still done in service of the ultimate goal of "cross-cause prioritization". This jumped out at me when reading the following sentence:

A quick reading of EA history suggests that when the movement was born, it focused primarily on identifying the most cost-effective interventions within pre-existing cause-specific areas (e.g. the early work of GiveWell and Giving What We Can)

I think an important part of the story here is that early GiveWell (et al.) found that a lot of "standard" charitable cause areas (e.g. education) didn't look to be very promising given the available evidence. So they actually started with a kind of "cause prioritization", and simply very quickly settled on global poverty as the most promising area. This was maybe too quick, as later expansions into animal welfare and x-risk suggest. But it's still very different from the standard (non-EA) attitude of "different cause areas are incommensurable; just try to find the best charity within whatever area you happen to be personally passionate about, and don't care about how it would compare to competing cause areas."

That said, I agree with your general lesson that both broad cause prioritization and specific cross-cause prioritization plausibly still warrant more attention than they're currently getting!

Have you applied to LTFF? Seems like the sort of thing they would/should fund. @Linch @calebp if you have actually already evaluated this project I would be interested in your thoughts as would others I imagine! (Of course, if you decided not to fund it, I'm not saying the rest of us should defer to you, but it would be interesting to know and take into account.)

Since AI X-risk is a main cause area for EA, shouldn't significant money be going into mechanistic interpretability? After reading the AI 2027 forecast, the opacity of AIs appears to be the main source of risk coming from them. Making significant progress in this field seems very important for alignment.

I took the Giving What We Can Pledge, I want to say there should be something like it but for mechanistic interpretability, but probably only very few people could be convinced to give 10% of their income to mechanistic interpretability.

A lot of EAs wanted to slow down AGI development to have more time for alignment. Now Trump's tariffs have done that - accidentally and for the wrong reasons - but they did slow it down. Yet no EA seems happy about this. Given how unpopular his tariffs are maybe people don't want to endorse them for PR reasons? But if you think that AI is by far the most important issue that should easily lead you to say the unpopular truth. Scenarios where China reaches AGI before the US became more likely, but that was always an argument against AI slowdown and it didn't seem to convince many people in the past.

Thoughts?

Maybe this post should be placed in some AI safety thread, but I wasn't sure where exactly.

Executive summary: This exploratory reanalysis uses causal inference principles to reinterpret findings from a longitudinal study on meat reduction, concluding that certain interventions like vegan challenges and plant-based analog consumption appear to reduce animal product consumption, while prior findings suggesting that motivation or outdoor media increase consumption may have stemmed from flawed modeling choices rather than true effects.

Key points:

1. Causal inference requires co-occurrence, temporal precedence, and the elimination of alternative explanations—achievable in longitudinal studies with at least three waves of data, as demonstrated in the case study.
2. The original analysis by Bryant et al. was limited by treating the longitudinal data as cross-sectional, leading to potential post-treatment bias and flawed causal interpretations.
3. The reanalysis applied a modular, wave-separated modeling strategy, using Wave 1 variables as confounders, Wave 2 variables as exposures, and Wave 3 variables as outcomes to improve causal clarity.
4. Motivation to reduce meat consumption was associated with decreased animal product consumption, contradicting the original counterintuitive finding of a positive relationship.
5. Vegan challenge participation and plant-based analog consumption had the strongest associations with reduced consumption and progression toward vegetarianism, though low participation rates limited statistical significance for the former.
6. Some results raised red flags—especially that exposure to activism correlated with increased consumption, prompting calls for further research into the content and perception of activism messages.

This comment was auto-generated by the EA Forum Team. Feel free to point out issues with this summary by replying to the comment, and contact us if you have feedback.

Executive summary: This detailed update from the Nucleic Acid Observatory (NAO) outlines major expansions in wastewater and pooled individual sequencing, air sampling analysis, and data processing capabilities, emphasizing progress toward scalable biosurveillance systems while acknowledging ongoing technical challenges and exploratory efforts.

Key points:

1. Wastewater sequencing has scaled significantly, with over 270 billion read pairs sequenced from thirteen sites—more than all previous years combined—thanks to collaborations with several research labs and support from contracts like ANTI-DOTE.
2. Pooled swab collection from individuals has expanded, with promising Q1 results leading to a decision to scale up; a public report is expected in mid Q2 detailing the findings and rationale.
3. Indoor air sampling work has resulted in a peer-reviewed publication, and the team is actively seeking collaborations with groups already collecting air samples, potentially offering funding for sequencing and processing.
4. Software development continues, with improvements to the main mgs-workflow pipeline and efforts to enhance reference-based growth detection (RBGD) by addressing issues with rare and ambiguous sequences.
5. Reference-free threat detection is being prototyped, including tools for identifying and assembling from short sequences with increasing abundance—efforts recently shared at a scientific conference.
6. Organizationally, the NAO has grown, adding two experienced staff members from Biobot Analytics and securing a $3.4M grant from Open Philanthropy to support wastewater sequencing scale-up, methodological improvements, and rapid-response readiness.

