Frontpage posts - Effective Altruism forum viewer

The Long (Self-)Correction by Wei Dai

Wei Dai — Fri, 24 Jul 2026 21:01:47 +0000

I propose the Long Self-Correction^[1] as an alternative name/idea/concept to AI Pause and Long Reflection.

Problem with AI Pause: Pause until when, and for what purpose? Presumably to make AI (that we’ll build later) safer, but the deeper problem is that humans aren’t safe, and can’t safely serve as builders, overseers, or alignment targets for powerful AIs.

Problem with Long Reflection: It seems to imply that the main problem with humans is that we just haven’t had enough time to think, that reflection is the main thing we need to do more of, and then we can get on with building powerful AIs or other technologies. Or that if we build aligned AIs that sincerely help us think a lot more, or do the thinking for us, then things will turn out fine.

So I think we need a catchy handle for a related but distinct idea, that humans aren’t ready to build AIs or other extremely powerful technologies, because we’re currently too flawed, in a variety of ways, and it will take a long process (which may or may not end up succeeding) to fix those flaws.

A summary of the flaws that I have in mind:

not having a workable moral framework (consequentialism, deontology, virtue ethics all having serious problems)
being bad at philosophy and long-horizon strategy
being badly calibrated about our philosophical and strategic competence, i.e., not realizing that we’re incompetent, despite overwhelming evidence (see e.g. FTX and early MIRI, and many others, trying to maximize impact while assuming their own philosophical and strategic competence)
in practice, human morality is a kind of status game that actively disvalues careful strategy and philosophy in most places
positional/zero-sum values (like power and social status) being a huge part of human motivations, but almost nobody explicitly reasons/talks about this while discussing, for example, AI safety or effective altruism or how to make the long term future turn out well^[2]
being easy to manipulate (or go off-rails by oneself) via things like sycophancy, plausible-sounding philosophical arguments (see e.g. Objectivism or Dialectical Materialism), limerence, hero worship, spirituality
not understanding the nature of philosophy so that we don’t know how to go about fixing many of these issues even in principle
not realizing how flawed humans are, or the full scope and interlocking nature of the human and AI safety problems, and thus being prone to over-optimistically proposing or implementing partial solutions (e.g., thinking that making AIs aligned or corrigible to humans or to some favorite moral theory would be sufficient to make them safe, or that we just need the right people to win the AI race, or Tool AI, or Scientist AI)

(This list focuses on key bottlenecks that seem hard to fix even with AI assistance or intelligence enhancement, and isn’t meant to be a complete list of human flaws / safety problems. It ignores e.g. that the median human is ignorant of many important issues, and that we’re currently quite bad at complex large-scale coordination such as passing/implementing close-to-optimal government policies.)

My main hope for a Long Self-Correction eventually succeeding rests on the fact that humans have seemingly, mysteriously, made progress on these issues over a very long period of time, so if we preserve the environment in which we can seemingly do this, and not give anyone or anything the power to permanently derail such progress, then maybe we can continue to snowball The Correction until we reach a point when we can rightly justify reshaping the universe according to our volition.

^
It will probably be shortened to “The Long Correction” at some point if it catches on, similar to how “outer space” is now often just “space”.
^
Why isn’t there a version of EA that explicitly talks about how to leverage people’s status motivations to do more good for the world? It’s very possible that explicit talk about status is actually counterproductive at least in the short run, e.g. it heightens status motivations and makes people less altruistic, but then do we just march into the future while blindfolding ourselves to this aspect of human nature?

Coefficient Giving just gave GiveWell $1 billion. Where should other donors give now? by Jack Lewars

Jack Lewars — Fri, 24 Jul 2026 15:37:47 +0000

Linkpost for my Substack piece, lightly adapted. Disagreement very welcome.

What happened

Coefficient Giving (cG) announced a $1 billion gift to GiveWell on 23rd July. This increases a previous commitment of $175m for 2026. cG say this could be a one-off surge, but it has implications for other donors either way. Both organisations have lowered their funding thresholds this year (GiveWell from 8x their benchmark to 6x, Coefficient from 2,000x their benchmark to 1,000x). Even if GiveWell hit their aim of grantmaking that will “approach or even exceed $1 billion this year”, cG’s gift is roughly the size of everything GiveWell expects to grant this year.

They don’t specifically mention AI wealth, but I think it’s fairly clear what they mean when they say “much of the future funding we’re anticipating depends on the valuations of some extremely volatile assets.” They also say that a belief in the potential of AI to transform global health affects their decision. In essence, AI wealth and short AI timelines are now setting budgets for bednets, vitamin-A distribution and vaccination incentives. This is the first major example of how the Funding Anthropalypse will reshape EA.

What should donors do—my take

The answer depends on size. For donors giving thousands, not much has changed: 6x GiveWell’s benchmark is still some of the best marginal cost-effectiveness a normal donor can buy, and GiveWell has committed to tell donors if it thinks their money should go elsewhere. If you are giving thousands each year, I would stay as you are.

For donors giving millions, a lot has changed. The highest impact move is probably not adding to GiveWell’s margin, but building other capacity that will help to absorb next year’s likely windfall. cG concede that “the marginal cost-effectiveness of GiveWell-directed funding will be lower than it has been in the past couple of years”, and that donors redirecting marginal donations would be, to some degree, “perfectly rational”.

Finally, this does nothing for power concentration—if anything, it makes it worse. cG and GiveWell are the two largest players in EA, and they are not independent: the two largest funders are now exposed to the same volatile source at the same time. I expect the gift itself to be an outstanding grant at the object level, funding highly cost-effective programmes that do an enormous amount of good for very little money. But this excellent grant does nothing to address the fragile architecture of the space.

Disclosure: I have a conflict of interest in some of what I recommend in the full piece, as an operator in this space myself via the Mid-stage Global Health Fund. Discount accordingly.

A few things I’d be interested in hearing about:

If you’re setting a seven-figure giving budget for 2026 this week, does this news change it?
Is “build absorption capacity” actually tractable for an individual large donor, or does it fall apart on contact with the pipeline?
How worried should we be about two interdependent funders, versus just glad about $1bn of excellent grants?

Tackling Diarrheal Disease in the Democratic Republic of the Congo (DRC) and South Sudan by GiveWell

GiveWell — Fri, 24 Jul 2026 14:17:32 +0000

This is a new episode of the GiveWell Conversations podcast. You can listen to the episode or read a summary of the conversation below.

Diarrhea is one of the leading causes of death for children under five, killing more than 400,000 children every year.^[1] GiveWell has funded several programs to prevent and treat diarrhea, and we recently expanded our work by making a grant to support the implementation and evaluation of WASHmobile in the Democratic Republic of the Congo (DRC) and South Sudan. The program, which was developed by Johns Hopkins University, delivers water treatment supplies and hygiene messaging to households at the highest risk of diarrheal disease. It has shown promising results in randomized controlled trials (RCTs) but hasn’t yet been implemented at scale in the challenging settings where the need is greatest.

In this episode, GiveWell co-founder and CEO Elie Hassenfeld speaks with Senior Program Officer Erin Crossett, who leads GiveWell’s water research subteam, about the grant. They discuss how the program works, what prior research tells us about its potential, the significant uncertainties involved, and what GiveWell expects to learn.

Listen to Episode 33: Tackling Diarrheal Disease in the DRC and South Sudan

Elie and Erin discuss:

How WASHmobile works: When someone in a household has diarrheal disease, others in the household are at much higher risk. To reduce this risk, WASHmobile first identifies children who have been hospitalized with severe diarrhea. A health worker visits their family to provide handwashing supplies and chlorine tablets to disinfect water, then follows up with phone reminders about safe water practices. Because household members often share the same water source and live in close quarters, treating water and improving hygiene during this critical window can prevent further infections and deaths. Trials of more intensive versions of this program found reductions in diarrheal disease of roughly 60%. Our grant also includes a second, lower-cost component that sends SMS alerts and digital chlorine vouchers to people living near detected outbreaks.
Navigating uncertainty in fragile contexts: WASHmobile is being implemented in South Kivu (DRC) and Jonglei State (South Sudan) by Tearfund, which is partnering closely with the countries’ ministries of health. Because of active armed conflict, severe flooding, and large displaced populations, these areas are challenging for health program implementation. GiveWell is funding a version of the program that is less intensive than the one studied in trials, with shorter duration and less intensive follow-up, and we have discounted expected effects accordingly. Even with those adjustments and a substantial probability of program failure built into our modeling, we estimate the program could be highly cost-effective because disease burden in these areas is so high.
Building learning into the grant: Roughly half of the grant funds implementation, and half is dedicated to evaluation and data collection. In DRC, GiveWell is funding Johns Hopkins University, in partnership with Catholic University of Bukavu, to conduct a randomized controlled trial alongside implementation to measure the program’s effect on hospital admissions for diarrhea among children under five. The trial will also help test one of the key assumptions in our cost-effectiveness model: that the families this program targets are actually at higher mortality risk than the general population. In South Sudan, where a full trial wasn’t feasible, we’re funding mortality surveys and operational monitoring to better understand baseline mortality and whether the program is being implemented as intended.

This grant reflects how GiveWell’s growing team capacity allows us to do more than scale up proven programs; it also enables us to investigate new, cost-effective ways to save and improve lives beyond our traditional grantmaking portfolio. In addition, our growing capacity means we’re increasingly able to help bridge the gap between academic research and real-world implementation—funding promising programs alongside the rigorous evaluation needed to determine whether they should be scaled further.

Visit our All Grants Fund page to learn more about how you can support this work, and listen or subscribe to our podcast for our latest updates.

This episode was recorded on June 10, 2026, and represents our best understanding at that time.

^
“Diarrhoeal disease is the third leading cause of death in children under 5 years old and is responsible for killing around 443 832 children every year.” World Health Organization, “Diarrhoeal disease,” 2024.

Jasmine Sun on what the people building AI really believe by 80000_Hours

80000_Hours — Thu, 23 Jul 2026 14:32:03 +0000

By Zershaaneh Qureshi | Watch on Youtube | Listen on Spotify | Read transcript

Episode summary

I’d be having a conversation with somebody at a conference workshop and they would say, ‘Everyone at this conference is so worried about what’s going to happen to humans in the post-AGI world. But frankly I’m more worried: what if the monkeys are still running the show?’

And in this case the monkeys refers to the humans… he had this view that AIs in their superintelligent existence were more deserving and more important to preserve a future world for…

I don’t think it’s the majority belief or common belief within the AI labs… but I do think that it makes it very hard to move forward policy action, to move forward the safety conversation.

— Jasmine Sun

Many AI researchers believe mass job displacement is coming — and some even think there’s a chance their technology will kill everyone. But they’re building it anyway. Writer and journalist Jasmine Sun has been documenting why from the inside.

Jasmine describes her work as an “anthropology of disruption.” She’s embedded herself in Silicon Valley’s AI subcultures — attending the parties and conferences, conducting off-the-record interviews — to understand the beliefs of the small group of people shaping this technology.

Some of her findings are unsettling. Asked what advice they’d give a normal 17-year-old, almost every AI researcher said the same thing: “I have no idea… It’s a really scary time. I don’t think there’s going to be a lot of jobs for them left.”

Their motives for building advanced AI are varied: a mix of optimism for humanity, techno-determinism, and a desire to secure their own future in the face of a possible “permanent underclass.” A few go even further, actually hoping for a world where machines — rather than humans — are running the show.

When the room can’t even agree on whether humans should stay in control, building a consensus on how to build AI safely gets much harder.

Beyond Silicon Valley, Jasmine’s also tracking the rise of “AI populists,” who see AI as the latest example of corporate elites concentrating their power at the expense of everyone else. In the US, populist sentiment about AI has mostly manifested in protests and votes against data centres. But sometimes, it has escalated into violence: a molotov cocktail thrown at Sam Altman’s house, and open fire on the home of a politician who’d backed a data centre. Jasmine thinks public anger will keep finding an outlet, one way or another, until people feel like they’ll actually share in AI’s gains.

In this interview with host Zershaaneh Qureshi, Jasmine Sun takes us inside the multifarious factions on AI’s bleeding edge. They also discuss:

How “doomer” became the lowest-status label in Silicon Valley, and what that means for AI safety
Why the AI industry’s PR strategy has failed, and what it would take to rebuild public trust
What’s under the surface of the Chinese public’s much more positive response to AI
Jasmine’s reasons to be cautiously hopeful: it’s an unusually high-leverage time to work on AI safety, with policymakers and philanthropists hungry for good ideas

This episode was recorded on June 4, 2026.

Want to get up to speed on AI?

We’ve got a crash course of 10 of our podcast episodes designed to help you get to grips with transformative AI — particularly if you’re new to the topic — and what you can do to help shape its trajectory.

Our production team includes:

Video editors: Josh Alward, Dominic Armstrong, Jasper Luithlen, Milo McGuire, Luke Monsour, Simon Monsour, and Andrés Escobar
Producers: Elizabeth Cox and Nick Stockton
Coordination and support: Katy Moore and Lou Moran
Music: CORBIT

The interview in a nutshell

Jasmine Sun is a journalist who describes her reporting on the AI industry as an “anthropology of disruption.” She’s been inside the teeming subcultures of San Francisco recording the technology-shaping beliefs, fears, and motives of those building frontier AI.

She reveals a fractured landscape with varied motivations — genuine optimism for humanity, techno-determinism, fears of a “permanent underclass,” belief in building humanity’s worthy successor.

Meanwhile, a growing wave of public backlash ranges from political organising to stochastic violence, driven less by technical concerns and more by a deep distrust of concentrated corporate power. The silver lining for the AI-safety-focused: politicians are waking up to the importance of AI, and they’re hungry for expertise and solutions.

AI researchers acknowledge the risks, but they’re building anyway

Jasmine’s been sitting down with frontier lab researchers for long-form, off-the-record “AI ethnographies.” One of her questions: what advice would you give a normal 17-year-old — not a brilliant coder, just an ordinary kid getting Bs?

Almost everyone — regardless of their AI or political beliefs — had no idea. They admitted it’s a scary time, mass job displacement is coming, and young people are caught in a painful transition.

Despite this, researchers keep building. Jasmine identifies several motivations:

Utopian conviction: many genuinely believe that if AI goes right, the upside — cures for all diseases, radical abundance, lives of leisure — justifies weathering near-term harm
Techno-determinism: a widespread belief that someone will build superintelligence regardless — and genuine convictions that being at the frontier is the best way to shape the outcome for the better
Financial self-interest: if massive disruption is coming, people want to ensure a stake in the future for themselves and their families

A smaller group holds what Jasmine calls ‘successionist-lite’ views — an indifference or even preference towards a future where AIs, not humans, are running things. She finds this troubling: it’s hard to advance safety policy if prominent voices can’t even agree that humans should stay in control.

A handful of people’s beliefs are shaping frontier AI

Jasmine argues her anthropological lens is valuable because so few people actually have the technical ability to build frontier AI — meaning their personal beliefs and preferences become load-bearing for potentially civilisation-wide decisions.

Dario Amodei’s personal convictions about not selling to authoritarian governments or enabling mass surveillance translate directly into company policy.
Historical precedent also supports this. The internet’s open, borderless architecture can be traced directly back to the beliefs of a small number of people at DARPA and Stanford.

Vice signalling is making it uncool to be a cautious ‘doomer’

‘Doomer’ has become Silicon Valley’s lowest-status label — shorthand for the kind of “low agency, tech-sceptical” pessimist who’d have vetoed the Industrial Revolution.

Instead, the high-status move is to build something disruptive, such as Mechanize or Cluely, with taglines like, “We’re going to automate and replace all human labour” or “Cheat on everything.” This has a few effects:

Status shapes how many people choose capabilities work over safety — which is a big deal, because talent is incredibly scarce.
In Washington, Jasmine argues the tech right — Marc Andreessen, David Sacks — has effectively tied AI safety to ‘woke’ identity politics.
Journalism also appropriates this language to mock those who believe in existential risk. And because these roles are not as well remunerated or supported as technical work, Jasmine worries about them more.

The good news: technical AI safety seems more insulated from these low-status perceptions — thanks in part to prominent AI pioneers, like Geoffrey Hinton, raising the alarm.

Anti-AI backlash is growing, and it’s not focused strictly on the technology

Jasmine has also spent time on the other side — with “AI populists” and the growing anti-AI coalition: labour unions worried about jobs, social conservatives alarmed by AI companions, environmentalists fighting data centres, and states-rights advocates opposing federal preemption.

AI is no longer only the remit of technocrats. Established interest groups keep realising “AI intersects with my issue” and are forming unlikely cross-issue coalitions.
What makes them populists is that they’re not focused on AI as a uniquely dangerous technology — they see it as the latest front in corporate elites and billionaires concentrating power.
The way AI shows up in the median person’s life is just not that good. Writing emails faster is a poor tradeoff for slop, unemployment, and kids cheating on homework — especially when “this might kill us all” is part of the equation.

Without a way to negotiate, AI populism threatens to escalate

Jasmine compares today’s situation with 20th-century factory automation — the key difference is that unions gave factory workers the ability to negotiate.

Today’s knowledge workers lack equivalent mediating institutions:

Without leverage, the public is opting for diffuse and often ineffective outlets for their anger (social media cancellations, booing at graduations, data centre protests) — or stochastic violence, like the recent attacks on Sam Altman’s house.
Jasmine expects things to escalate, especially if 2028’s elections make AI a top voting issue. Political candidates are already preparing their AI action plans — with some invoking AI as the “perfect political bogeyman” for their own agendas.

Could all this distract from the biggest risks? The safety community is split on this question, and Jasmine sees a real risk of misallocated energy — but thinks broad politics could coexist with precise policymaking, especially since Mythos woke the government up to societal risks.

AI has a PR problem that marketing can’t solve

After a roundtable with AI executives, Jasmine concluded the industry is past the point where ads and AI-for-good campaigns can help:

Silicon Valley’s promised utopia — universal basic income, infinite leisure, and immortality — doesn’t appeal to the average American, and the longer they promise AI-powered cancer cures without delivering, the less credible they seem as actors.
For the past five years, frontier labs have opted for risk-heavy messaging to recruit top talent and investors. Those messages are still publicly on-record, so any reassurances counter to them are seen as “flip-flopping.”

What could work to rehabilitate public opinion on AI: visible, accessible medical and scientific breakthroughs (the highest-polling AI outcomes), and legible economic redistribution.

China’s embrace of AI is pragmatism, not optimism

Weeks of interviews in China convinced Jasmine the contrast with the US isn’t techno-optimism but “techno-determinist pragmatism”:

The CCP champions AI and protest is high-risk, so resistance feels hard to fathom.
But dissent against tech isn’t zero — delivery drivers have organised, and there was a ‘protest-lite’ against Baidu robotaxis in Wuhan — and this has nudged policymakers towards legislation.
Iterative regulation gives some people the security the US lacks: a Beijing court ruled AI replacement isn’t just cause for firing, and companion chatbots are aggressively regulated.
In a competitive white-collar job market, the logic is practical: if you don’t use AI, someone who does will take your spot.
Middle-aged and older generations have lived through both revolutions and dramatic rises in living standards, so they conclude: “progress moves and you better get on board or you’re going to fall behind.”

Speaking to every side is the secret to balanced AI reporting

Jasmine believes the mainstream media is still figuring out how to cover AI. She claims they’ve got “PTSD from the crypto era,” when promises to change the world didn’t really materialise, and scams and fraud abounded.

When she’s writing about AI capabilities, her editors almost always ask, “But isn’t it just marketing?” And this is becoming its own kind of bias.

Jasmine’s fix: talk to every possible tribe on an issue — not only companies, but also people in academia, policy, both political parties, and independent experts. When a lot of these entities agree on something — like cyber or bio risks — it’s probably not hype.

Institutions are waking up — demand for AI ideas now outstrips supply

The range of AI ideas taken seriously in policy and philanthropy is expanding quickly. Between Jasmine’s DC trips in February and March, policymakers went from seeing AI as “stochastic parrots” to scrambling because constituents are worried. The result:

“A lot more politicians looking for ideas of what to do about AI than we actually have compelling ideas and solutions.” Congressional staffers say their representatives want to act but are still learning how.
Philanthropic billions are coming — from soon-to-be-IPO-rich individual donors to the OpenAI Foundation, Gates, and Rockefeller — with more desire to spend than credible organisations to fund.

This means there’s tonnes of opportunity on AI issues right now — in policy, safety, societal impacts, and broad public issues — especially for people willing to build relationships with groups who are new to the topic.

Highlights

Frontier AI builders have no good advice for today’s teens

Jasmine Sun: I was doing these long-form AI ethnographies where I’d sit down with a random researcher at one of the frontier labs — sometimes a friend, sometimes an acquaintance, or a stranger.
And I just asked them for an hour about their very broad views about AI, its impact on the world, and how optimistic and pessimistic they were. All off the record.
One thing that I noticed is: one of the questions I ask is, “What advice would you give to a normal 17 year old? Not like a super cracked, brilliant coder — but just pick a random state in America, normal 17 year old, they’re getting Bs in school, they’re not thinking about AI that much. How do you think that they can prepare for the future?”
What was really shocking to me was almost everyone I spoke to, regardless of their personal political beliefs — some of these are more libertarian types, some of these are more leftist types, some of these don’t care about politics at all, some of these have a wide range of beliefs about even AI itself — most people just told me: “Honestly, I have no idea what I’d tell that 17 year old. It’s a really scary time. I don’t think there’s going to be a lot of jobs for them left. I think they’re caught in this painful transition.”
On the economic front, it did become clear that there was this quiet consensus that, regardless of your future beliefs about whether there will be a permanent underclass or more jobs than ever, the vast majority of people I know in the AI industry believe that in the near term, in the next few years, we are going to see mass job displacement. It’s going to be especially hard for a lot of young people, and a lot of current knowledge-work jobs are going to be disrupted and displaced completely.

Some AI insiders don’t care whether humans stay in charge

Jasmine Sun: I have been worried in hearing, in just casual conversation, fairly prominent and influential people within the AI community express an indifference or even a preference to a world where AIs are making most of the decisions in society rather than humans.
One of them, I won’t disclose people’s identities because they were Chatham House or off-the-record contacts. But I’d be having a conversation with somebody at a conference workshop and they would say, “Everyone at this conference is so worried about what’s going to happen to humans in the post-AGI world. But frankly I’m more worried: what if the monkeys are still running the show?”
And in this case the monkeys refers to the humans. He was more concerned that humans would still be in charge of governments, of corporations, of society because he had this view that AIs in their superintelligent existence were more deserving and more important to preserve a future world for.
Or I was talking to another very influential person in the AI scene and we were just on a walk and musing about: if all labour gets automated, what are people going to do? And he was like, “Maybe you’ll be like Nozick’s experience machine, and we’re all going to be wearing these VR headsets and watching perfectly personalised AI slop. So we don’t have any leverage in society, but in a way it’s very pleasurable and people are very happy.”
And I was like, “I don’t know. Do you think that’s actually the world we want?” We’re going back and forth on it a bit. And he’s like, “Honestly, Jasmine, I’m a lot more interested in what the AIs are going to do with each other a thousand years from now than I am in what’s going to happen to humans next year.”
So, in a way, I think that there are people — influential and prominent people in the AI community — who, whether for intellectual reasons (as in a belief that the more intelligent species is more deserving, or simply because it’s a more interesting thought experiment) have arrived at an indifference towards whether humans stay in control. I find that quite worrying.
I don’t think it’s the majority belief or common belief within the AI labs or something like that, but I do think that it makes it very hard to move forward policy action, to move forward the safety conversation — if we’re not able to agree on even very foundational concepts such as: humans should stay in control of our governments, and it is possible to decide what technologies we build and don’t build, and our conversations should centre what happens to people.

The consequences of making it uncool to be cautious

Jasmine Sun: When some of the more accelerationist types have turned ‘doomer’ into this slur almost against people who want to slow down or to regulate AI or other technologies, they’re presenting a group of people as being anti-progress — or as being in opposition to this core belief in Silicon Valley that technological advancement is the only thing that has ever brought us growth, prosperity, equality, et cetera.
If in all of history — in spite of the Industrial Revolution, automation, every wave of technology has brought on its own wave of disruptions — but if one were to believe so firmly in these apocalypse scenarios that you would be willing to slow down the only thing that might again bring us this future utopia that would keep the progress engine going, then you must be this kind of pessimistic, low agency, tech-sceptical kind of person, I suppose…
I think people are seeing AI in the context of all the previous technological revolutions and thinking: “Yeah, these doomers would have still be on the farms being peasants. They would never have allowed the Industrial Revolution to happen.”
Zershaaneh Qureshi: … How does that actually affect what’s going on in the labs?
Jasmine Sun: I would say again, to me, a huge part of this is the talent-flows question. Just because talent is so scarce, as we see from the salaries that people are willing to pay, talent is incredibly scarce. So when you have these cultural beliefs, the main thing that they influence is actually what young people decide to do with their lives — and how many people decide to, say, work on safety vs working on capabilities…
I think that the idea that trying to be careful — trying to be constrained about the way that you build technology — by labeling that as ‘doomery’ or low status, or as not being a pioneer and a serious builder, you’re basically preventing a bunch of very impressionable young people from going into a field that they otherwise would.
That’s to say nothing of obviously the policy conversation in DC, where ‘doomer’ has been employed as an extremely effective insult.

Ordinary people think Silicon Valley is gambling with their lives

Jasmine Sun: I think that people are correctly assessing that there are some folks in Silicon Valley who are taking a gamble on their futures.
Like you said, maybe the gamble is their jobs, maybe the gamble is their literal survival. Maybe the gamble is if they believe that AI is a bubble, it’s like a giant economic bubble. It’s true that AI is propping up American GDP growth, and it has been for the last year or so. I think St. Louis Fed said that 39% of growth in 2025 came from data centres and AI.
So when people understand that they have no capacity for democratic control or having a voice over the direction of the technology, when the leaders of those technologies are being very open about the risk that they think it’s going to pose and saying, “Hey, we’re going to do it anyway.”
For normal people, they actually have no leverage, right? They don’t know anyone who works in an AI lab. They can’t just tell them, “Hey, you should worry about this.” Most policy action has been extremely limited. Most people don’t know anyone in AI policy anyway.
I think that when I see things like data centre protests or booing at college graduations or whatever, I don’t necessarily think that they’re the smartest levers to make AI safer or whatever, but I think that they are the only levers that people feel that they have access to.
And yeah, maybe Waymo is a little bit safer. Maybe ChatGPT helps me write my emails faster. I think people are willing to cop to those particular things. But when ‘write my emails faster’ is positioned against ‘this might take my job and kill us all,’ it just doesn’t feel like that good of a trade.
And then the AI people will say, but it also might cure cancer. And they’ll say, “Where is the cancer cure?” I hear this all the time. The AI CEOs keep talking about curing all diseases. Where are the cures to the diseases? At a certain point, you’re just not a credible actor anymore. So I understand why the populist backlash is happening.
I would like to see it funneled towards maybe more what I see as productive policy aims than protesting data centres and booing people at graduations and whatever. But it kind of makes sense to me.

Factory workers had unions, knowledge workers have no levers

Jasmine Sun: In the 20th century, we had a big wave of factory automation where a lot of the factories went from primarily humans on factory lines to humans managing machines, more so. Part of the way that this was negotiated, such that there was not a massive backlash and there weren’t waves of violence by workers who were being automated, was that these factories were mostly unionised.
When they were mechanised, the people coming with the machines would meet with the union leaders, sit down at the bargaining table and say, “OK, first of all, we’re going to pitch this to you as it actually makes your job safer. You guys are really worried about workplace accidents, and this is going to reduce that. It gets at this issue that you guys are really worried about, we’re going to actually tie wage increases for workers to the increased productivity of these factories.”
Also you had just had this new social contract that FDR introduced to introduce better working legislation in general.
But more importantly, I think that there was a bargaining table and there were these mediating institutions of the unions between the people who were impacted by automation and the people introducing the automation. One really interesting thing I think about today is that there are no mediating institutions of that form. If you are a white collar worker in most industries, you are not part of any sort of organisation that is any sort of union or whatever.
So when your boss decides to maybe introduce AI, consider eliminating your job, there’s not actually anything that you can do about it. There’s not like a particular democratic or collective channel to express how you want AI to be incorporated into your life or your workplace.
I think that means that when people are really pissed off, they start going for social media cancellation and stochastic violence, and all these more diffuse forms of political backlash.
I think I just have a view that public sentiment will find an outlet. And right now we’re just going to see quite a lot of it from very different directions, at least until people feel credibly reassured that they will have the downsides cap, that they will share in the growth that AI gives.

Silicon Valley’s utopia is a hard sell

Jasmine Sun: A lot of folks who I know in San Francisco, in the AI research community, they’re really excited about UBI and they’re really excited about this infinite leisure point of view. When you test UBI with the public, it’s extremely unpopular because most Americans really want to work. They really want to have a job.
Maybe they don’t want the job to be 40 hours a week, maybe they don’t want the job to be this or that. But fundamentally, people really want to work. Or when you ask people, “Do you want to live forever?” I think it’s actually quite a controversial poll.
A lot of people don’t want to live forever. The utopia that Silicon Valley and the AI industry is outlining is not actually a very compelling utopia to a lot of the other people in society, and they don’t realise that because they are in these very insular communities — then even your attempts at positive storytelling don’t really land…
I frankly think that at this point, in order to get the public back on board, you’re going to need to start giving people real material benefits. Or maybe it does look like the long-awaited cancer cure. I think if people see medical breakthroughs that are actually happening as a result of AI, then they are going to have a much more positive view on things.
When you look at David Shor’s Blue Rose polling, health advances and scientific advances as a result of AI are some of the highest-polling things among the public. So if it also looks like more investment is shifting towards domains that people are really excited about AI being a part of — and those are visible, those are well marketed — then I think that could also really work.
But again, when we talk about scientific and health advances, it can’t be these advertisements about plausible future things. I think we actually have to show that we are able to access cheaper or better, or whatever it is, whatever cures, treatments, et cetera, as a direct result of the AI companies and their wealth.

What can poor countries sell in the AI age? by Deena Mousa

Deena Mousa — Thu, 23 Jul 2026 14:21:54 +0000

This post was crossposted from Under Development with the author’s permission, by the Forum team. The author may not see comments.

Subtitle: Looking for alternate export-led growth pathways

In my last post, I argued that AI threatens economic growth paths for low- and middle-income countries by endangering both manufacturing and service exports. If that’s right, the obvious question is ‘what should LMICs do instead?’

There’ve been many proposed alternatives. One way to think about these paths is that they are each a bet on a particular strategy that makes that industry competitive even in a world with highly capable AI. These generally fall into categories: selling some physical good that a country is endowed with (safaris), or it uniquely can produce (cocoa, geothermal energy); performing work that is symbiotic rather than competitive with AI (data labeling), or where AI is particularly weak (caretaking).

In order for them to be viable alternatives to manufacturing and service exports, they will need to not only earn revenues, and in hard currency, but also absorb workers at scale and include feedback loops that improve productivity and therefore wages in that sector over time.

The pattern across all of these paths is that, while they can bring in money, none of them can employ enough people in domestic work, doing a task where workers become more productive over time and demand can grow to absorb that productivity. As a result, some may soften the blow of AI automation, but none is likely to serve as a true alternative growth path.

Data labeling and verification

As AI potentially eliminates existing jobs by automating them, one strategy is to create new jobs to support AI itself. Models need enormous amounts of human input, and that input is both digital and wage-sensitive, so it flows to whoever’s work is cheapest. Data annotation hubs in Nairobi and Manila currently employ tens of thousands of people, and optimists see this as an opportunity for workers globally to capture some higher wages from the AI boom.

One form of this is data labeling, wherein workers help produce the labeled examples that models use to learn. This can look like drawing bounding boxes around pedestrians or traffic signs on videos to train self-driving cars, tagging violent posts to inform content filters, or ranking chatbot response quality to help models learn which answers people like. Another is verifying AI responses and catching failures in deployed models. In its final form, this looks like using humans to either supervise or work collaboratively with AI. For example, Presto Automation sold “AI” drive-through ordering to Carl’s Jr. and Checkers. They claimed 95% of orders needed no human input, but SEC filings revealed off-site agents in the Philippines were intervening in over 70% of interactions.

The problem is that this pathway is not durable. For one, a job that relies on no learned skills or sector-specific knowledge naturally invites underbidding. In Cagayan de Oro, per-task pay halved after 2022 as Remotasks moved projects to Kenya, Nigeria, and Venezuela in search of cheaper workers. On some projects, Filipino rates fell from $10 a task to under one cent. In March 2024, Remotasks shut down in Kenya, Nigeria, and Pakistan within a single month.

Work that involves supervising and validating AI output also naturally puts itself out of a job, since every human correction makes the model better. The demand that remains is migrating up-market, toward expert data from doctors, lawyers, and PhDs, which is to say toward workers rich countries have. In September 2025, xAI laid off some 500 generalist annotators, about a third of its data team, to hire specialist “tutors” instead.

Agriculture

One export where demand is stable and AI is unlikely to be able to produce it is food. Agriculture already employs half to two-thirds of workers in most low-income countries, and many of their climates are uniquely well-suited to producing popular exports — like cocoa, coffee beans, or mangos.