This comment was auto-generated by the EA Forum Team. Feel free to point out issues with this summary by replying to the comment, and contact us if you have feedback.

I'm not asking EA to focus singularly on democracy. I'm asking EA to give any resources at all to the cause of democracy. Prove my ignorance wrong. Is any organization in EA involved with democracy at this moment? Is any organization bothering to evaluate potential interventions? What work has been done? What papers have been written? Is there some work saying, "Look, we've done the work, yes it turns out democracy has a terrible ROI!" How about you guys? Are you making any consideration or analysis on potential pro-democracy interventions? If you have, I'd love to see the analysis. My search for it, I've seen nothing. I hear crickets. 

Here's the thing about evidence. You have to look for it. Is EA bothering to look for it? Is your organization bothering to look for it? Otherwise, you have no idea how tractable it is or is not. 

With Chollet acknowledging that o1/o3 (and ARC 1 getting beaten) was a significant breakthrough, how much is this talk now outdated vs still relevant?

Oh, great, thanks so much! I'll check this out.

Fascinating post! A quick technical tip - marking your post as a link post highlights the original and makes it easier for readers to get to your Substack and subscribe. (Also, looks like the 'Thanks for reading! This post is public so feel free to share it.' from the original didn't get cleaned up when you pasted it into the forum - same thing happened to me my first time!)

Great post! One of the biggest mistakes the EU has made is essentially making snus illegal to sell in stores, and severely restricting access to it online (although it is handled a little bit differently in different countries). I think one of the most impactful and tractable things the THR movement could do is to target this policy. 

How can we unhypocritically expect AI superintelligence to respect "inferior" sentient beings like us when we do no such thing for other species? 

How can we expect AI to have "better" (more consistent, universal, compassionate, unbiased, etc) values than us and also always only do what we want it to do? 

What if extremely powerful, extremely corrigible AI falls into the hands of a racist? A sexist? A speciesist? 

Some things to think about if this post doesn't click for you

haha, yes, people have done this! This is called 'vignette-adjustment'. You basically get people to read short stories and rate how happy they think the character is. There are a few potential issues with this method: (1) they aren't included in long-term panel data; (2) people might interpret the character's latent happiness differently based on their own happiness

Always enjoy your posts, you tend to have fresh takes and clear analyses on topics that feel well-trodden. 

That said, I think I'm mainly confused if the Easterlin paradox is even a thing, and hence whether there's anything to explain. On the one hand there are writeups like Michael Plant's summarising the evidence for it, which you referenced. On the other hand, my introduction to Easterlin's paradox was via Our World in Data's happiness and life satisfaction article, which summarised the evidence for happiness rising over time in most countries here and explain away the Easterlin paradox here as being due to either survey questions changing over time (in Japan's case, chart below) or to inequality making growth not benefit the majority of people (in the US's case).

Plant's writeup says that 

The reply that Easterlin and O’Connor (2022) make is that Stevenson, Wolfers, and co. are looking over too short a time horizon. They point out that the critique looks at segments of ten years and to really test the paradox requires looking over a longer time period, which is what Easterlin and O'Connor (2022) do themselves. Easterlin and O'Connor (2022) write that they don't really understand why Stevenson and Wolfers are using these short time segments rather than the longer ones.

But the chart for Japan above, which is from Stevenson and Wolfers (2008), spans half a century, not ten years, so Easterlin and O'Connor's objection is irrelevant to the chart. 

This OWID chart (which is also the first one you see on Wikipedia) is the one I always think about when people bring up the Easterlin paradox:

But Plant's writeup references the book Origins of Happiness (Clark et al., 2019) which has this chart instead:

So is the Easterlin paradox really a thing or not? Why do the data seem to contradict each other? Is it all mainly changing survey questions and rising inequality? On priors I highly doubt it's that trivially resolvable, but I don't have a good sense of what's going on.

I wish there was an adversarial collaboration on this between folks who think it's a thing and those who think it isn't.

I already registered! This is an exciting opportunity to learn more about Animal Welfare Economics, and who knows, perhaps meet some fellow EAs during the breaks?

Thank you for this deep and thought-provoking post! The concept of the "power-ethics gap" truly resonates and seems critically important for understanding current and future challenges, especially in the context of AI.

The analogy with the car, where power is speed and ethics is the driver's skill, is simply brilliant. It illustrates the core of the problem very clearly. I would even venture to add that, in my view, the "driver's skill" today isn't just lagging behind, but perhaps even degrading in some aspects due to the growing complexity of the world, information noise, and polarization. Our collective ability to make wise decisions seems increasingly fragile, despite the growth of individual knowledge.

Your emphasis on the need to shift the focus in AI safety from purely technical aspects of control (power) to deep ethical questions and "value selection" seems absolutely timely and necessary. This truly is an area that appears to receive disproportionately little attention compared to its significance.

The concepts you've introduced, especially the distinction between Human-Centric and Sentientkind Alignment, as well as the idea of "Human Alignment," are very interesting. The latter seems particularly provocative and important. Although you mention that this might fall outside the scope of traditional AI safety, don't you think that without significant progress here, attempts to "align AI" might end up being built on very shaky ground? Can we really expect to create ethical AI if we, as a species, are struggling with our own "power-ethics gap"?

It would be interesting to hear more thoughts on how the concept of "Moral Alignment" relates to existing frameworks and whether it could help integrate these disparate but interconnected problems under one umbrella.

The post raises many important questions and introduces useful conceptual distinctions. Looking forward to hearing the opinions of other participants! Thanks again for the food for thought!

Are these disagreements representative of the general disagreements between people with long and short AI timelines?

Great post! One of the biggest mistakes the EU has made is essentially making snus illegal to sell in stores, and severely restricting access to it online (although it is handled a little bit differently in different countries). I think one of the most impactful and tractable things the THR movement could do is to target this policy. 

Can confirm; Zipline is ridiculously cool. I saw their P1 Drones in action in Ghana and met some of their staff at EA conferences. Imo, Zipline is one of the most important organizations around for last-mile delivery infrastructure. They're a key partner for GAVI, and they save lives unbelievably cost-effectively by transporting commodities like snakebite antivenom within minutes to people who need it. 

Their staff and operations are among the most high-performing of any organization I've ever seen. Here are some pics from a visit I took to their bay area office in October 2024. I'd highly recommend this Mark Rober video, and checking out Zipline's website. If any software engineers are interested in high-impact work, I would encourage you to apply to Zipline!

If you’ve found the 'snapshot of EA' section particularly valuable, please flag it under this comment so we can gauge how much we should invest in updating it or expanding it in the future. To clarify:
- Vote agree for "particularly valuable".
- Vote disagree for "not that valuable".
Feel free to add any comments.

Some people might find that this post is written from a place of agitation which is fully okay. I think that even if you do there are two things that I would want to point out as really good points:

1. A dependence on funders and people with money as something that shapes social capital and incentives, therefore thought in itself. We should therefore be quite vary of the effect that has on people, this can definetely be felt in the community and I think it is a great point.
2. That the karma algorithm could be revisited and that we should think about what incentives are created for the forum through it. 

I think there's a very very interesting project of democratizingthe EA community in a way that makes it more effective. There are lots of institutional design that we can apply to ourselves and I would be very excited to see more work in this direction!

Edit:

Clarification on why I believe it to cause some agitation for some people:

1. I remember that some of the situation around Cremer being a bit politically loaded and that the emotions were running hot at that time and so citing that specific situation makes it lack a bit of context.
   1. There are some object level things that people within the community disagree with when it comes to these comments that point at deeper issues of epistemics and cause prioritization that is actually difficult to answer.
   2. The post makes it seem more one-sided than that situation was. Elitism in EA is something covered in the in-depth fellowship for example and there's a bunch of back and forth there but it is an issue that you will arrive at different consequences on depending on what modelling assumptions you do.
   3. I don't want to make a value judgement on this here, I just want to point out that specifice piece of Cremer's writing has always felt a bit thorny which makes the references feel a bit inflammatory?
   4. For me it's the vibe that it is written from a perspective of being post EA and something about when leaving something behind you want to get back at the thing itself by pointing out how it's wrong? So it is kind of written from a emotionally framed perspective which makes the epistemics fraught?
2. There's some sort of degree where the framing of the post in itself pattern matches onto other critiques that have felt bad faith and so it is "inflammatory" that it raises the immune system of people reading it. I do still think it is quite a valuable point, it is just that part of the phrasing makes it come across more like this than it has to be?
   1. I think that might be because of LLMs often liking to argue towards a specific point but I'm not sure? 
      (You've got some writing that is reminiscent of claude so I could spot the use of it: e.g): 

This isn’t just a technical issue. This is a design philosophy — one that rewards orthodoxy, punishes dissent, and enforces existing hierarchies.

I liked the post, I think it made a good point, I strong upvoted it but I wanted to mention it as a caveat.

With Chollet acknowledging that o1/o3 (and ARC 1 getting beaten) was a significant breakthrough, how much is this talk now outdated vs still relevant?

Always enjoy your posts, you tend to have fresh takes and clear analyses on topics that feel well-trodden. 