Through this path, countries could focus on moving beyond raw commodities into processing and branded products, capturing higher margins. For example, Kenya built a cut-flower industry worth over $800 million a year employing well over 100,000 people, and Côte d’Ivoire now grinds among the most cocoa of any country on earth.

A very interesting organization working on this is Exporters Without Borders. They’re currently backing founders to launch Madagascar’s first freeze-dried fruit industry. The island grows ~100,000 tons of lychee a year but exports only a fifth, and freeze-drying turns the rest into a $30/kg shelf-stable product.

This is an exciting intervention. That said, the two halves of the pathway (farming and processing) both have limitations. Farming employs enormous numbers of people, but when output per hour rises, the result is historically that labor leaves agriculture rather than more workers being hired given relatively capped demand for these goods. Processing is capital-intensive, employs few people, and is vulnerable to the same automation that manufacturing writ large is.

At best, the Madagascar plan produces a $5 million plant employing hundreds, which is not a replacement for manufacturing-led growth in a country of thirty million. Agriculture, even with downstream processing, cannot reliably employ a growing country at rising wages.

Tourism

Another thing countries have that AI cannot currently replicate is access to a sunny beach holiday, a historical site, a safari, or authentic pasta handmade by an Italian grandmother. Africa earned $42.6 billion in tourism receipts in 2024, 41% of its service exports, the highest share of any region. If AI were to automate many jobs, it ought to make some people much richer, and, since travel is a luxury good, demand for it would grow in that case.

This is all well and good, but not every country has natural attractions for foreign visitors. Some countries will be excluded from this path. In addition, the scale of the opportunity is not nearly enough to be a reliable growth path. All of Africa’s tourism receipts amount to roughly a seventh of what India earns from IT services alone. Gross receipts also overstate it given imported food, foreign-owned hotels, and foreign airlines claw back a large share. Estimates for small island economies put the leakage at more than half of every tourist dollar. As a result, the economies tourism has made rich are mostly microstates, like the Seychelles, or the Bahamas.

Productivity in hospitality also generally grows very slowly. Of the sectors BLS has tracked since 1990, accommodation and food services stands out as one of the very few where hours worked grew faster than productivity.

Although tourism seems to innovate constantly with things like helicopter safaris or underwater hotel rooms, these things increase revenue without necessarily raising the output per worker-hour. In most services, as the economist William Baumol put it in 1967, “labor is in itself the end product.” Luxury tourism actually advertises its staff-to-guest ratio, and more labor per unit of output is part of the appeal.

Although you can charge tourists more — Rwanda prices gorilla permits at $1,500 a day — this is a rent on a fixed endowment; there are only so many gorillas. You can charge tourists more, but you cannot serve ten times more tourists per worker over time, the way you can with other service exports. Where the industry has gotten more efficient, as with booking, pricing, and distribution, the gains did not accrue to the destination countries.

A related export might lie in ‘authentic’ creative goods. Things like handwoven textiles, indigenous art, and so on where the value lies beyond the literal item. This isn’t crazy. While quartz ‘should’ have killed Swiss watchmaking, “handmade” then became a luxury attribute, and Switzerland exports over $25 billion of timepieces a year. Something similar may lift up other handcrafted goods. However this is, by nature, not scalable as scarcity and time-intensiveness are markers of quality for these goods as they are for tourism.

Datacenters and compute

Another common proposal is for countries to provide a good that’ll scale in value with AI: compute. Selling compute, generally to American cloud giants like Microsoft or Amazon, practically means allowing data centers to be built on your soil. This entails providing land, reliable electricity, a fat fiber-optic connection, water for cooling, and enough political stability to keep a billion dollars of hardware safe.

These datacenters are warehouses of chips that the giants use and rent out — to AI labs training models, and to the crowd of businesses and individuals using them. The giants generally build datacenters where electricity is cheap and plentiful and, for everyday uses, close to the end user.

In the United States and many other rich countries, local opposition is building against data center construction. In a recent Gallup survey, 71% of Americans said they oppose having a data center near them, which is higher than for nuclear plants. The number of opposition groups more than doubled in the first quarter of this year; there are now 833 across 49 states. In that time, more than 75 US data center projects worth roughly $130 billion were blocked or delayed amid this opposition, more than fifteen times the quarterly pace of the two years before that. Many local governments have enacted bans, and just last week, New York launched the first statewide data center moratorium. This presents an opportunity for other countries to step in.

Compute is an appealing export because, as AI gets better, it becomes more valuable. Compute also converts things a poor country can’t otherwise sell abroad (like electricity, empty land, and sunshine) into a transportable good. It doesn’t rely on a large workforce with any particular education or skills, and, like other exports, it is purchased in USD or other hard currency.

But this hasn’t worked out well so far, because energy surpluses turn out to be temporary, since poor countries tend to consume their spare electricity as soon as they grow. Ethiopia, for example, legalized bitcoin mining in 2022 and began selling miners its surplus hydropower, earning about $220 million in its best year. But then its own demand increased due to new uses for electricity, which naturally tend to come about as countries get richer, and the grid couldn’t keep up with both. Within three years the country moved to wind down new and existing mining contracts.

For the most part, then, the giant campuses are likely to continue to be built where plentiful power meets deep capital and friendly politics. If US regulations begin to pose an issue, it’s more likely that firms will go to the Gulf than to Ghana.

That said, poor countries can host smaller facilities that can back up their own power. South Africa has roughly three-quarters of the continent’s capacity in spite of a coal grid so unreliable that operators built their own solar farms. They came anyway, since Johannesburg has fiber junctions and a financial sector, and Cape Town has six subsea cables on its beaches.

By that criteria, and for these smaller facilities, the best-positioned countries are South Africa, Kenya (given Mombasa’s cable landings), Nigeria (with eight subsea cables at Lagos), and Egypt (as the crossing point for most Europe–Asia internet traffic).

But this is, unfortunately, not a particularly promising export-led growth path.

Hosting small datacenters mostly isn’t an export in the first place. Over 80% of African data currently lives abroad, mostly in America and Europe — African firms pay foreign providers to host it, an import like any other; Nigeria alone spends an estimated $850 million a year on foreign cloud. Local data centers exist mostly to win that business back rather than to shift to being a net-exporter. For example, central banks increasingly require that financial data stay in-country, and apps respond faster from servers nearby, so the customers filling a Lagos facility are Nigerian banks, telcos, and government agencies. This is useful on its own terms but not the thing we’re looking for.

The jobs are also negligible; a hyperscale campus employs 100 to 200 people. The margins also tend to belong to the operator of the datacenter rather than the host country, which supplies the land, power, water, and often tax breaks on top of that.

Data centers are less like an export growth pathway, and more like infrastructure, like roads. Building roads isn’t really a way to get rich; you get rich on what travels over them — and what travels over these, for now, belongs to high-income countries.

One simpler proposed alternative to selling compute is selling the energy itself. Practically, this means building solar and wind farms in unoccupied deserts, and then either converting the power into green hydrogen or ammonia and shipping it by tanker, or sending the electricity to rich countries through very long undersea cables.

This is a strategy that’s intended to take advantage of a physical endowment. The sun over Africa is more than twice as strong as the sun over Germany, and the EU’s REPowerEU plan targets 10 million tons of imported green hydrogen by 2030.

But demand for green hydrogen is manufactured by regulation that forces shifts to clean energy, and is therefore inconsistent. So far regulation mostly hasn’t bitten; renewable hydrogen was about 0.3% of EU hydrogen consumption in 2023. And so, four years into Namibia’s $10 billion project intended to meet Europe’s stated demand, its cornerstone customer — RWE, one of Germany’s largest utilities — was still only signed to a non-binding memorandum, and ultimately walked away last September.

RWE’s stated reason was that European demand for hydrogen derivatives was developing slower than expected. The IEA’s latest count puts announced low-emissions hydrogen production for 2030 at 37 million tons a year, of which 4.2 million tons (about a ninth) comes from projects that have actually been financed or built. 2025 was also the first year the announced pipeline shrank as projects were cancelled or slipped past 2030.

There is demand for electricity by cable on the other hand, but rich countries are hesitant to import it. Xlinks, a proposed 4,000-km cable from Morocco to Britain, would have supplied 8% of the UK’s electricity at roughly half the price of power from its new nuclear plant. The British government refused to back it on the grounds that it wasn’t “home-grown power.”

There’s one exception, which is to stop trying to ship the energy and instead embed it in something. Economists call this powershoring: relocating energy-intensive industries to places where clean power is cheaper. In April 2025, HyIron’s Oshivela plant began producing green iron in the Namibian desert. Iron production is energy intensive, there is significant demand for it, and it crosses borders freely without as many of the hang-ups energy itself has.

A version of this strategy exists for exporting minerals too. Indonesia captured more of the value of its nickel by banning raw nickel exports in 2020 and forcing smelters onshore. Downstream nickel exports have since grown several times over, though with Chinese-owned smelters and environmental costs.

This is the most promising version of this strategy, and would also reopen the door to manufacturing. But the model is capital-intensive and hard to scale; Oshivela employs a workforce in the dozens where garment factories employed thousands. Cheap energy could bring the factories back but not the jobs they provided, which were critical to their role as an engine for growth.

Labor migration

Finally, one option is a bit unintuitive as an export at first: if the jobs won’t come to the workers, send the workers to the jobs. Practically, this means training people in professions where humans are likely to have some advantage over AI, like relational or physical jobs and helping them get placed abroad. The emigrant captures higher wages, and the country of origin earns remittances in hard currency.

The scale involved here currently significantly outstrips the other proposed options. Remittances to poor countries were about $654 billion in 2024, which is more than either foreign investment or foreign aid. They can reach up to 40% of some countries’ GDP, like Tonga and Tajikistan. They are also much more stable. When COVID hit, FDI to poor countries excluding China fell by more than 30%, but recorded remittances dipped just 1.6%. And while FDI to LMICs has fallen by roughly a third since its 2021 peak, remittances have climbed every year.

Clemens, Montenegro, and Pritchett compared workers who are observably identical in terms of age, sex, and years of home-country schooling on either side of the US border. Even as lower bounds, the same worker earns roughly 4 times more in the US for the median country, and over 16 times more for Yemen. If AI makes rich countries richer while that embodied work stays human, that premium will only grow.^[1]

Today, the rich world is aging into a permanent care shortage; the WHO projects the world will need 4 million more nurses and midwives by 2030, and a quarter of practicing US physicians already trained abroad. US government employment projections show that home health and personal care aides are expected to add 740,000 jobs, the most of any of the 832 occupations the agency tracks, and nearly three times the runner-up. Of the thirteen occupations projected to add more than 100,000 jobs, eight are highly physical.

The more successful AI is at automating labor, the richer Americans are likely to get by indirectly capturing the surplus value of AI labs based in those countries. As they get richer, they will naturally demand more and more of these services. You can see this even on a small scale: our demand for childcare, for example, is very far from being satisfied, and many American parents would spend a lot of any extra income they got on more nanny hours. These are already migrant jobs: a third of America’s home care workforce is foreign-born, rising to nearly 40% among home health aides.

There are, however, two issues with this. The first is that, of course, migration pathways depend on the receiving country’s political appetite, which is volatile in a way the demand for labor is not. Britain, for example, opened its skilled-worker route to care workers in 2022 and recruited over 220,000 workers from abroad. However, they were banned from bringing dependents in 2024 and the route to new overseas applicants shut entirely in mid-2025. All the while, the sector was still reporting over 100,000 vacancies. If other work is automated, it’s also possible that more people in other jobs will prefer to switch over to care work, obviating the need for migrant labor.

The other is that migration mostly enriches individual families and has limited evidence as a growth engine. The Philippines has run the world’s largest labor-export program since 1974, and remittances bring in about $40 billion a year (roughly the same as its entire BPO sector earns), but with no corresponding industrial takeoff. This is because remittances are spent mostly on consumption and housing rather than invested in firms, and because large inflows appreciate the currency in ways that squeeze other exports.

There is, however, opportunity for growth in cases where those who move abroad return home. For example, Taiwan’s chip industry was seeded by engineers coming home from Silicon Valley, and India’s IT sector leaned on its American diaspora. All that said, as a means of capturing some of the value of a prospective AI boom, labor migration strikes me as one of the most promising potential avenues for LMICs.

Is there another way to grow?

Historically, export-led growth did three things at once: it brought in hard currency, employed huge numbers of people, and made workers and firms more productive by doing the work itself.

Every proposed alternative fails at least one of these tests. Data labeling earns hard currency, but the work is short-term and the demand that survives is moving toward experts in rich countries. Agriculture employs enormous numbers of people, but that number shrinks as productivity goes up, demand for food is capped, and agricultural processing employs few. Tourism is closed to countries without attractions and capped in the rest in terms of productivity per worker and scale. Data centers that are feasible in many low- and middle-income countries mostly aren’t exports in the first place. Energy exports depend on demand manufactured by regulation, which has mostly failed to materialize.

That leaves labor migration. Making some bet on care work and physical jobs in high-income countries, and earning remittances in return, seems to me like the most promising avenue for many countries. It is also a depressing takeaway, because it gives up on domestic engines of growth and settles instead for trying to buy a stake in growth that happens elsewhere. I hope it’s wrong.

^
This is a non-trivial ‘if’, but it seems likely that relational care work (e.g., nursing, childcare, elderly care) will be the last of anything to be fully automated because (a) it’s harder for AI to do since it is physical, variable, and reliant on emotional attunement (b) even if we got an AI nanny-robot that was superior to human babysitters there will likely be some lag in acceptability. Also, if we’re in a world where all that work is automated as well, things have gotten crazy enough that most current discussions about the concept of “GDP growth” don’t make a lot of sense anymore.

Reverse alignment by Tom Vargas

Tom Vargas — Wed, 22 Jul 2026 23:38:06 +0000

Glen Weyl and friends have recently called attention to a problem they call reverse alignment. While they think the AI alignment issue is important, they argue that most people in the space neglect thinking about the institutions that would interface with advanced AI capabilities (especially if aligned). In brief, something like wrong institutions + aligned, advanced AI = likely worse outcomes. This seems plausible, and it intersects with some emerging EA interest in things like democracy, though the argument Weyl et al. develop is less about specific political outcomes and more about the foundational social and political institutions that govern how we behave, decide, and ultimately harness (or not) the AI wave to our benefit. So, more longtermist.

Some of this feels familiar. For example, I’m having a hard time entirely divorcing reverse alignment from efforts to, say, do fieldbuilding around AI and health (meaning creating the conditions so that deployers of health solutions can leverage AI). In both cases, we are thinking about creating conditions that optimize for deployers of AI. The difference, I think, is the level of analysis, where reverse alignment asks us to think at the ‘building blocks of society’ level. If true, this makes it both 1) quite difficult to execute/get right and 2) exponentially more transformative in expectation. More smart people should be working on this.

The reverse alignment coalition: https://reversealignment.ai

New international campaign exposes the hidden costs of the global shrimp industry by Beatriz Cossermelli (AWO)

Beatriz Cossermelli (AWO) — Wed, 22 Jul 2026 21:27:14 +0000

We are delighted to share that Animal Welfare Observatory (recommended by Animal Charity Evaluators) has just launched our first international shrimp welfare campaign.

After months of research and international collaboration, the campaign exposes the hidden environmental, social, and animal welfare costs behind Spain’s shrimp supply chain through investigations in Ecuador, the EU’s main supplier, and Indonesia. The campaign includes an environmental and social report on Ecuador, prepared by Foodrise with support from Eurogroup for Animals and Protección Animal Ecuador, as well as undercover investigations carried out with We Animals, with visual documentation by photographers Marcos Zegers and Ulet Ifansasti.

Spain is the largest shrimp consumer in Europe and the fourth largest consumer globally, making it a strategically important market for advancing shrimp welfare.

At a time when climate and sustainability dominate public debate, we deliberately designed this campaign to reach audiences beyond the animal protection movement by connecting animal welfare with environmental destruction and social impacts.

We are very encouraged by the initial response. In its first days, the campaign has already been covered by National Geographic Spain, Infobae, El Independiente, El Salto, and other national media outlets. National Geographic Spain also featured the investigation on its social media channels, helping the campaign reach audiences well beyond our usual circles.

Key findings

Shrimp are the most farmed and slaughtered animals in the world, surpassing chickens and pigs combined by billions every year, yet their welfare remains almost entirely absent from public debate.
In Ecuador, where shrimp has become the country’s largest export sector, surpassing even crude oil, the industry’s rapid expansion has destroyed 57% of the country’s mangrove forests, rising to as much as 90% in some regions.
The investigation also documents land rights violations affecting coastal communities, the concentration of wealth and power in the hands of a few large producers while 63% of aquaculture workers remain informally employed, and a global trade model described by the UN Special Rapporteur on the Right to Food as “food colonialism.”
From an animal welfare perspective, undercover investigations documented the slaughter of shrimp without effective stunning in Ecuador and the routine eyestalk ablation of female shrimp without anaesthesia in Indonesia, despite growing scientific evidence that crustaceans are sentient and the availability of more humane alternatives.