That said, I think I'm mainly confused if the Easterlin paradox is even a thing, and hence whether there's anything to explain. On the one hand there are writeups like Michael Plant's summarising the evidence for it, which you referenced. On the other hand, my introduction to Easterlin's paradox was via Our World in Data's happiness and life satisfaction article, which summarised the evidence for happiness rising over time in most countries here and explain away the Easterlin paradox here as being due to either survey questions changing over time (in Japan's case, chart below) or to inequality making growth not benefit the majority of people (in the US's case).

Plant's writeup says that 

The reply that Easterlin and O’Connor (2022) make is that Stevenson, Wolfers, and co. are looking over too short a time horizon. They point out that the critique looks at segments of ten years and to really test the paradox requires looking over a longer time period, which is what Easterlin and O'Connor (2022) do themselves. Easterlin and O'Connor (2022) write that they don't really understand why Stevenson and Wolfers are using these short time segments rather than the longer ones.

But the chart for Japan above, which is from Stevenson and Wolfers (2008), spans half a century, not ten years, so Easterlin and O'Connor's objection is irrelevant to the chart. 

This OWID chart (which is also the first one you see on Wikipedia) is the one I always think about when people bring up the Easterlin paradox:

But Plant's writeup references the book Origins of Happiness (Clark et al., 2019) which has this chart instead:

So is the Easterlin paradox really a thing or not? Why do the data seem to contradict each other? Is it all mainly changing survey questions and rising inequality? On priors I highly doubt it's that trivially resolvable, but I don't have a good sense of what's going on.

I wish there was an adversarial collaboration on this between folks who think it's a thing and those who think it isn't.

I basically agree with this critique of the results in the post, but want to add that I nonetheless think this is a very cool piece of research and I am excited to see more exploration along these lines!

One idea that I had -- maybe someone has done something like this? -- is to ask people to watch a film or read a novel and rate the life satisfaction of the characters in the story. For instance, they might be asked to answer a question like "How much does Jane Eyre feel satisfied by her life, on a scale of 1-10?". (Note that we aren't asking how much the respondent empathizes with Jane or would enjoy being her, simply how much satisfaction they believe Jane gets from Jane's life.) This might allow us to get a shared baseline for comparison. If people's assessments of Jane's life go up or down over time, (or differ between people) it seems unlikely that this is a result of a violation of "prediction invariance", since Jane Eyre is an unchanging novel with fixed facts about how Jane feels. Instead, it seems like this would indicate a change in measurement: i.e. how people assign numerical scores to particular welfare states.

Thanks for engaging!

This means, specifically, a flatter gradient (i.e., 'attenuation') – smaller in absolute terms. In reality, I found a slightly increasing (absolute) gradient/steeper. I can change that sentence.

I don't think that is necessary - my confusion is more about grasping how the aspects play together :) I'm afraid I will have to make myself a few drawings to get a better grasp.

I think rescaling could make it steeper or flatter, depending on the particular rescaling. Consider that there is nothing that requires the rescaling to be a linear transformation of the original scale (like you've written in your example). A rescaling that compresses the life satisfaction scores that were initially 0-5 into the range 0-3, while leaving the life satisfaction score of 8-10 unaffected will have a different effect on the slope than if we disproportionately compress the top end of life satisfaction scores.

Sorry if I expressed this poorly -- it's quite late :)

I think rescaling could make it steeper or flatter, depending on the particular rescaling. Consider that there is nothing that requires the rescaling to be a linear transformation of the original scale (like you've written in your example). A rescaling that compresses the life satisfaction scores that were initially 0-5 into the range 0-3, while leaving the life satisfaction score of 8-10 unaffected will have a different effect on the slope than if we disproportionately compress the top end of life satisfaction scores.

Sorry if I expressed this poorly -- it's quite late :)

Charlie Harrison @ 2025-04-16T09:19 (+2)

Thanks for this example, Geoffrey. Hm, that's interesting! This has gotten a bit more complicated than I thought.

It seems: 

1. Surprisingly, scale stretching could lead to attenuation or magnification depending on the underlying relationship (which is unobserved)

Let h be latent happiness; let LS be reported happiness. 

Your example:

$P\left(h\right)=0.40-h/100$

$t=1,h\equiv LS=>dP/dLS=-1/100$

$t=2,h=2LS=>dP/dh\ast dh/dLS=2\ast -1/100=-1/50$

So yes, the gradient gets steeper. 

Consider another function. (This is also decreasing in h)

$P\left(h\right)=1/h$ 

$t=1,h=LS=>dP/dLS=dh/dLS=dP/dh\ast dh/dLS=-1/h^2\ast 1=-1/\left(L{S}^2\right)$

$t=2,h=2LS=>dp/dLS=dP/dh\ast dh/dLS=-1/h^2\ast 2=-2/\left(4L{S}^2\right)=-1/\left(2L{S}^2\right)$

i.e., the gradient gets flatter. 