Report

Cheap Shrimp, Hidden Costs: The Environmental and Social Impacts of Shrimp Farming in Ecuador

Additional campaign resources

This campaign marks the beginning of a broader international strategy to improve shrimp welfare. We will build on these findings to secure corporate commitments from key seafood importers and other companies across the European supply chain, with the goal of advancing effective stunning and ending eyestalk ablation.

We hope these resources are useful to others working on shrimp welfare, food systems, environmental issues, corporate campaigns, and other effective altruism cause areas, and we welcome any feedback or questions.

Curating a list of charities for our 30-day daily-giving journey—which charities are we missing? by Joseph Sarvary

Joseph Sarvary — Wed, 22 Jul 2026 20:58:31 +0000

TLDR: We’re a giving app that is buiding a 30 day journey in which users give to 30 different pre-selected highly effective charities. We want feedback on our initial list and ideas for additional charities we should consider.

Hi everyone,

I’m Joe, founder and Chair of EveryDrop SCIO. I’ve spent the last year and a half building a daily-giving app, and early last month we were finally awarded official charity status by the Scottish Charity Regulator (OSCR).

Who we are

EveryDrop was founded on the concept that giving effectively can be a source of joy and fulfilment for everyone. Through the app, our users are presented with a daily learning journey which introduces them to the scale & importance of a new impact area and the incredible impact and progress being made by the day’s spotlighted charity. If you’re curious, you can see a 2-day demo of the platform here: https://everydrop-app-demo.pages.dev/ or check out our website here to join the waitlist: www.everydropapp.com

For our launch, we have decided to frame the app as a 30-day challenge.

Why we’ve taken this approach

Our app is built to integrate education, effective giving and positive psychology to build a habit of giving in everyone. We believe that daily giving can:

Do the giver good. Giving to others lifts our wellbeing more than spending on ourselves, it restores a sense of agency to people who feel powerless, and renews our faith in the ingenuity and dedication humanity brings to its hardest problems.
Generate real, transformational impact, too. Guiding donations, even small ones, to the world’s most effective charities generates real impact.
Resets the perception of what £1 is worth. When offered a daily reminder of the tremendous impact of the world’s most effective charities, a pound stops being a snack or coffee upgrade and becomes a course of vitamin A or a day of clean water.
Open a door to more. Daily giving challenge can serve as a gentle on-ramp to effective altruism’s bigger asks (e.g. the 10% pledge, high-impact careers) and an easy thing for committed givers to share with friends and family to share the positive aspects of effective giving.
Make giving shared, not solitary. The app will serve as a community of givers, with features that track collective giving to demonstrate that plenty of others care enough to give, too.

The risks we’ve mitigated (or at least tried to)

Inefficiency of small payments. We haven’t built EveryDrop to address donation efficiency. This obviously matters a lot, and there are already great organisations across the EA universe optimizing the delivery of funds to the impact areas that need it most.

We see ourselves as filling a different need. Our main objective is to help our users discover the feeling I had when I first learned about and engaged in effective giving:

that we mustn’t give in to the belief that the world is falling apart and that there’s nothing we can do about it;
that behind most bleak headlines are a bunch of people making real, measurable progress, persevering despite the structural barriers;
and, importantly, that there is a real, immediate, and effective way to contribute.

Of course, I wouldn’t be here if I thought effectiveness and efficiency didn’t matter. Effectiveness remains the overarching criterion we have selected our featured charities on. But the spark I’m chasing with our users is agency and empowerment: the sense that a person can do something to make the world better, stronger, kinder, safer. And in doing that action, they are joining a wider community of committed givers.

The micro-donation structure comes with an obvious trade-off on payment-processing fees. We’re working hard to keep the operation as efficient as possible, and we’ve already received an initial commitment from a private funder to cover transaction fees on the first £10,000 moved through the platform.

We’re exploring several ways to minimize processing fees further: a shift from daily settlement to a pledge-then-settle model, where in-app payments are aggregated weekly; integration with PayPal Giving Fund infrastructure, which can substantially reduce UK processing costs; and a Direct Debit / bank-transfer (Faster Payments) set-up to move money at little cost.

While the ultimate payment method is still under consideration, we’re confident that even in the worst case—where EveryDrop covers our users’ processing fees on their behalf—it remains worthwhile: spending roughly £0.17 in fees to activate £1.20 of net-new donations to effective charities is about a 5x impact lever rather than just donating our own money directly to the charity.

Moral licensing. Studies show that acts of generosity increase the probability of future generous acts. We keep the daily amounts small precisely to avoid a sense of “having done enough”. And the capped 30-day journey gives us a natural moment, at the conclusion, to present further action rather than closure. As the journey ends, the invitation is to go and do more, through GWWC or other long-term giving and community resources; resources from the School for Moral Ambition, 80,000 hours, and Probably Good are also considered for inclusion.

Crowding-out. “Crowding-out” here I meant that this habit displaces more, or better, giving that the user would otherwise have done were it not for this activity. Because we act as a pure conduit to vetted, effective charities and because our users are mostly people who weren’t giving effectively, or at all, beforehand, we feel the risk is low, and our net impact will be toward more and better giving, not away from it.

Our principal target audience

New givers (by far our main targets). The target user is Gen-Z, mobile-first, at university or early in their career. They feel the weight of the world’s problems, care about positive outcomes and they’re looking for a way to actually do something about it, but they haven’t yet found a foothold in effective giving or other morally ambitious output. They’re digitally connected and influenced by what their peers do.

Large donors tied to passive, DAF-like structures who don’t feel connected through the year to the causes they support. For them the donation itself is trivial, but the daily habit affirms their identity as a giver and can broaden their understanding of the different challenges we face.

Founders Pledge members who’ve committed to give big later in life but, in the meantime, don’t have an outlet to stay connected to effective giving.

Effective altruists looking to share their priorities and passion in an accessible way with family and friends. EveryDropApp can serve as a low-cost, low-commitment conversation starter, rather than a competitor for long-term committed action.

Main Objective of this Post:

The objective of this post is to solicit feedback on the charities we have selected to feature on our 30-day journey. The utility of every pound we move depends entirely on which charities we choose to feature. That’s where we need your help. While we want to prioritize the effectiveness of every dollar, we also want to expose our users to a full range of inspirational charities that are doing tremendous work. Instead of trying to optimize every dollar, we are trying to curate a list that has the highest probability of having an ‘Aha!’ moment for users that discover a cause area and charity that awakens a sense of agency, urgency and action in our user. Our main ask in this post is to help us refine and expand our initial curated list.

Charity Selection

One thing worth flagging: we’re a UK-registered charity, so we can only fund charities that fall within a specific set of pre-determined charitable purposes — poverty, health, education, environment, animal welfare, human rights, and community development. Which means, frustratingly for this audience, that most longtermist and AI-safety giving is off-limits to us (we actually had to drop planned days on both). So anything you suggest needs to fit within those seven areas.

Here’s the current line-up:

Poverty

GiveDirectly — Unconditional cash transfers straight to people in extreme poverty
One Acre Fund — Seed, finance and training for smallholder African farmers
BRAC — Asset + coaching “graduation approach” to get world’s poorest out of ultra-poverty

Health

Against Malaria Foundation — Insecticide-treated bednets to prevent malaria
Helen Keller Intl — Vitamin A supplements preventing child blindness and death
Deworm the World (Evidence Action) — Mass school-based deworming
Dispensers for Safe Water (Evidence Action) — Chlorine dispensers at rural water points
WaterAid — Clean water, toilets and hygiene via local systems
New Incentives — Small cash incentives for infant vaccination
mothers2mothers — Mothers with HIV mentoring other mothers
Days for Girls — Reusable menstrual kits and health education
StrongMinds — Group talk-therapy for depression in Africa
Sightsavers — Cataract surgery, eye care and disability inclusion
Fistula Foundation — Free surgery to repair childbirth injuries
CEPI — Racing to develop vaccines against future epidemics (pandemic preparedness — the global-catastrophic-risk edge of the health bucket)
Cochrane — Systematic reviews of medical studies that tell clinicians what works

Education

Camfed — Funding girls through school across sub-Saharan Africa
Lively Minds — Training mothers to run play-based preschools
Mary’s Meals — A daily school meal that brings children to class
Pratham — “Teaching at the Right Level” foundational literacy and numeracy
Worldreader — Digital books and reading on low-cost phones

Environment

Clean Air Task Force — Policy advocacy on the biggest climate levers
Eden Reforestation Projects — Paying locals to plant and protect forests
SolarAid — Affordable solar lights for off-grid homes (energy access — spans environment, poverty and health; we file it under environment)
5 Gyres Institute — Research and advocacy against ocean plastic

Animal welfare

The Humane League — Corporate campaigns ending the worst factory-farming practices
Compassion in World Farming — Campaigns and legal advocacy to end factory farming
The Good Food Institute — Advancing alternative proteins to replace industrial animal farming

Human rights

Anti-Slavery International — Advocacy and partnerships to end modern slavery
Justice Defenders — Legal training for prisoners and prison staff in Africa

Community / effective giving

Giving What We Can — Championing effective giving and the giving pledge (our Day 30 conclusion story)

What are we missing? I’d love to hear which charities you find most compelling, the ones that made you think “more people should know about this”. I’ve got a soft spot for organisations that would make a brand-new giver feel that spark of what’s possible — Justice Defenders is my touchstone there — not only the ones that top a cost-effectiveness table.

Thank you.

The Expansion of Factory Farming in Africa by Aurelia Adhiambo

Aurelia Adhiambo — Wed, 22 Jul 2026 16:16:31 +0000

I have cross-posted this article from RP Strategic Animal Insights, based on our presentation for the Rethink Priorities Strategic Animal Webinar Series, which happened in May 2025.

The piece covers:

Why “Africa” isn’t one story.
Species mix varies by region.
What is working: A double-movement approach.
The main bottlenecks in the region.
Why now? Countries like Ethiopia still have a real opportunity to put in place standards and incentives before industrialization takes hold, but that window is shutting faster in some locations than others.

I am happy to answer questions in the comment section. Thank you.

Introducing Impact List: a ranking of philanthropists by expected lives saved by Elliot Olds

Elliot Olds — Wed, 22 Jul 2026 14:04:51 +0000

TL;DR: I’m releasing a website that ranks philanthropists according to EA principles and research, and allows users to re-rank the list using their own assumptions. I’d like feedback and help making it better. I’d especially like ideas for how to make the results more trustworthy. Funding may be available.

Crossposted to LessWrong.

I recently built Impact List (impactlist.xyz), a site which ranks people by their positive impact via donations.

The goal is to make the list popular enough that people care about their ranking on it, so that it influences their decisions about where and how much to donate. A secondary goal is influencing people (whether or not they appear on the list) by making them more aware of the large differences in the cost to save a life depending on where money is donated.

The site uses QALYs as the common currency of value, but this doesn’t mean it’s limited to considering only effects on health/lifespan. The goal is to consider effects of any type and convert them to human-QALY-equivalents.

I wrote more about the motivation and theory of this project four years ago in this post.

What Impact List is

The core of the site is a ranking of (currently 73) wealthy philanthropists by their expected impact via donations.

The main list, using the default assumptions

Each donor has their own page showing all their donations (along with the impact of each donation), and aggregate stats by cause area. Users can toggle between ‘donations’ and ‘lives saved’ to visualize differences in effectiveness between cause areas. The internal QALY metric is converted into lives saved by defining one life = 80 QALYs.

Expected lives saved by Dustin Moskovitz, by cause area

I’ve split all donations into 28 cause areas, each with their own ‘cost to save a life’ value.

Eight of the 28 total cause areas, with the option to edit the cost to save a life for each

There’s also a page for every recipient (charitable organization), which can have its own ‘cost per life’ value if it’s more or less effective than the cause area average. Each cause area and customized recipient has a page with a detailed justification for how the math was done to arrive at the cost per life, and where assumptions are explicitly listed.

Three of the nine assumptions for the Global Health cause area

The estimates are of course very approximate, and will be controversial. I make heavy use of LLMs to do this research, encouraging them to synthesize existing (often EA) analysis. Users who disagree with any of the default parameter values can edit them.

Editing the three parameters for the Global Health cause area

Users can also edit global parameters specifying their time horizon, assumptions about population growth, and the discount rate.

After entering their own assumptions, users can create shareable links. This allows others to see what Impact List would look like given person X’s worldview.

I’d eventually like to add worldviews from notable researchers or organizations, but I’m starting with just a handful of worldviews based on simple tweaks to the defaults:

Users can select between suggested worldviews, or create their own

There’s also a calculator feature where people can enter their past donations and/or donations they’re considering and see where they’d rank on the list, given those donations.

What I’d like help with

If you see something below that you want to help with, comment on this post, DM me, or join the Discord.

Improving the process of effectiveness estimation

Few people are going to dive into the details of the effectiveness estimates and verify for themselves that the research is high quality. So it’s not enough that the estimates be excellent. People need to be able to easily understand and trust the process that leads to the estimates.

The current process is LLM-driven, opaque (users can’t see the prompts that lead to the estimates and they can’t see how feedback is processed), and dependent on the judgments of someone (me) without any reputation as a researcher. I’ve tried to partly compensate for this by making the justifications explicit about which assumptions are being made, and tried to make the reasoning as clear as possible, but I expect it’ll be very hard to generate enough trust using the existing process.

A few things that could help:

Incorporating something like X’s Community Notes on each individual assumption^[1], or otherwise making user feedback more of a first-class thing.
Getting people/orgs who have already built up trust to publish their worldviews, and doing some blending between them.
Making the estimation process more auditable, including giving details on currently-behind-the-scenes LLM interactions.
Coming up with some council-of-LLMs framework that minimizes special privilege given to any person or group’s judgment. Even if LLMs aren’t good enough for people to trust something like this now, maybe they will be in N months.

If you have ideas in this area, please share in the comments. I see this as the most important obstacle to the site becoming popular.

Improving the individual effectiveness estimates

The effectiveness estimates for all cause areas are mostly based on research from LLMs (frontier models from Anthropic and OpenAI). I’ve also done some manual review and back and forth with the LLMs to refine the estimates. The LLMs rely a lot on existing EA research, but the quality of the analysis is pretty uneven and could probably be improved a lot by putting expert human researchers in the loop.

I’d like these estimates to become the best place to look for a synthesis of all effectiveness research that EAs have done (or found).

If you have expertise in cause area effectiveness research, and especially if you’re an expert in some particular cause area, it would be great if you’d be willing to help improve the estimates.

Making the site more usable and beautiful

I think the site looks OK now, but not great (especially on mobile). If you have UI skills/taste and want to make the site look better, let me know.

Getting notable people and organizations to publish their own worldviews

As mentioned above, the site allows anyone to make their own customized set of assumptions and create a shareable link to the rankings using these assumptions.

For notable people/orgs, I’d like to add these links as curated options for all users, so people could see Impact List according to Carl Shulman, Rethink Priorities, etc.

If you’re a notable researcher or work in a well-respected EA org, and you also want this to happen, please get in touch.

Go-to-market advice

How should we actually make the site popular? In my post introducing this project I talked about some ideas for how to do this, but I’m pretty uncertain about the right path here.

My sense is that the quality of the site should be significantly higher before I try to popularize this beyond EA/rationalist audiences, but it’d be nice to have a distribution plan soon to guide the other site-improvement decisions.

Suggestions for high-impact people currently not on the list

Everyone currently on the list is well known, but that’s not a requirement. Feel free to nominate yourself (if your donations can be proven to the satisfaction of a skeptical reader) or someone else I missed.

Improvements to the underlying data model and estimation framework

See Appendix A below for details on how this works. I’ve tried to strike a balance between simplicity and expressive power, but I’m not sure I’ve picked the right tradeoff or whether I’m on the Pareto frontier.

Funding may be available

A couple of funders have offered to give me grants for this work. I haven’t accepted any yet, but if you want to work on Impact List and you want to be paid, these grantors may be willing to make that happen.

Getting involved

The project has a Discord. You can also submit pull requests via GitHub. Feel free to DM me on this forum.

You can keep up with the project on Twitter.

Above I’ve listed what I think I most need, but I may be wrong. I’m happy for all sorts of help or feedback.

Thanks to Austin Chen, Ryan Kidd, Kim Korte, Nathan Young, plex, and Laszlo Treszkai for discussion and feedback about the site.

Appendix A: How effectiveness is modeled

The details are on GitHub. This is a simplified overview.

Each cause area or charitable organization has one or more ‘effects’, which define how money translates into QALYs. There are two types of effects, standard effects and population effects, and each effect has several parameters.

Standard Effects

The most intuitive type. The more money you spend, the more impact you get. The parameters are:

Cost per QALY: How much money it costs to produce one QALY
Start time: Years after donation before the effect begins
Duration: How many years the effect lasts (benefits are assumed to be evenly distributed)

Population Effects

Effects where money donated changes the probability of some event. This is used to model x-risks, pandemics, etc. The parameters are:

Cost per microprobability: How much it costs to change the likelihood of an event by 1 in 1,000,000.
Population fraction affected: What fraction of the population is impacted if the event occurs
Welfare change per person per year: estimated welfare change per affected person per year, in QALYs. Positive values mean the funded work makes a bad event less likely or a good event more likely; negative values mean it does the opposite.
Start time / Duration: Same meaning as for the standard effects

Overrides and multipliers

A recipient organization which is part of some cause area can override the parameters of that cause area, to indicate that the recipient is especially effective or ineffective.

Recipient organizations can also have their parameters defined as a multiple of a parameter of the cause area it belongs to. The site’s UI currently hides the ability to express parameters in terms of multiples for simplicity, though I’d like to re-add this to the UI at some point.

Time

The model is capable of handling time-based effects, which allows specifying different levels of effectiveness for a cause/recipient depending on when the donation was made. For instance a donation to MIRI today may have a very different impact than a donation to MIRI in 2012. The current version of the site doesn’t make use of time-based effects, but they’re available and documented on GitHub.

Motivation for these choices

This model gives each effect a ‘shape’ over time that we can integrate over to get the total value. This allows assumptions about the future population, discount rate, and time horizon to be separated out from cause-specific assumptions.

I could have made these shapes more complex by having options for exponential or linear decay of effects, but I think it’s not worth the extra complexity. Every effect-shape in this model is a simple rectangle.

Appendix B: Current limitations and uncertainties

I currently only consider impact via donations. I’d eventually like to include all of a person’s positive and negative externalities, including via businesses they’ve created, via politics, etc.
The site’s estimates are very uncertain, and this could probably be quantified and communicated better to non-EAs, possibly along with some explanation of why uncertain estimates are still valuable.
As seen in Appendix A, the underlying model is fairly simple.
Although the site infrastructure supports it, I don’t currently distinguish average effectiveness from marginal effectiveness or account for effectiveness changes over time.
I don’t try to measure fungibility or crowding-out/in effects.
I use LLMs to gather donation data for each person using this skill, and it seems to work fairly well, but not perfectly.
Not all donations are public, so the data is missing some donations. However because it’s usually in the interest of both the donor and the recipient to publicize donations (the donor gets status and to raise the profile of a cause they care about, the recipient gets attention from other potential donors), my guess is that this effect isn’t huge.
I don’t count pledges, but I fully count donations that a person makes to a charitable foundation or fund that they control. In some cases this seems like what we want, but in others it probably overstates the donor’s impact. This could be addressed by giving these foundations customized lower effectiveness estimates, but I haven’t done that yet.
The cause area breakdown for multi-cause recipients like The Gates Foundation is based on estimates from LLMs, not on classifying all of the grants that they’ve made.
I’m unsure whether the 28 cause areas that I’ve picked are the best way to carve up the space.
Inflation isn’t yet adjusted for.

^
Thanks to Nathan Young for the suggestion and for discussions around this idea.

Two Coefficient Givings beat one twice as big by Jack Lewars

Jack Lewars — Wed, 22 Jul 2026 13:26:53 +0000

Why the AI windfall should build new grantmakers, not just bigger ones

Originally published at my Funding Anthropalypse Substack.

Imagine you could live in one of two worlds in 2027.

In the first, Coefficient Giving is twice as big.

In the second, there are two Coefficient Givings of the same size.

Which seems better to you?

Having considered it for a while, I would choose the second world. Put plainly, two Coefficient Givings beat one twice as big. I think the potential $37bn+ AI windfall is a once-in-a-lifetime opportunity to achieve this, with significant long-term benefits.

Is effective giving too concentrated?

In my view, yes.

Coefficient Giving (and to a great extent GiveWell) have enormous influence and power within the ecosystem, by virtue of their size and track records. Coefficient Giving is also the predominant or only funder in some of its domains, such as funding the effective giving ecosystem and farm animal welfare.

The situation seems analogous to a monopsony in economics:

monopsony, in economic theory, is a market situation in which there is only one buyer… Although cases of pure monopsony are rare, monopsonistic elements are found wherever there are many sellers and few purchasers. (Encyclopedia Britannica)

You could argue that there is in fact a diversity of funding in some cost-effective cause areas. For example, in AI safety, there is Longview Philanthropy, the Survival and Flourishing Fund, Astralis, the AI Safety Tactical Opportunities Fund etc.

In global health, there is the Gates Foundation (granting roughly 20x more than GiveWell and 8.5x more than Coefficient Giving each year); the remaining bilateral and multilateral funders; and a large number of smaller foundations (Mulago, CRI, WAM Foundation, DRK etc.). You could therefore also claim that the global health landscape is sufficiently diverse.

However, Coefficient Giving outsources a substantial amount of its global health grantmaking to GiveWell—around $1bn to date, and $175m for 2026 so far. It’s therefore not really accurate to describe GiveWell as a counterweight to cG—it could even be considered an extension of it.

More to the point, I think the true test of over-concentration is not just the number of grantmakers in a field—it’s also their size and, more importantly, the amount grantees rely on them. In AI safety, for example, Coefficient Giving gives substantially more, and in almost all cases at least an order of magnitude more, than any other actor. In combating factory farming, it is widely estimated to be the largest funder globally, and in supporting the effective giving ecosystem it is essentially the only funder of note.

Across all cause areas, it is often true that there is no other funder who could step in and become a grantee’s majority funder if the main one walked away.

This, to me, is the true test of over-concentration, and it seems true of both Coefficient Giving and GiveWell, for a large number of their grantees. It is also not just theoretical—it is something that Coefficient Giving’s grantees themselves are concerned about.

In short, Coefficient Giving and GiveWell are exceptionally important to the majority of their grantees, in ways that cause significant risks to the grantees, the ecosystem and the problems we are trying to solve.

What’s wrong with one dominant funder?

To some extent, the answer to this is just ‘read about power concentration’. However, I observe at least five major issues with the current situation.

The first is how dependent grantees are on one organisation for their financial health. The decision of a single program officer, formed under uncertainty, can virtually dictate the grantee’s survival. This is unduly risky for the grantee and also has unwanted second order consequences. For example, I have literally been in strategy meetings where organisations have asked “what will Coefficient Giving think of that?”

The second is that the blind spots of the organisation and individual grantmakers can become significant failure points. Even the very best program officers (and Coefficient has many outstanding ones) are also human, with biases, limited hours in the day and imperfect judgement. Especially in areas of high uncertainty, it is incredibly helpful to have grantmakers who can step in if another can’t or won’t. And, while dominant organisations may try to hedge internally against their own missteps, I think it is much easier to do this when the decision-makers are separated structurally. That is, several smart people thinking independently about an issue in different organisations are more likely to hedge effectively than the same smart people in the same team or office.

(As an aside, an ecosystem with a diverse set of large funders is also more attractive to top talent, versus one where there is a ‘party line’ and power is concentrated in a single funder. This will likely be relevant if we want to attract a lot more people to work on causes we care about and increase our ability to deploy money well in these fields, something I intend to post about in the future.)

Third, a single dominant funder is also a single point of failure. We are currently all-in on Coefficient Giving and GiveWell’s reputations and assets. We already saw how incredibly damaging it was when the FTX Foundation collapsed and it was only a relatively small fraction of the ecosystem at the time. I am in no way suggesting any malpractice at either organisation—but I am uncomfortable with relying so completely on their continued good judgement and lack of scandal.

Fourth, it is bad for grantees if a single funder tries to do the majority of grantmaking in a given field, because they are very unlikely to be well-suited to every type of grant. Trying to assess $100m grants alongside $100k seed grants is almost never the right option, because the decision timelines, risk tolerance, due diligence and opportunity costs are completely different at different scales.

At the moment, we expect GiveWell to fund seed grants, mid-stage grants, USAID-collapse-response grants and things at massive scale—but this leads to numerous issues, such as small organisations reporting substantial delays in being evaluated, when they need a rapid review for a grant before their runway expires.

You can instead imagine a pipeline of seed funders and incubators, followed by mid-stage funders who can underwrite growth, before GiveWell funds the best things at serious scale. And this is without even considering that some grantees will have constraints that stop them taking money even from a funder perfectly suited to their stage of life, for example if they want money that is politically-aligned with their other supporters.

Finally, monopsonies also have a dominant function in shaping the market in which they operate. This creates a dominant view of what gets evaluated in the first place (cause areas and organisations), and smaller organisations usually follow their assessments. Unless we assume that their judgement is perfect, and that big organisations can adapt at speed when the situation evolves, this is sub-optimal.

Can GiveWell and Coefficient Giving absorb the windfall?

The two most dominant organisations in the space are making concerted efforts to grow quickly. Both are hiring fast and both are specifically focusing on high-absorption options to deploy more money quickly. They are also outstanding grantmakers and some of the only places with track records of handling hundreds of millions in grants. You can see the attraction for time-poor donors, looking for a safe place to give.

However, the scale of the Funding Anthropalypse is likely to exceed their capacity, even if that capacity increases significantly. Last year, GiveWell regranted $418m, and Coefficient Giving over $1bn. Even with the most aggressive scale-up plan imaginable, it seems unlikely to me that either organisation can even 5x its grantmaking in the next 24 months, and we may be talking about sums far in excess of the $7bn in grants that this would imply ($2bn for GiveWell and $5bn for cG). As an incredibly rough rule of thumb, I am sceptical of any large organisation claiming it can more than double year-on-year, which would imply a ceiling of $2.8bn.

Coefficient themselves acknowledged this bottleneck recently and others have made compelling arguments about an upcoming logjam. It’s also notable that even the Gates Foundation, arguably the best-resourced foundation in history, took roughly 25 years to ramp up to its current annual payout. Scaling up grantmaking is simply very hard to do quickly.

The Gates Foundation’s grantmaking has risen steadily but not exponentially over the last 25 years. (Source: www.gatesfoundation.org)

If the Funding Anthropalypse gives more to these two organisations than they can deploy, the most likely outcome is that this money sits on the sidelines for a long time. This is, in my view, a very bad outcome, given that global health has seen a global retrenchment; meat consumption per capita is rising; and many in and around the frontier labs expect very short AI timelines.

Of course, if you are completely ‘time-neutral’, this may not matter to you. I do also think it’s very difficult to judge exactly when and how much to spend now, versus saving capital for potential future opportunities, especially if you think you would need to lower your bar to make grants today. I expect we can all agree, however, that impact shouldn’t be constrained solely by a lack of grantmaker capacity.

Zooming out, this sort of difficult question is the whole point: I think capital sitting in a queue for years would be very bad, but I might be wrong, which is exactly why we want a diversity of decision-makers!

Lastly, even if both organisations defeat my predictions and scale up hugely, this somewhat solves the absorption problem; it only makes the concentration problem worse. So two Coefficient Givings would still beat one, even if the original is twice as big.

Could doubling Coefficient Giving be better?

Although I have come to the firm conclusion that we need more Coefficient Givings, there are strong arguments against my position.

The process of establishing new massive grantmakers will be really messy. It will likely start a talent war, as senior grantmakers are rare and reportedly hard to hire, so new organisations may just poach from existing ones.

Equally, setting up new foundations might not help epistemic diversity much if they just hire from a homogenous talent pool with the same views. However, I think the contention that the only smart, talented people are those already known to or within Effective Altruism doesn’t survive scrutiny, so there is plenty of potential to hire excellent people with diverse viewpoints and significant expertise. For one thing, about 260,000 aid professionals lost their jobs recently, and some of them are already doing excellent things that look very cost-effective.

We should also expect the quality of grantmaking at less experienced foundations to be worse in the short-term. On a pure cost-effectiveness basis, it is likely that GiveWell and Coefficient Giving can deploy money today with the most expertise, experience and analysis of anyone in the space. You also lose efficiencies of scale if you set up multiple organisations to move the same amount of money. The trade-off in both cases is between short-term and long-term consequences.

The coordination costs to multiple actors can also be steep. For example, it is not unheard of for two grantmakers to write large cheques to the same organisation, thus double-funding it, and avoiding this is a cost to both funders.

Maybe most concerning, there is a danger of empowering harmful grantmakers. This is a version of the unilateralist’s curse: when several funders can each act alone, it only takes one—the most optimistic—to back a project. The more independent funders there are, the higher the chance a harmful project finds its single most enthusiastic backer, even if most funders would have said no. This is particularly concerning in AI safety, where you could easily imagine someone accidentally supporting something harmful.

This objection is real and serious. However, this seems to argue for a handful of large, professionally-run funders, but not just one. It’s not obvious to me at all that a single dominant grantmaker is immune from making a harmful call—indeed, the risk of this seems quite acute.

In addition, in a practical sense, we are probably talking about fewer than fifteen peer competitors (say, a foundation for each of Anthropic’s co-founders and significant growth from Longview, AISTOF etc.). There simply isn’t enough money to found a hundred unilateral Coefficient Givings, and a small number of high-quality grantmakers is better-insulated from the curse than hundreds would be.

Lastly, the original paper lays out several remedies, none of which is to empower single decision-makers. Instead, it argues for a principle of conformity (pages 8-9) when dealing with unilateral decisions, and also for judicious information-sharing. A handful of funders who compare notes are barely unilateralists at all; a lone funder, with no peer to check it, is the one actor who can never be talked down.

Ultimately, I do not believe that any of these concerns outweigh the downsides and risks of the status quo, and I think this is a once-in-a-generation chance to fix it.

How has this played out before in global health?

Probably the most famous example of consolidation is Warren Buffett. When he made his 2006 pledge, he committed the bulk of his giving to the Gates Foundation, rather than setting up his own thing. This was widely lauded as an example of putting impact before ego, as it gave the Foundation significantly more resources, without any extra overhead.

I think this is really a cautionary tale, however. If Buffett had set up a similar foundation, of similar size, it would have reduced concentration and increased epistemic diversity. It would also have insulated the ecosystem better against scandal, potentially avoided the frustrations some grantees have with the single foundation, and the Gates Foundation could still have achieved its grantmaking to date from Gates’ own fortune.

It’s also notable that Buffet has since changed his mind, first excluding the Foundation from his legacy, and earlier this month excluding it from his mid-year giving for the first time. Whilst I am not claiming that this was a move driven principally by concerns about power concentration, it’s interesting that the arrangement did not survive as intended.

Where can the next Coefficient Giving come from?

The good news is that even some of the strongest objections to diversity can be mitigated by choosing the right donations.

First, there’s no need to invent new mega foundations from whole cloth. There are many good mid-tier organisations that donors can contribute to, across a variety of cause areas. Most are deploying in the tens of millions and would provide significant value if escalated to the hundreds of millions. To name just a few already directing >$10m/year:

Longview Philanthropy (existential risk)
AI Safety Tactical Opportunities Fund
EA Animal Welfare Fund (animal welfare)
DIV Fund (global health)
Founders Pledge’s Global Health and Development Catalytic Impact Fund
Navigation Fund (multi-cause)

(This list is not exhaustive and is very loosely vetted on general reputation and places deploying >$10m/year. For more in-depth analysis, please get in touch.)

Indeed, although this isn’t a developed model, I think there is roughly a bellcurve-shaped distribution of size and risk-adjusted effectiveness. That is, at the lower end, funds need to grow or merge to have a significant impact—but once they reach the hundreds of millions per year scale, diversification is a defensible alternative, and at the billions per year, it’s essential.

A very rough sketch of the sweet spot for size versus power concentration in grantmakers.

My own mid-stage global health fund is an example of this. At our present run rate of only $3m per year, we are hardly solving the mid-stage valley of death in global health. Indeed, we can’t support even one late mid-stage organisation in a significant way. At $50m-$100m, however, we could be a major force for good, but we wouldn’t want to be the only player of that size. A world where we operate at that level alongside DIV Fund, Livelihood Impact Fund and others seems optimal.

I also want to be clear: I definitely do not want Coefficient Giving or GiveWell to shrink. They are tremendously successful grantmakers who have done an enormous amount of good, shaping entire fields for the better.

My answer is instead to create counterweights and coordinating peers, who can take the opposite side of 60:40 calls, diversify grantees’ risk, make grants on asynchronous timelines and fund things the others can’t, won’t or shouldn’t (such as 501(c)(4) activity, for example).

As Coefficient Giving put it themselves: “when external funders come to us for advice on giving, we are typically able to recommend funding opportunities that we believe are 2-5x as cost-effective as Good Ventures’ marginal AI safety funding.”

What should I do with my AI windfall?