2. Scale shifting should always lead to attenuation (if the underlying relationship is negative and convex, as stated in the piece)

Your linear probability function doesn't satisfy convexity. But, this seems more realistic, given the plots from Oswald/Kaiser look less than-linear, and probabilities are bounded (whilst happiness is not).

Again consider:

$P\left(h\right)=1/h$

T=1: LS = h => P(h) =1/LS

T=2: LS = h-5 <=> h = LS+5 => P(h) = 1/(LS+5)

Overall, I think the fact that the relationship stays the same is some weak evidence against shifting – not stretching. FWIW, in the quality-of-life literature, shifting occurs but little stretching. 

Thanks for sharing, some very interesting ideas.

I'm skeptical about the biodiversity point, at least at that level of generality. It makes sense there are some species that are important for human welfare, maybe in ways that are not initially appreciated, but it seems like a big jump to go from this to biodiversity in general being important.

The improvements to flooring and noise pollution make a lot of sense to me. One interesting intervention I've heard of for the latter is improving the regulations about backup warning alarms on trucks and other vehicles.

Thank you for sharing this – it's also pretty informative as i'm soon moving to DC! 

Love this.Has there really not been an RCT on floor replacements yet? That surprises me as it would be a relatively easy RCT to do. EarthEnable from Rwanda just won the 2 million dollar Skoll award doing this at scale. 

GiveWell must have considered it I would have thought? 

Is there a good list of the highest leverage things a random US citizen (probably in a blue state) can do to cause Trump to either be removed from office or seriously constrained in some way? Anyone care to brainstorm?

Like the safe state/swing state vote swapping thing during the election was brilliant - what analogues are there for the current moment, if any?

Are these disagreements representative of the general disagreements between people with long and short AI timelines?

Always enjoy your posts, you tend to have fresh takes and clear analyses on topics that feel well-trodden. 

That said, I think I'm mainly confused if the Easterlin paradox is even a thing, and hence whether there's anything to explain. On the one hand there are writeups like Michael Plant's summarising the evidence for it, which you referenced. On the other hand, my introduction to Easterlin's paradox was via Our World in Data's happiness and life satisfaction article, which summarised the evidence for happiness rising over time in most countries here and explain away the Easterlin paradox here as being due to either survey questions changing over time (in Japan's case, chart below) or to inequality making growth not benefit the majority of people (in the US's case).

Plant's writeup says that 

The reply that Easterlin and O’Connor (2022) make is that Stevenson, Wolfers, and co. are looking over too short a time horizon. They point out that the critique looks at segments of ten years and to really test the paradox requires looking over a longer time period, which is what Easterlin and O'Connor (2022) do themselves. Easterlin and O'Connor (2022) write that they don't really understand why Stevenson and Wolfers are using these short time segments rather than the longer ones.

But the chart for Japan above, which is from Stevenson and Wolfers (2008), spans half a century, not ten years, so Easterlin and O'Connor's objection is irrelevant to the chart. 

This OWID chart (which is also the first one you see on Wikipedia) is the one I always think about when people bring up the Easterlin paradox:

But Plant's writeup references the book Origins of Happiness (Clark et al., 2019) which has this chart instead:

So is the Easterlin paradox really a thing or not? Why do the data seem to contradict each other? Is it all mainly changing survey questions and rising inequality? On priors I highly doubt it's that trivially resolvable, but I don't have a good sense of what's going on.

I wish there was an adversarial collaboration on this between folks who think it's a thing and those who think it isn't.

To synthesize a few of the comments on this post -- This comment sounds like a general instance of the issue that @geoffrey points out in another comment: what @Charlie Harrison is describing as a violation of "prediction invariance" may just be a violation of "measurement invariance"; in particular because happiness (the real thing, not the measure) may have a different relationship with GMEOH events over time.

Ah I missed the point about the relationship getting flatter before. Thanks for flagging that.

I think I'm more confused about our disagreement now. Let me give you a toy example to show you how I'm thinking about this. So there's three variables here:

• latent life satisfaction, which ranges from 0 to infinity
• reported life satisfaction, which ranges from 0 to 10 and increases with latent life satisfaction
• probability of divorce, which ranges from 0% to 100% and decreases with latent life satisfaction

And we assume for the sake of contradiction that rescaling is true. One example could be:

• In t=1, latent life satisfaction = 1 * reported life satisfaction
  • Both are bounded from [0,10]
• In t=2, latent life satisfaction = 2 * reported life satisfication.
  • Reported life satisfaction still ranges from [0,10]
  • But latent life satisfaction now ranges from [0,20]

Let's say that's true. Let's also assume people divorce less as they get happier (and let's ignore my earlier 'divorce gets easier' objection). One example could be:

• In t=1 and t=2, probability of divorce = 0.40 - latent life satisfaction/100. That implies:
  • In t=1, probability of divorce ranges from [0.40, 0.30]
  • In t=2, probability of divorce ranges from [0.40, 0.20]

And so if I got the logic right, rescaling should accentuate (make steeper) the relationship between probability of divorce and reported life satisfaction. But I think you're claiming rescaling should attenuate (make flatter) the relationship. So it seems like we're differing somewhere. Any idea where?