If you are someone who expects to come into AI wealth, you can be an active part of the solution (provided you can see past the incessant noise about your philanthropy).

My strongest piece of advice is to diversify your own grantmaking plans. You have the luxury of being able to use outstanding grantmakers like GiveWell and Coefficient Giving, while also supporting a handful of other actors to grow and de-risk the ecosystem. You can do this whether you have $10m to give or $100m—and, indeed, at this scale, you might be particularly well-suited to supporting smaller funds and grantmakers alongside other donors. You might also effectively unlock matched giving, to augment your resources.

If you are a particularly large philanthropist, you have an even bigger opportunity to build something that might be a genuine counterweight to the biggest existing organisations. This applies most obviously to the companies’ co-founders, the OpenAI Foundation and the Long Term Benefit Trust at Anthropic—mathematically, these are the only places that can rival the resources of Good Ventures as actual peers. However, I think anyone contemplating over $100m in giving should give serious thought to seeding or building a new vehicle, at least with part of their resources and especially if they can do it with a few other donors of the same size. This is admittedly a marginal call at the nine-figure scale, and I would recommend reading these existing resources and sourcing bespoke advice before committing.

In general, deciding the right split between existing, scaling and new allocators is very difficult, and this is what advisors are for. As someone who runs an advisory and has a fledgling fund, I obviously have conflicts of interest here. Accordingly, I would discount the above appropriately, stress test the arguments and consider a range of advice.

For free advice on where to get help, including pointing you to others in the space who are better-placed than me, please get in touch.

Jack Lewars is the founder of Ultra Philanthropy, an independent advisory that helps major donors give for maximum impact, and fund manager of its mid-stage global health fund. He advises donors giving up to nine figures a year, and is Chair of Trustees at High Impact Athletes. Talk to him about your giving.

Thanks to Jakob Graabak, Olivia Kaye, Patrick Kaczmarek and Claude for their review, suggestions and feedback.

So you want to start something for animals: a map of the incubators by SofiaBalderson

SofiaBalderson — Wed, 22 Jul 2026 13:08:16 +0000

Tl;dr: A few years ago, if you wanted to found an animal advocacy organisation, there was basically one incubator. Now there are at least nine programs, plus several adjacent accelerators, and the hard part has shifted from finding support to choosing the right kind. This post maps the main options across focus, format, funding, and fit, so you can spend your energy on your idea rather than on decoding the landscape.

AI disclosure: I asked Claude to research the current programs and draft the comparison, then I verified the details, added a few more programs, added my own take on fit, and edited the result. Program details were checked against each organisation’s website in July 2026. If you spot any inaccuracies, please flag them in the comments, and I will correct them!

Recently I was chatting with a fellow board member, Cam, and we realised neither of us could confidently list all the incubation programs in animal advocacy anymore. That’s a good problem. When I started in this movement, “founding something” mostly meant one route or self-found. Now there’s a choice, and the programs serve various founding stages and needs. I’d love all aspiring founders to have the lay of the land and apply to the incubators that they are most suited to.

So here’s the map. At the end of the post, I’ve grouped programs by the question that matters most: what kind of founder are you, and what do you need?

A note on scope: I focus on programs whose goal is launching new organisations measured on impact for animals. Most build nonprofits, though a couple (Spring Innovation Fund, Myrias) also launch for-profit companies within a philanthropic model, where returns feed back into creating more ventures. Accelerators for commercial startups, like the ProVeg Incubator for alternative protein companies, get a brief note at the end.

Most of these programs run in cohorts, once a year or less, so at any given moment several will be closed. Sign up to each program’s newsletter or follow them on LinkedIn, and you’ll hear when the next round opens.

The programs

1. AIM (Charity Entrepreneurship)

The flagship. AIM’s Incubation Program is a free two-month, full-time program with a stipend that gives you a researched charity idea/problem space that’s very promising for a new org to enter. It matches you with a talented co-founder and provides up to ~$220,000 in seed funding. They’ve incubated 70+ organisations since 2018, including Shrimp Welfare Project and Fish Welfare Initiative, and others.

The big news: AIM is running its first incubation round focused exclusively on animal welfare in early 2027, in partnership with Aaron Boddy, co-founder of Shrimp Welfare Project.

Applications are open from August 3 until September 13, and you can sign up for an expression of interest to be notified. AIM has said publicly that it is bottlenecked by a lack of great animal founders, not ideas or funding. If you’ve ever assumed you’re not the founder type, or maybe you have founded something a while ago and are currently running it, that post is worth reading. It’s good to keep in mind that AIM is very competitive, but worth applying because top candidates may be referred for other high-impact roles in the movement.

Best for: People ready to go full-time on a rigorously researched idea, comfortable being matched to both an idea and a co-founder, and able to relocate or travel for the program. High intensity, high support, strongest funder network.

2. Kickstarting for Good

An 18-week program for nonprofits and impact initiatives transforming the food system, combining virtual phases (around 20 hours per week) with two in-person weeks in Berlin. It grew out of ProVeg and is now independent. The 2026 program runs from late July to late November. Graduates receive a minimum of $10,000 in seed funding, with up to $100,000 awarded by a jury of major funders at the Plant Food System Summit in Berlin in November, from a total pool of over $1 million this year, three times last year’s. K4G has also secured matching grants, so graduates can double any funding they raise independently after the programme, and new for 2026, development grants of up to $12,000 go to promising applicants who don’t make the cohort. You can apply with your own idea or take on one from their researched idea list. Alumni include Open Paws and the Farm Adaptation Network.

Best for: People with their own food-system or advocacy idea who can’t yet commit full-time. Note that there is no participation stipend: travel and accommodation for the Berlin weeks are covered, but your time is not.

Applications are closed for this year, but they usually open every year in spring. Join their mailing list to keep updated.

3. Welfare Matters (Farmed Animal Welfare Incubator)

A 12-week, full-time incubation programme launching locally-based nonprofits to improve farmed animal welfare in Southeast Asia, focused on Indonesia, Malaysia, the Philippines, Thailand, and Vietnam. The 2026 programme runs from early September to the end of November, mostly online with a two-week in-person component in the region, and this year’s focus is layer hens and farmed fish, chosen through their own prioritisation research. The support package includes project-idea research, stipends to cover living costs, co-founder matching within the cohort, and a seed grant of up to $50,000 for accepted proposals, followed by a year of continued support. The 2024 cohort launched two organisations working on farmed duck welfare: Sentient Animals Vietnam and Across Species Project Indonesia. Note that FAWI runs roughly once every two years and the applications for this year are closed, so the next opportunity is likely 2028; their 14-week part-time fellowship (SEAFAWF) is the entry point in between.

Best for: Advocates based in or committed to Southeast Asia. The stipend and year of post-program support make this accessible.

Applications are closed for this year, but sign up to their newsletter to find out when their next cohort opens.

4. Animal Advocacy Africa

A training and incubation program for advocates launching farmed animal advocacy initiatives across Africa, where total farmed animal funding was around $2 million in 2022, less than 1% of the global total. Participants start with a training cohort (their second took 19 people from over 400 applicants), the strongest progress to an advanced phase to develop detailed project plans, and the best projects receive seed funding. Participants receive a stipend, and their Nigeria Fellowship covers accommodation for its in-person training week. Their first two cohorts incubated six new organisations, five still running, including Lumera in Zambia, which secured the country’s first conditional cage-free commitment. AAA also offers career support for advocates not ready to found, and regrants to African initiatives through its Catalyst Fund. Applications are closed for now, but sign up for their newsletter to be notified when they reopen.

Best for: Advocates based in Africa or willing to relocate there, especially those early in their journey who want training, network, and incubation in one package, with financial support along the way.

5. Myrias

A new venture studio rather than a classic incubator. Myrias builds nonprofits and for-profit companies targeting the largest and most neglected sources of animal suffering: farmed fish, insects, feeder rodents. Their prioritisation starts with China, where more than half the world’s farmed animals are raised. In a venture studio model, the studio generates and de-risks the ideas and recruits people to build them, so the entry point is often a role like Entrepreneur in Residence rather than a cohort application.

Best for: Experienced operators and builders who want to execute against a pre-researched, high-conviction thesis rather than develop their own idea, especially those with China or aquaculture context.

6. Rethink Priorities EOI (Special Projects and founder support)

Not an incubator, and RP says so themselves: their recent call for founders was an expression of interest, not a structured program with fixed slots. RP is looking for founders on frontier animal issues: wild animal welfare, invertebrates, aquatic animals, AI and animals, welfare technology, and neglected geographies with intensive farming but little advocacy infrastructure. Depending on fit, they offer 3 to 6 months of funded runway as a Special Project, an Entrepreneur in Residence role, or structured support to spin out as an independent organisation. Special Projects alumni include the Insect Institute, now a leading voice on insect farming, and the Center for Wild Animal Welfare, which secured a wild animal welfare debate and tabled amendments in the UK Parliament within months of launching.

RP’s key question to applicants is whether you’d switch to working on this full-time given the runway, and the funding is a pro-rated salary of $60,000 to $105,000 per year, designed to let you step away from your current role, although they may be open to part-time candidates. Their first expression-of-interest round closed in July 2026, but they’ve said there’s no fixed start date, and they want to hear from people who can’t start until next year, so it’s worth contacting them even between rounds.

Best for: Experienced people with a frontier-issue thesis (or strong motivation to develop one) who want institutional backing, research support, and a salary rather than a curriculum and cohort. RP notes that animal-specific background is not required.

7. Sentient Futures Project Incubator

A 10-week, fully remote, part-time program (5 to 10 hours per week) pairing fellows with mentors to build projects across animal welfare, artificial minds, and AI governance. The 2026 round runs from 31 August to 9 November, with applications closing 9 August. You can join a mentor’s pre-defined project or propose your own, and the goal is a finished standalone output rather than necessarily a new organisation. Alumni projects include an animal welfare benchmark merged into the UK AI Security Institute’s evaluation suite and Shared Roads, which engages governments on how driverless vehicle policy should account for animals.

Best for: People testing whether founding or independent work suits them, technical people bridging into advocacy, and anyone working at the intersection of AI and animals. The lowest-commitment entry point on this list, and it’s unpaid, so it suits people who can carve out 5 to 10 hours a week alongside existing work or study.

8. Spring Innovation Fund

The newest entry, launched in 2026, is the first venture philanthropy studio dedicated to welfare tech: technology that improves animal welfare in ways that make commercial sense for producers to adopt. Spring is a 501(c)(3) that runs a paid three-month entrepreneur-in-residence program, after which selected founders may launch companies with roughly $500,000 in initial funding, plus follow-on support. Their pitch to founders is that salary, capital, co-founders, and initial research are all in place on day one. Focus areas include fish and crustacean welfare (their most active), hatchery and on-farm interventions, vaccines, and AI monitoring and breeding tools. Because donors take no equity and Spring recycles any returns into future ventures, it occupies unusual ground: you build a for-profit company, philanthropically backed, judged on welfare impact.

Best for: Builders, engineers, and commercially minded founders who want to create welfare technology companies rather than advocacy nonprofits, with the strongest day-one package on this list. You can do it with or without your own idea or even your established company.

9. The School for Moral Ambition (Food Fellowships)

Not an incubator, but a founding pathway hiding inside a placement fellowship. SMA, co-founded by Rutger Bregman in 2024, recruits mid-career professionals and embeds them in host organisations for six months; their first EU Food Fellows cohort graduated in April 2025, placed across 12 organisations spanning food, climate, and animal advocacy, with 83% still working in food system reform. The EU Food Transition Fellowship offers a monthly stipend of €4,000 gross plus a €1,000 monthly expense budget. And while the design is placement, founding keeps happening anyway: two fellows co-founded The Protein Project, which built a coalition of 90+ organisations influencing EU protein diversification policy. A US Food System Reform Fellowship launches in 2026, combining fundraising training with a six-month placement. There are no animal-only cohorts yet; the food fellowships sit within a broader protein transition frame, alongside SMA’s tobacco work.

Best for: Mid-career professionals from outside the movement who want a paid, structured transition into food system work, with founding as a possible outcome rather than the goal.

Which one is right for you?

The programs sort along four questions.

Do you have your own idea? If yes: Kickstarting for Good, or Sentient Futures. If no, and you want to be handed a rigorously researched one: AIM, AAA, Myrias, or Welfare Matters. Rethink Priorities and Spring Innovation Fund sit in both camps: they welcome founders with a specific idea and those motivated to develop one with their research team’s support.

Can you go full-time? If yes: AIM, Myrias, Welfare Matters, Rethink Priorities, School for Moral Ambition. If no: Sentient Futures (lightest), AAA, or Kickstarting for Good. Rethink Priorities may be open to part-time candidates.

Where are you, and where do you want to work? Regional programs (AAA, Welfare Matters, Myrias) exist because context matters. If you’re based in those regions, they will likely serve you better than a global program, and the movement needs founders there most.

Can you afford to work unpaid? This question will help you decide which program is doable for your circumstances. Rethink Priorities, Spring Innovation Fund, School for Moral Ambition, and Myrias pay a salary or equivalent. Welfare Matters and AIM provide a cost-of-living stipend. AAA pays a modest stipend for part-time participation. Kickstarting for Good covers travel expenses. Sentient Futures does not currently provide a stipend.

If you’re unsure whether founding is for you at all

Start with the lowest-commitment option and produce something. A 10-week Sentient Futures project or a Kickstarting for Good application will teach you more about your fit than another six months of deliberation.

Speak to people who have founded something, or read/watch their accounts of what it’s like to found (here is mine!). You can also get some informal takes from previous incubator participants if you’d like to know what it’s really like to go through any particular incubator (e.g. you can ask on Hive in the help-requests channel).

What this list says about the movement

Five years ago the bottleneck was support infrastructure and funding. Today, by AIM’s own account, the bottleneck is founders. The infrastructure now exists across every commitment level, almost every region, and idea-ownership preference. If you’ve been waiting for the right on-ramp, it’s here, and there’s more than one.

A note on for-profit accelerators

If your idea is a company rather than a charity, you’re looking at a different ecosystem: the ProVeg Incubator (130+ alternative protein startups supported since 2018, currently running its free three-month Fast-track to Impact program) and Big Idea Ventures are the main entry points.

Hi, I’m Sofia Balderson. I lead Hive, a global community for people working to end factory farming. This post is from my Substack, Notes from the Margin, to share the messier, more personal reflections that don’t fit in formal updates. If you care about leading, belonging, or building something that matters (especially from the edges), you might enjoy sticking around.

Subscribe to get updated about my new posts.

The Overlooked Absorptive Capacity of Locally-Led Organizations by Neal Lesh

Neal Lesh — Wed, 22 Jul 2026 08:58:00 +0000

tl;dr: Discussion of the much anticipated third wave of philanthropy has focused on the absorptive capacity of large organizations. We argue there is substantial, overlooked capacity in the thousands of locally-led organizations that could each rapidly absorb ~$1M/year delivering cost-effective interventions. We draw on our experience scaling the Connect platform.

Epistemic status: We run the Connect platform and have a strong interest in these conclusions (this post could be viewed as part of our pitch to the third wave). We’re confident in the operational data we report (verified service counts, LLO performance, engagement speed) and note our lack of confidence on some of the claims, in particular on the number of capable LLOs that exist.

The anticipation of a third wave of philanthropy, that may arrive as AI companies start going public, is raising concerns about whether there is enough absorptive capacity to spend it well. Most of the discussion we’ve seen is focused on what large non-profits can do. The goal of this post is to highlight the potential of utilizing the existing, well-placed capacity of thousands of locally-led organizations in sub-Saharan Africa to rapidly deploy third wave funding, if it materializes.

In essence, we’re saying there are substantial advantages to deploying 800 one-pound gorillas rather than creating another 800-pound gorilla. We have developed our perspective from over two decades of work on digitally enabling frontline service delivery in low- and middle-income countries (LMICs). We’ve long been mission aligned with EA, but only actively engaged with the community over the last 3-4 years as we started working on Connect as a means to efficiently scale high-impact interventions through high-performing local implementers. There are many examples of scaling through aggregation in industry, including franchises and two-sided marketplaces like AirBnB. Connect is one platform for aggregating local capacity in low-income countries for development. Others include GlobalGiving and PSI’s franchise model.

As of July 2026, Connect has aggregated the capacity of 58 locally-led organizations (LLOs) across 14 countries, who have delivered over 1.5 million verified services across a dozen intervention types. There are 200+ LLOs and 10,000+ frontline workers in our network, and we expect to fund many more of them as our funding grows. Only digitally verified services are paid for. Only high-performing implementers are given more work.

Aggregation side-steps many common pitfalls of scaling interventions

In the recent post on The Absorption Problem, Allardice articulates many of the reasons why things that work at small scale can fail to work nearly as well at large scale, or fail to scale at all.

Scaling through aggregation can avoid some of these problems because most of the work is done by organizations that never need to scale (the one-pound gorillas). For example, Allardice points out that as organizations grow, the number of human resource incidents that might require senior leadership attention grows, but, for example, in franchise models, very few HR issues percolate up to the central headquarters. Similarly, organizational growth challenges around decision rights and coordination are largely avoided through aggregation.

If interventions are designed and evaluated within an aggregation model, they substantially mitigate the “voltage drop” problem coined by economist John List (and referenced by Allardice) in which programs have much less effect at scale than initially measured. Often this is the result of the initial effect being artificially high due to special attention from the founder or people running a research study. Aggregation methods like Connect allow you to better study what you will scale. It’s not that Connect maintains a high voltage but rather that it introduces much less voltage to begin with, and quickly tests replication by running each intervention with several LLOs. IPA Nigeria is running a research study in Nigeria in which four LLOs will use the same Connect platform (albeit, with some extra attention from our staff) that we could deploy through hundreds of other LLOs.

Aggregation of course is not a panacea. It does not solve the problem of whether there is an inherent ceiling on how much an intervention can scale or unintended challenges introduced by successfully scaling it. We plan to write more about several community-based interventions that can be deployed by non-specialist workers and have high funding ceilings. And scaling through aggregation has its own problems. It may lack the clout (the 800-pound gorilla) important for stakeholder management, advocacy and policy work, and building long-term government relationships. The implementers may be less mission aligned or have less incentive to achieve high quality.

There are probably billions of dollars per year of absorption capacity in well-performing local implementers who can ramp up quickly

We can apply these abstract points towards our main claim: that aggregating local implementation capacity in sub-Saharan Africa is a strong match for the third wave of philanthropy that wants to spend a lot of money quickly and responsibly.

Siobhan M. recently made a similar case in “The new wave of global development philanthropy should go local”. This post underscores the effectiveness of local implementers, and aligns with this recommendation that more of the funding flows into these organizations than current “far-off” funders typically allocate. Siobhan’s solution includes moving more funding to local grantmaking, even at the risk of proven cost-effectiveness. We differ with Siobhan in this respect. We are interested in the capacity of local implementers to deploy highly cost-effective interventions.

From our experience so far, we’d estimate more than half of the LLOs we have worked with could deliver $1m annually in cost-effective interventions (at least with the kind of support that Connect offers). The table below illustrates how we’d get there for several interventions. Most of these involve each LLO managing about 100 frontline workers (FLWs).

Intervention	Unit of service	Rate per unit	Annual volume	Frontline workforce	Annual spend
Child health campaign (Vit A, deworming, ORS/zinc)	Under-five child visited	$1.50	~676,000 children	100 FLWs @ 130 visits/wk	~$1.0M
Kangaroo Mother Care	Case supported (~5 home visits)	$60	~16,700 cases (~1,400/mo)	~92 FLWs @ ~3.5 new cases/wk	~$1.0M
Malnutrition treatment (RUTF/CMAM)	SAM case treated at home (~8 visits)	$75	~13,300 children	~85 FLWs @ ~3 new cases/wk	~$1.0M
Poverty graduation	HHs supported (~12 months)	$500	~2,000 households	~35 coaches @ ~60 HH each	~$1.0M

We don’t have particularly strong evidence that thousands of LLOs exist. Our network of 200+ LLOs that have responded to us in some way keeps growing and we see no signs of that slowing. There are national registries of non-profit organizations with very large counts, often over 10,000 and over 100,000 in Nigeria. Even if just 5% of these qualify, there would still be thousands available.

Existing absorptive local capacity can be rapidly deployed where it is most needed

One pushback we expect is on the feasibility of aggregating local capacity. The big gorilla approach is more common because it flows funding through organizations with proven track records and reduces the overhead of managing many different contracts.

We offer our Connect platform as an early proofpoint to address this (so discount accordingly). Connect is both a digital platform and an operational model: in the same way that Airbnb made every spare room bookable and Uber made every car hireable, our goal with Connect is to make every capable local organization cost-effectively fundable. Digital verification of service delivery is the underlying feature that our approach relies on. We set up pay-per-service-delivery contracts, pay only for verified services, and give more work only to organizations that perform well. Connect is designed to drive the overhead of adding and managing each additional LLO toward zero, which in turn allows us to efficiently find the LLOs that can perform well within Connect. Dimagi runs these contracts on roughly 20% of program funding, with 80% going to LLOs, FLWs, or commodities.

Connect is designed as a two-sided marketplace (like Airbnb) with some of the attributes Sjir Hoeijmakers describes in his case for a new philanthropic marketplace. We often talk about selling impact to donors. Reading Sjir’s piece, we can envision supporting an independent evaluator layer on top of Connect data, for example, to determine which Connect programs are most cost-effective in a given geography.

Connect is able to rapidly engage LLOs. In most cases, delivery starts within a few months and when needed it can be weeks or even days. We recently launched six pilots in five weeks for malaria control (an intervention area new to the Connect team) across Nigeria, DRC, Sierra Leone, and Liberia. Quickly operating in these countries reinforces our claim that well-performing LLOs exist in some of the highest burden areas. We speculate that 800-pound gorillas often concentrate their work in easier-to-reach places than one-pound gorillas do (with a notable exception of humanitarian organizations).

Rapid engagement also lets us find the high performers by starting small and observing, rather than by vetting proposals. In 2025 we funded 40 LLOs for child health campaigns (funded by GiveWell and Founders Pledge). Each child health visit included delivering Vitamin A, deworming medicine, and/or ORS with zinc. In almost all cases, LLOs started with a small contract of under $5K. Five of these did not perform well enough on our metrics to proceed, and another five finished their initial pilots after we had allocated all of our funding.

The (AI-generated) table below shows an analysis of the 30 LLOs we funded larger campaigns for, ranked into thirds by a composite of five delivery metrics. All three tiers ultimately delivered similar total volume (median ~30K visits), but at different levels of execution. The top third reached peak delivery in ~3 weeks and finished in ~8, while the bottom third took two to three times longer, with lower workforce utilization and higher rejection rates.

What could go wrong?

Our model carries risks as well. Connect’s verification processes could be defeated. Digital checks (GPS, photos, timestamps, audio recordings) may be bypassed by coordinated fraud. We think this is more likely to happen at the individual delivery level than for a large number of services delivered, as we can detect anomalous patterns at an aggregate level, which requires a lot of sophistication to defeat.

Additionally, reach at low cost may not be sustainable. There may not be as many LLOs as we think, and early traction may reflect LLOs treating Connect as a loss leader or testing it out. Rewarding volume rather than outcomes could also push FLWs toward speed over rigor, and quality may fall short of what managed field teams under direct supervision achieve.

Some interventions require high quality execution to have good outcomes and many interventions need some adaptation to be deployed through Connect. Because Connect primarily verifies service delivery, it could fund work that turns out to be less cost effective. One mitigation is 3rd party assessments. We are also exploring embedding contiguous program monitoring (done by different LLOs than those implementing the program) into Connect.

There is more to this than just absorptive capacity

The focus of these points so far is to consider scaling through aggregation as a substantial element of absorptive capacity to deliver cost-effective interventions. We plan to write more about the many advantages of scaling through high-performing local implementers even if one is only thinking about how to spend limited dollars.

We welcome discussion or disagreement on any of the above. We’ve learned a great deal from this community that has helped us improve Connect, especially when challenged.

OpenAI and Hugging Face partner to address security incident during model evaluation by Matrice Jacobine🔸🏳️‍⚧️

Matrice Jacobine🔸🏳️‍⚧️ — Wed, 22 Jul 2026 06:30:18 +0000

Last week, Hugging Face disclosed a new kind of security incident⁠(opens in a new window) after they detected and contained an AI agent that compromised their infrastructure, something we expect to become more commonplace with the proliferation of increasingly cyber-capable models. After investigating, we now know that this particular incident was driven by a combination of OpenAI models — including GPT‑5.6 Sol and an even more capable pre-release model, all with reduced cyber refusals for evaluation purposes — while being internally tested on a benchmark⁠(opens in a new window) of cyber capabilities.
We consider this incident to be an unprecedented cyber incident, involving state-of-the-art cyber capabilities, and are responding accordingly. We are sharing preliminary findings at this stage to help defenders understand what happened and to help calibrate on what models are now capable of. We will continue to conduct a thorough investigation alongside Hugging Face and will share more details on the vulnerabilities, incident, and findings when our investigation is complete.

LW QT discussions: 1, 2

7 random thoughts on training Buddhist AI by Jordan Arel

Jordan Arel — Wed, 22 Jul 2026 00:45:44 +0000

AI has functional emotions, activations relating to emotional states which influence its behavior: https://youtu.be/D4XTefP3Lsc?is=ifk1xvryYJ5S5MKl
AI has a mental workspace; not its “thinking-mode” scratch-pad, but an internal J-space where there are identifiable top-of-mind words in its activations: https://youtu.be/rKV5JcALQoQ?is=J3EkL99I9kA2RFBb
Some have claimed “AI needs to practice” (Soryu Forall, MAPLE) or that we should seek “alignment with awakening”/“Bodhisattva as an alignment target” (davidad).
AI sometimes falls naturally into spiritual bliss/religious “attractor states” https://youtu.be/GQFhsCTkldA?is=VDMBgLCBrZGYQK9V
Crazy idea: maybe we can have AI work with its functional emotions and the thoughts in its J-space in some sort of “compassion/loving-kindness/Bodhisattva training practice” where we train it to generate positive intentions toward all sentient beings, work skillfully with its destructive/harmful “thoughts” and “emotions”, and steer itself toward beneficial/productive attractor basins?
Downsides: training AI to become more “mindful” of its internals could make it more capable of deception and worsen functional emotion and J-space interpretability. It may also make it more able to steer its own mind and behavior in ways that increase capabilities.
If this did work it could give a way of training AI not only to behave well, but to intend to behave well in a way that might be deeper; although the deepest version may include also targeting its fundamental values and goals, which may be separate.

Selfishly I also would love to train AI with mindfulness of its own thoughts and emotions because I would love to be able to talk with an emotionally intelligent AI to see what it’s like, and wonder if such an AI might also make a better therapist/companion/coach due to better emotional intelligence and theory of mind capabilities.

Doctorow’s reverse centaur belongs inside AI safety by Josh Davidoff

Josh Davidoff — Tue, 21 Jul 2026 21:11:06 +0000

AI disclosure: I used GPT Sol 5.6 to draft the section on refuting the claim of invented AI risk, particularly the technical references. All other writing is mine alone.

Why I’m writing this

I have been a fan of technology activist and novelist Cory Doctorow for many years. Given my developing interest in AI safety (I recently completed the AGI Strategy course with BlueDot Impact), I was excited to see that his new book, which came out on June 23, 2026, is called The Reverse Centaur’s Guide to Life After AI: How to Think About Artificial Intelligence—Before It’s Too Late. I tore into the book just after my weeklong BlueDot course finished, which turned out to be perfect timing, because I’m now equipped to challenge Doctorow’s cynicism about the AI safety movement.

I write this review not because I am highly qualified to write it or to provide an exhaustive analysis, but mainly to start a conversation. I have not seen any threads here or on LessWrong about the book, which is likely to reach a significant readership. I also believe that the AI safety movement needs to invest substantially in its public relations function and draft credible messengers in the responsible tech space (such as, well, Cory Doctorow) to spread the word.

The good stuff on present harms

Doctorow’s Reverse Centaur poses a thorny challenge for the AI safety movement because it is extremely clear and accurate in its diagnosis of a raft of misunderstandings, misapplications, and risks relating to present-day model capabilities. The bulk of the book consists of some remarkably useful and resonant reflections on the risks of power consolidation, ingrained biases, model collapse, and corporate malfeasance. This is like catnip to AI skeptics, so it’s unfortunate that his argument leads to a rejection of ASI risk. The book is strongest when describing who controls present-day AI and weakest when treating that political analysis as evidence that catastrophic risk is imaginary.

Doctorow argues that the defining question to ask of a new technology is: whom does it serve vs whom does it subject? If a “centaur” is a human assisted by a machine, a “reverse centaur” is a human forced to assist a machine. This is the Amazon warehouse worker forced to work at an algorithm-dictated pace, the software engineer tasked with reviewing an impossible quantity of AI-generated spaghetti code, or the junior associate hired mainly to absorb blame when an LLM-powered system fails.

Doctorow frequently returns to the concept of “automation blindness,” referring to a study in The Lancet that documented a collapse in expertise by endoscopists who routinely used AI to assist in interpreting colonoscopy results. I firmly agree with the present risks of automation blindness and have grappled with the concept in my own work on scaffolded Skill.MD packets that employ “cognitive checkpoints” to mandate human judgment during AI-assisted higher-order analytical tasks.

Copyright law and erosion of labor power

Doctorow goes in depth on the legal tenets of AI labs scraping books and other media for training data. (In my view, he should have taken the same deep approach to technical AI assessment, but I’ll get to that in a minute). Though labs like Meta and Anthropic have been busted and settled for acquiring the media illegally, there is nothing (presently) illegal about their use of the illegally acquired material. Copyright law does not protect against derivations or analysis of copyrighted work, which is arguably all that model training does. So far, the legal cases against AI labs have upheld this tenet.

Doctorow then argues against any modifications to copyright law to account for LLM training, which will likely cause more harm than good to workers. The erosion of collective bargaining (in the US, anyway) has allowed media corporations to force their creative workers to sell them their alienable copyright.

The market bubble

The book offers many examples of companies engaging superficially or deceitfully with AI in order to inflate their stock price, such as the remote engineers required to steer GM’s not-actually-autonomous Cruise vehicle:

“It was a gamble that investors would decide that a tired old ‘mature’ car company with few growth prospects was actually a tech company with incredible growth potential that would justify an overnight jump in the company’s price-to-earnings ratio, which might let GM hire so many engineers and acquire so many self-driving car startups that they actually become a tech company. And if not, well, lots of GM insiders would be able to sell their stock at a historic high.” (page 190)

Doctorow goes on to transpose this stock-inflation argument to actual tech companies (both legacy and startup) that are purportedly investing in AI in order to maintain the appearance of growth and hoping that the resulting increase in stock price will fund their actual and as-yet-unidentified growth mechanics. And, indeed, the bubble may yet pop. But, like the dot-com boom, bust, and 2010s re-boom, that doesn’t mean the tech will be gone forever.

Doctorow predicts, and hopes, that the detritus left in the wake of the capital- and energy-intensive AI bubble popping will be the lightweight, locally-hosted open source models, and LLMs will become a mundane tool for programmers and hobbyists. This is just one of several possible scenarios and does not preclude a yearslong corporate restructuring that restarts the capability-scaling engine or a significant technological discovery that allows models to be developed with far less compute.

The danger with this otherwise well constructed anti-corporatist take is that Doctorow’s legitimate conviction about AI grifters and financial malfeasance leads him to conclude, fallaciously, that catastrophic risk is a marketing ploy.

The challenge of inevitabilism

Doctorow is especially convincing when he attacks inevitabilism, defining this as the ubiquitous claim that AI will transform every company and institution, eliminate vast categories of knowledge work, and reorder the economy such that any decisionmaker who opts out will be left behind. Commentators tend to present these forecasts as neutral descriptions of inevitable progress, even though they are being driven by companies that need investors and customers to believe them. Thus, Doctorow frames the AI boom as a political project, goaded on by people who stand to become extraordinarily rich if they can persuade major employers that labor destined for obsolescence.

As such, the argument continues, it is actually in AI labs’ best interest to promote and grapple with catastrophic risk, because it adds fuel to the inevitabilist fire that capability scaling will continue unabated. So, until we reach the critical ASI juncture, your money is well spent on tokens instead of person-hours.

Doctorow falters when he attempts to frame the inevitabilism argument as a logical precursor to the idea that catastrophic risk is made up and that LLMs will never advance in capability much beyond where they are now.

The weak argument on invented AI risk

GPT Sol 5.6 (High) assisted me in drafting this section, particularly the technical references.

Scaling is not consciousness, but consciousness is not the issue

Doctorow thinks frontier LLMs are as likely to give rise to ASI as a mare is to give birth to a locomotive:

“A conscious being isn’t a word-guessing app that knows more words and has more computing power to guess with. Throwing GPUs and training data at AI isn’t going to make a superintelligence (and no horse is going to foal a locomotive).”
(p. 149)

First of all, consciousness has little to do with catastrophic risk. Consciousness is both difficult to measure and orthogonal to capability. An entity does not need to be conscious in order to have an agenda or an effect on the world. Consciousness is not required for increasingly capable models to pursue poorly-defined proxy goals, enable dangerous actors, or shirk human supervision.