The phenomenon you describe as "rescaling" is generally known as a (violation of) measurement invariance across in psychometrics. It is typically tested by observing whether the measurement model (i.e., the relationship between the unobservable psychological construct and the measured indicators of that construct) differ across groups (a comprehensive evaluation of different approaches is in Millsap, 2011).

I would interpret the tests of measurement invariance you use.... 

If people are getting happier over time — but reporting it on a stretched or stricter scale — then the link between how happy someone says they are, and what they do when they're unhappy, should weaken over time.

In other words: if life satisfaction is increasing, but the reporting scale is stretching, then big life decisions — like leaving a job or ending a relationship — should become less predictable from reported happiness

....to actually be measures of "prediction invariance": which holds when a measure has the same regression coefficient with respect to an external criterion across different groups or time.

But as Borsboom (2006) points out, prediction invariance and measurement invariance might actually be in tension with each other under a wide range of situations. Here's a relevant quotation:

In 1997 Millsap published an important paper in Psychological Methods on the relation between prediction invariance and measurement invariance. The paper showed that, under realistic conditions, prediction invariance does not support measurement invariance. In fact, prediction invariance is generally indicative of violations of measurement invariance: if two groups differ in their latent means, and a test has prediction invariance across the levels of the grouping variable, it must have measurement bias with regard to group membership. Conversely, when a test is measurement invariant, it will generally show differences in predictive regression parameters.

This is stretching my knowledge of the topic beyond its bounds, but this issue seems related to the general inconsistency between measurement invariance and selection invariance, which has been explored independently in psychometrics and machine learning (e.g., the chapters on facial recognition and  recidivism in The Alignment Problem). 

I don't see how you can 'multiply by 1000+ other species' given these two examples were likely selected for being unusually large. 

I agree that not everyone already knows what they need to know. Our crux issue is probably "who needs to get it and how will they learn it?" I think we more than have the evidence to teach and set an example of knowing for the public. I think you think we need to make a very respectable and detailed case to convince elites. I think you can take multiple routes to influencing elites and that they will be more receptive when the reality of AI risk is a more popular view. I don't think timelines are a great tool for convincing either of these groups because they create such a sense of panic and there's such an invitation to quibble with the forecasts instead of facing the thrust of the evidence. 

Thank you for writing this comprehensive proposal. I agree with your conclusion it's not a case of if but when and we should be improving our pandemic planning now.

Industrial animal agriculture creates conditions where pathogens can evolve and spread rapidly between densely housed animals, potentially creating new zoonotic diseases that can jump to humans. This factor alone raises the likelihood of future pandemics and strengthens the case for robust early detection systems.

The comparison to fire protection spending provides a compelling perspective. It's striking that New Zealand spent nearly 3 times more on fire protection than pandemic preparedness, despite COVID-19 costing the country roughly 50 times more than annual fire damage. This kind of data-driven comparison makes a strong case for increasing pandemic surveillance investment.

I hope you're able to get this information to MoH!

I have the opposite intuition for biodiversity. People have been studying ecosystem services for decades and higher biodiversity is associated with increased ecosystem services, such as clean water, air purification, and waste management. Higher biodiversity is also associated with reduce transmission of infectious diseases by creating more complex ecosystems limiting pathogen spread. Then we have the actual and possible discovery of medicinal compounds and links with biodiversity and mental health. These are high level examples of the benefits. The linked article gives the possibility of impact by considering two effects from bats and vultures. Multiply that effect by 1000+ other species, include all the other impacts previously mentioned and I can see how this could be high impact. 

I just learned about Zipline, the world's largest autonomous drone delivery system, from YouTube tech reviewer Marques Brownlee's recent video, so I was surprised to see Zipline pop up in a GiveWell grant writeup of all places. I admittedly had the intuition that if you're optimising for cost-effectiveness as hard as GW do, and that your prior is as skeptical as theirs is, then the "coolness factor" would've been stripped clean off whatever interventions pass the bar, and Brownlee's demo both blew my mind with its coolness (he placed an order on mobile for a power bank and it arrived by air in thirty seconds flat, yeesh) and also seemed the complete opposite of cost-effective (caveating that I know nothing about drone delivery economics). Quoting their "in a nutshell" section:

In December 2024, GiveWell recommended a $54,620 grant to Zipline for a six-month scoping project. Zipline will use this time to review ways that they could use drones to increase vaccination uptake, especially in hard-to-reach areas with low vaccination coverage and high rates of vaccine-preventable diseases. ...

We recommended this grant because:

• Drones are an intuitive way to bring vaccines closer to communities with low coverage rates, especially when demand for vaccination exists, but conditions like difficult terrain, poor infrastructure, weak cold chain, or insecurity make it difficult for families to access immunizations.
• This grant aligns with our strategy of making several scoping grants to high-potential organizations to source promising ideas for solving bottlenecks in the routine immunization system, and then testing these concepts.

Okay, but what about cost-effectiveness? Their "main reservations" section says

Evidence on cost-effectiveness of drones for vaccine delivery is limited, and we have not modeled the cost-effectiveness of the types of programs that Zipline plans to consider, nor the value of information for this scoping grant. 

An internal review conducted in 2023 focused on drones for health was generally skeptical about there being many opportunities in this area that would meet GiveWell’s bar, although this scoping grant will focus on the most promising scenario (remote areas with high rates of vaccine-preventable diseases and low vaccination coverage rates).

Is there any evidence of cost-effectiveness at all then? According to Zipline, yes — e.g. quoting the abstract from their own 2025 modelling study:

Objectives: In mid-2020, the Ghana Health Service introduced Zipline’s aerial logistics (centralized storage and delivery by drones) in the Western North Region to enhance health supply chain resilience. This intervention led to improved vaccination coverage in high-utilization districts. This study assessed the cost-effectiveness of aerial logistics as an intervention to improve immunization coverage. 

Methods: An attack rate model, adjusted for vaccination coverage and vaccine efficacy, was used to estimate disease incidence among vaccinated and unvaccinated populations, focusing on 17 022 infants. Incremental cost-effectiveness ratios of US dollar per averted disability-adjusted life-year (DALY) were evaluated from societal and government perspectives, using real-world operations data. ...

Results: In 2021, aerial logistics averted 688 disease cases. Incremental cost-effectiveness ratios were $41 and $58 per averted DALY from the societal and government perspectives, respectively. The intervention was cost-saving when at least 20% of vaccines delivered by aerial logistics replaced those that would have been delivered by ground transportation, with potential government savings of up to $250 per averted DALY. Probabilistic sensitivity analysis confirmed the robustness of these findings. 

That's super cost-effective. For context, the standard willingness-to-pay to avert a DALY is 1x per capita GDP or $2,100 in Ghana, so 35-50x higher. Also:

... we calculated that aerial logistics facilitated the completion of an additional 14 979 full immunization courses... We estimated that 4 children’s lives (95% CI 2–7) were saved in these districts during 2021. ... the intervention averted a total of $20 324 in treatment costs and $2819 for caregivers between lost wages and transport. 

At a cost of $0.66 per incremental FIC (fully immunized child), this approach outperforms other delivery methods analyzed in the review, including the most cost-effective category of interventions identified, namely “Delivery Approach” interventions, such as monthly immunization by mobile teams in villages and the enhancement of satellite clinic immunization practices. 

(GW notes that they'd given Zipline's study a look and "were unable to quickly assess how key parameters like program costs and the impact of the program on vaccination uptake and disease were being estimated". Neither can I. Still pretty exciting)

Thanks for sharing, some very interesting ideas.

I'm skeptical about the biodiversity point, at least at that level of generality. It makes sense there are some species that are important for human welfare, maybe in ways that are not initially appreciated, but it seems like a big jump to go from this to biodiversity in general being important.

The improvements to flooring and noise pollution make a lot of sense to me. One interesting intervention I've heard of for the latter is improving the regulations about backup warning alarms on trucks and other vehicles.

Sorry – this is unclear. 

"If people are getting happier (and rescaling is occuring) the probability of these actions should become less linked to reported LS"

This means, specifically, a flatter gradient (i.e., 'attenuation') – smaller in absolute terms. In reality, I found a slightly increasing (absolute) gradient/steeper. I can change that sentence.

I could imagine thinking about "people don't settle for half-good any more" as a kind of increased happiness

This feels similar to Geoffrey's comment. It could be that it takes less unhappiness for people to take decisive life action now. But, this should mean a flatter gradient (same direction as rescaling)

And yeah, this points towards culture/social comparison/expectations being more important than absolute £.

(I know I'm late again replying to this thread.)

What surprises me about this whole situation is that people seem surprised at the executive leadership at a corporation worth an estimated $61.5B would engage in big-corporation PR-speak. The base rate for big-corporation execs engaging in such conduct in their official capacities seems awfully close to 100%.

Hm, good point. This gives me pause, but I'm not sure what direction to update in. Like, maybe I should update "corporate speak is just what these large orgs do and it's more like a fashion thing than a signal of their (lack of) integrity on things that matter most." Or maybe I should update in the direction you suggest, namely "if an org grows too much, it's unlikely to stay aligned with its founding character principles." 

I'm getting the sense that a decent number of people assume that being "EA aligned" is somehow a strong inoculant against the temptations of money and power.

I would have certainly thought so. If anything can be an inoculant against those temptations, surely a strong adherence to a cause greater than oneself packaged in lots warnings against biases and other ways humans can go wrong (as is the common message in EA and rationalist circles) seems like the best hope for it? If you don't think it can be a strong inoculant, that makes you pretty cynical, no? (I think cynicism is often right, so this isn't automatically a rejection of your position. I just want to flag that yours is a claim with quite strong implications on its own.)

Arguably the FTX scandal -- which after all involved multiple EAs, not just SBF -- should have already caused people to update on how effective said inoculant is, at least when billions of dollars were floating around.

If you were just talking about SBF, then I'd say your point is weak because he probably wasn't low on dark triad traits to start out with. But you emphasizing how other EAs around him were also involved (the direct co-conspirators at Alameda and FTX) is a strong point.

Still, in my mind this would probably have gone very differently with the same group of people minus SBF and with a leader with a stronger commitment and psychological disposition towards honesty. (I should flag that parts of Caroline Ellison's blog also gave me vibes of "seems to like having power too much" -- but at least it's more common for young people to later change/grow.) That's why I don't consider it a huge update for "power corrupts". To me, it's a reinforcement of "it matters to have good leadership."

My worldview(?) is that "power corrupts" doesn't apply equally to every leader and that we'd be admitting defeat straight away if we stopped trying to do ambitious things. There doesn't seem to be a great way to do targeted ambitious things without some individual acquiring high amounts of power in the process.(?) We urgently need to do a better job at preventing that those who end up with a lot of power are almost always those with kind of shady character. The fact that we're so bad at this suggests that these people are advantaged at some aspects of ambitious leadership, which makes the whole thing a lot harder. But that doesn't mean it's impossible. 

I concede that there's a sense in which this worldview of mine is not grounded in empiricism -- I haven't even looked into the matter from that perspective. Instead, it's more like a commitment to a wager: "If this doesn't work, what else are we supposed to do?" 
I'm not interested in concluding that the best we can do is criticise the powers that be from the sidelines.

Of course, if leaders exhibit signs of low integrity, like in this example of Anthropic's communications, it's important not to let this slide. The thing I want to push back against is an attitude of "person x or org y has acquired so much power, surely that means that they're now corrupted," and this leading to no longer giving them the benefit of the doubt/not trying to see the complexities of their situation when they do something that looks surprising/disappointing/suboptimal. With great power comes great responsiblity, including a responsibility to not mess up your potential for doing even more good later on. Naturally, this does come with lots of tradeoffs and it's not always easy to infer from publicly visible actions and statements whether an org is still culturally on track. (That said, I concede that you can often tell quite a lot about someone's character/an org's culture based on how/whether they communicate nuances, which is sadly why I've had some repeated negative updates about Anthropic lately.)

Thanks for sharing, some very interesting ideas.

I'm skeptical about the biodiversity point, at least at that level of generality. It makes sense there are some species that are important for human welfare, maybe in ways that are not initially appreciated, but it seems like a big jump to go from this to biodiversity in general being important.

The improvements to flooring and noise pollution make a lot of sense to me. One interesting intervention I've heard of for the latter is improving the regulations about backup warning alarms on trucks and other vehicles.

If we take the premise that income is the single most important factor correlated with happiness, then I think the acceleration effects do seem to imply that there is no happiness ceiling. However, I'm not sure how reasonable this premise is in the first place. I suspect we're zooming in on a well studied effect and if we zoom out a bit, there are many plausible hypotheses for why acceleration effects does not rule out a happiness ceiling, namely that other factors impact the high end of the scale more.

I notice this being muddied in the references to the happiness ceiling. On one hand, the happiness ceiling is only being defined or evaluated in the income sense, and on the other, the conclusions are described as though the income-happiness-ceiling is the only effect and therefore equivalent to all possible models of the happiness ceiling.

A separate question: how does the mathematical relationship relate in practice, e.g. if I have 9/10 happiness, then 10x my income, then am "only" 10/10 happy because I can't exceed 10? I haven't seen this explained before, and I have some concerns about whether it's valid to draw conclusions about this part of the curve without more complicated design. (In other words, I think the extreme end of the scale is an exception and that different study design is required to understand it more objectively.)
 

Hi Geoffrey,

Thank you!

It's possible that these 3 exit actions have gotten easier to do, over time. Intuitively, though, this would be pushing in the same direction as rescaling: e.g., if getting a divorce is easier, it takes less unhappiness to push me to do it. This would mean the relationship should (also) get flatter. So, still surprising, that the relationship is constant (or even getting stronger). 

Any comment or thoughts would be helpful. Happy to connect on linkedin and message as well https://www.linkedin.com/in/cadejs/