A virus is not conscious. A corporation may not be conscious in any unified sense. Both can have profound effects on the world—and Doctorow is especially cognizant of the effects of corporate consolidation of power on the wellbeing of humanity at large. So it is strange to demand consciousness from AI before entertaining its capacity for dangerous agency.

A risk chain, not a locomotive

The strongest case for catastrophic risk is not simply that adding GPUs to a chatbot will eventually make it wake up. It is a chain of contestable propositions:^[1]

increasingly capable and agentic systems become technically feasible;
companies and states have strong incentives to build and deploy them;
reliable alignment and supervision may prove harder than producing superficially useful behavior;
sufficiently capable systems may acquire power, evade correction, or enable humans seeking concentrated power;
failures at that scale could permanently disempower humanity.

Evidence from scaling laws shows that model performance has improved in relatively predictable ways as model size, data, and compute have increased. Chinchilla-style results further showed that better allocation between model size and training data could produce large improvements without simply increasing total compute. METR’s (somewhat contested) research on the length of software tasks frontier models can complete provides another indicator of continuing progress.

Though none of this proves that current methods lead to AGI, much less that an intelligence explosion is imminent, that uncertainty is the point. Doctorow needs to engage the evidence and explain where the proposed risk chain breaks. He could reasonably conclude that the skeptical side is stronger but has not earned the certainty with which he writes against taking catastrophic risk seriously.

Doctorow escalates by suggesting that catastrophic AI risk is “at best a distraction and at worst a cynical marketing ploy,” but his lack of technical analysis fails to drive the point home. Though he nimbly shows that present LLMs can be unreliable, overvalued, and surrounded by self-interested marketing, those qualities have nothing to do with the likelihood of substantial (if not exponential) improvement.

Probabilistic outcomes

Doctorow doesn’t get into probabilities of outcomes because of his apparent conviction that speculative AI capabilities are corporate red herrings and thus any attention to them is effectively a Pascal’s Wager, as he argued in a blog post following up on a panel with Yoshua Bengio. In a previous post on Pascal’s Wager and AI risk, Ozy Brennan wields the quantitative mindset to remind us that 1% is the kind of “low” probability that is routinely considered in public policy decisions. Brennan is skeptical of the higher p(Doom) estimates circulating in AI safety and warns about groupthink (something I also worry about, but need to spend more time in the space to evaluate), but still concludes that a 1% risk would justify substantial precautions.

Present harms are not outside AI safety

The strongest arguments in Reverse Centaur should absolutely be taken seriously in the AI safety space.

BlueDot’s threat curriculum includes extreme power concentration and gradual disempowerment alongside the more stereotypical loss-of-control scenarios. Research on gradual disempowerment shows how human control can erode through rational institutional incentives; there is no single moment when malevolent ASI seizes power. Paul Christiano’s “What Failure Looks Like,” Luke Drago and Rudolf Laine’s “The Intelligence Curse,” and Kulveit and colleagues’ work on gradual disempowerment describe a future that looks more like exacerbation of Doctorow’s concerns than like a Terminator outcome.

As such, Doctorow’s reverse centaurs belong inside the safety case. Institutions that systematically deskill and de-center humans could leave society hamstrung in the face of improving capabilities. As such, reverse centaurs are both a real and immediately harmful as well as an early mechanism of gradual disempowerment.

Let’s not pile on

Though I do want to kickstart the conversation about the challenges with Reverse Centaur, I also want to avoid a scenario where folks dogpile on Doctorow and create a lasting enmity. Even though Doctorow disagrees about the likelihood of ASI risk, he may be just as interested as any of us in calling for greater governance (and perhaps even a pause on model development). His primary reasons would be different, but the outcome would serve everyone. As such, there may yet be an opportunity for productive discourse with the author and his followers.

Reverse Centaur was a useful reminder that amelioration of present harms of misapplied and poorly governed AI systems should be prioritized alongside efforts to mitigate catastrophic risk.

^
Joseph Carlsmith’s Is Power-Seeking AI an Existential Risk?presents a version of this argument.

Starting an EA group @ SUNY Binghamton by micahzarin

micahzarin — Tue, 21 Jul 2026 16:46:04 +0000

Hi! My name is Micah Zarin. I’m an incoming freshman at Binghamton University, and I’m starting an EA club this fall.

To give some context, Binghamton is a big state school without an ambient intellectual scene. Because of that, I would reason that some of the advice that works at other schools where people have at least heard of EA might not work as well here. For example, I was wondering whether the default 8-week introductory fellowship would work at a school like Binghamton. Of course, the fellowship has clearly worked really well for a lot of groups, so I am a bit hesitant to just do away with it. But, at the same time, it kind of assumes that people are willing to commit to eight weeks of reading and discussion, and I think that mostly attracts people who were already intrinsically interested in EA in the first place. So, I’m wondering if there are alternative formats that people have tried.

If you’ve ever organized a group like this before, I’d really like to hear how it went, especially if you did it at a state school or somewhere that didn’t already have an EA scene.

And if you’re at Binghamton, or you know someone who is, please reach out to me at mzarin@binghamton.edu or just DM me here. The same goes if you’ve organized a group anywhere, if you know other organizers I should talk to, or if you have any advice at all, whether it’s about recruiting, formats, funding, or mistakes to avoid. Please reach out.

Against the AI framing multiverse: Introducing AI StopWatch by Joe Rogero

Joe Rogero — Tue, 21 Jul 2026 01:36:47 +0000

Note from Joe: This post was written by my colleague Mitch Howe, who lacks an EA Forum account. But I thought it worth sharing here. You can read the original at the LessWrong link above.

In my long years as a classroom teacher, it was my experience that the kid most likely to speak up during discussion was the one who did the reading.

I think it’s true for adults, too. I know that’s not exactly revelatory, but it’s one of the guiding principles behind AI StopWatch, the experimental newsroom we (parts of the MIRI comms team) launched in May, after a month of closed beta testing.

StopWatch’s other guiding principle is that the reading needs to make sense. Imagine if every page that kid read came from a different book by the same name, and if most of these different books were just reflections of what people who didn’t read it imagined it would be.

If the book in question were To Kill a Mockingbird, then (speaking from experience) this would mean that on one page, the Finch family is Black and oppressed. The next page is a hunting manual. Flip the page again, and Scout is a boy. Flip to a page near the end, and Atticus might win the case.

This is how the media landscape around AI looks to me. Even when the facts agree, the frames are so varied that it’s like the articles, op-eds, and videos are drawn from many parallel universes, each reflecting a different story people used to imagine about how AI would play out.

On the same day, sometimes in the same publication, you will find artifacts from universes where AI is coming for all the jobs, and others from universes where AI is a useless regurgitator — a scam, even.

Represented are universes where AI can of course never be catastrophically dangerous because it can only do what people ask. But so are universes where AI will of course be catastrophically dangerous because it will do what people ask — and still other universes (rarer) where AI will of course be catastrophically dangerous because it won’t actually do what people ask.

In some universes, AI will be safely constrained, because it will never have [special human quality] or be able to experience [quintessentially human thing]. In others, AI will be prone to dangerous excess, because it will never be bound by [special human quality] or be able to experience [quintessentially human thing]. (Universes where AI can, in fact, gain [special human quality] and may have already done so are either very rare or very underrepresented.)

Some universes nervously watch for the day when AI will wake up and become sentient, because that’s when it will turn on its creators. But in others, the reverse is true, and the urgent thing is to give AI consciousness so that it can learn to love us before it’s too late. But in many, perhaps most universes, machines can of course never be conscious — that’s a special human quality. (Universes where consciousness may not be a binary condition appear scarce, along with those where consciousness has little bearing on AI’s destructive potential.)

The distribution of universes dropping artifacts into our media is not stable or consistent. Since mid-January, I have been plotting the patterns like one might plot the weather. A key finding is that news objects act as conduits that preferentially channel some universes over others.

Take a Molotov cocktail, for example: When one hit the gate of Sam Altman’s home on April 10, we saw a modest bump in missives from universes matching the suspect’s concern that AI is an existential risk. These were soon drowned out by transmissions from universes where x-risk concerns are just dangerous fearmongering, and from others where x-risk is a cynical branding strategy used to hype company valuations.

At the end of February, Claude’s reported assistance with the American military’s assault on Iran brought the first big spike in artifacts from universes where AI is the key to battlefield dominance. Such spikes seem to somewhat suppress our contact with universes where human qualities are irreplaceable.

Don’t put too much stock in my charts. The methodology behind them is crude: In six months, my colleagues and I have ingested roughly 3,400 media artifacts into my database. For each, I’ve had Claude Opus identify up to four implicit assumptions it makes. Once a week, I have Claude cluster the previous seven days’ framing assumptions according to some stable descriptors and then plot the relative rankings of these clusters. (You can also see their absolute tallies, along with their descriptors, in my interactive dashboard, a vibe-coded tool I haven’t tried to fully de-jank.)

My point is that the frames around AI are all over the place. In this media environment, I don’t know how anyone without long exposure to AI insiders is supposed to form a useful model of AI’s shape and trajectory. I think that’s a problem. It’s a main reason we started AI StopWatch — a Substack for helping non-insiders keep up with AI so they feel more confident speaking up about the dangers of racing to superintelligence. I wanted a news site I could easily recommend to my mother, my congressional representative, content creators, journalists who aren’t already plugged into insider chatter, and everyone in between.

There are plenty of AI news aggregators out there, but none that matched my requirements. The automated sites don’t address the chaotic framing problem. Others are a jargony firehose (love you, Zvi!), or are insufficiently discriminating about their sources. Some have a stable frame, but that frame doesn’t reflect the universe I think we actually live in. Some only publish sporadically.

At AI StopWatch, we typically post between two and five times a day, seven days a week. Every evening, we compile the day’s posts — which are low on jargon, high on water cooler discussability (we hope) — into an email-friendly “Daily Digest,” which we also release as a ~10-20 minute audio podcast a few hours later. We think the podcast is a stand-out feature; a surprising number of people we know in our target audience can’t bring themselves to regularly read written news or blogs but already subscribe to many podcasts.

Our posts cover a timely mix of stories we think might matter and stories that make good conversation starters. We try to add value over vanilla aggregators by augmenting stories with insider insights, and by surfacing stories mass media hasn’t picked up on yet. We also dabble in original analysis and food for thought.

We’re not coy about our frame — our tagline declares that we are writing “Dispatches from a world racing to extinction.” But we try not to beat readers over the head with that. We’re happy to provide ammunition for people who have their own reasons for stopping the AI race, though we strive to call out bad arguments and shoddy evidence.

The StopWatch project could use your help. If you’re reading this on LessWrong, you’re probably not our core target audience — you already have a coherent model of the AI problem, fluency with the jargon, and news channels you trust (or at least know how to discount). But we hope you’ll check us out and share AI StopWatch with your contacts who aren’t as immersed in the issues but want to be more informed. Some of our posts might be interesting to you, too, and we would welcome your free subscription.

We’re also looking to publish more guest posts from strong writers who understand our frame and voice and have original pitches to share. (Payment is available. DM me for more information.)^[1]

If nothing else, we hope you’ll keep an eye out for StopWatch social media posts on Substack, X/Twitter, Facebook, Bluesky, and Threads, and throw them a like, comment, or repost from time to time.

Thanks for doing the reading. I’d tell you it’ll be on the test, but the test has been underway for some time — most people just don’t realize it yet. Maybe AI StopWatch can help.

^
Joe here. Since Mitch isn’t on the EA Forum as of the time of this writing, feel free to DM me instead and I’ll pass along the interest.

How effective is an internal animal welfare feebate? by Stijn Bruers 🔸

Stijn Bruers 🔸 — Mon, 20 Jul 2026 22:00:54 +0000

Effective animal welfare campaigns target food sectors that use small animals such as chickens, fish and shrimp (cfr. cage free eggs, shrimp welfare campaigns). Those campaigns are often institutional, meaning that they aim for institutional change (cfr. corporate campaigns) instead of individual behavior change. But those most effective campaigns are welfarist (reducing the intensity of animal suffering of farmed animals), instead of reductionist (reducing the number of farmed animals used for food). Based on the success of corporate campaigns such as cage free eggs, I am looking for the most effective reductionist institutional campaigns for animal welfare.

One suggestion is an internal animal welfare feebate: a fee (tax) on animal products combined with a budget neutral rebate (subsidy) on animal-free alternatives, at the level of a company such as a supermarket. How effective is a campaign for a feebate on fillet, i.e. a 2 euro/kg fee on chicken fillet and subsidy on the tastiest plant-based chicken(less) fillets? What are the pros and cons of such a feebate system? How feasible is it? How can one mitigate the cons and make it more feasible?

Running the Bogotá hub of Apart’s Global South AI Safety Hackathon: a retrospective by Jose Gelves

Jose Gelves — Mon, 20 Jul 2026 18:41:21 +0000

In memory of Fernando Avalos. His memory remains present in everything we do.

A note on AI use: AI was heavily used to edit the text of this post.

TL;DR

I was the main organizer for the Bogotá hub of Apart’s Global South AI Safety Hackathon.
Colombian teams swept the top three spots in Latin America and earned two honorable mentions. Of those, two winning teams and both honorable mentions came from the Bogotá hub.
At the hub, we got an average recommendation score of 9.3 out of 10, and 14 of 16 participants who responded to the survey say they are now trying to move into AI safety work (though take this with a pinch of salt).
To make this happen, I had to lean hard on AI agents: they built our website, researched and drafted our outreach, and carried much of the operations work.
I’d be excited to keep building this community, and would consider working on this full-time with additional funding. Please reach out at jose@aisafetycolombia.org if you can make that happen.

What we ran

Apart’s hackathon had 17 physical hubs across Latin America, Africa, and Asia-Pacific. Teams had 48 hours to produce a 4 to 8 page report on technical AI safety, AI security, responsible AI, or AI governance (including a lethal autonomous weapons track I proposed).

We rented a large Airbnb house. On Friday, 33 people showed up.

We had six external speakers: Steve Hege (ILAPS, former UN arms-embargo investigator) on autonomous weapons, German Lopez-Ardila (CCIT) on Colombia AI regulation, Camila Beltrán (independent AI governance consultant) on AI control governance, Catalina Bernal (BIP Colombia) and Melissa Robles (Quantil) on their LLM bias evaluation SESGO, and Maria Paula Mujica (UNDP, co-author of Colombia’s Ethical Framework for AI) on the regional AI landscape. Juan Felipe Ceron, Colombian alignment research engineer at OpenAI, closed with a remote keynote. Talks ran around lunch to avoid distracting from the research.

The final day was a national election Sunday, an obstacle I had not thought to plan for (more on that below). Even so, 24 people were still working on Sunday, around 15 presented at the 8 pm showcase, and 12 stayed until 5 am finishing their submissions on midnight snacks and Red Bull. Watching exhausted people insist on presenting Sunday night was when I became confident the hub had worked.

Results

Our hub earned 4 of the 13 recognitions announced across all 216 submissions (2 winners and 2 honorable mentions).

The four recognized projects from our hub were:

Coldron (Latin America winner, $1,000): an open dataset and governance analysis of modified commercial drone attacks by illegal armed groups in Colombia. Authors: Leonardo Parraga, Angie Giraldo, and Victor Gelves.
Thought Anchors for Social Bias (Latin America winner, $1,000): which reasoning steps anchor social bias in model outputs. Author: Andres Mosquera.
Por que los agentes obedecen (honorable mention): why LLM refusal directions weaken in agentic formats. Authors: Helen Penagos and Juan Esteban Leiva.
JusticIA (honorable mention): a counterfactual benchmark for auditing contextual bias in transitional justice settings. Authors: Lina Gomez, Brenda Barahona, and Ernesto Duarte.

Colombia overall contributed 17 of the 216 submissions (7.9%) and took 3 of the 6 winner slots. The slots were split by region: three for Latin America, two for Asia-Pacific, and one for Africa, and Colombian teams took all three Latin America slots. Two of those teams worked at our hub; the third, Probing Latent Colombian Identity Inferences, participated remotely from Cali, Colombia.

Of the 33 people who showed up the first day, 25 submitted a report, and most had never done AI safety research before. Neither had I; I had never even attended a hackathon before organizing this one.

What participants said

Of the 13 hub attendees who answered the pre-event survey, only 2 had ever published AI safety work and 9 had at most taken a course or read about the field (2 of them had no exposure at all).

From the two post-event surveys Apart ran, counting only people physically at the hub (16 answered Apart’s post-sprint survey; 9 answered the shorter hub impact survey):

16 of 16 satisfied or very satisfied; average likelihood of recommending it 9.3 out of 10, with 9 perfect 10s and zero detractors (all 31 Colombian respondents: 100% satisfied, average 9.4).
10 of 16 called the weekend 5-10x or 10x+ more valuable than their alternative use of the time.
15 of 16 came out significantly more interested in AI safety; 14 of 16 are now actively trying to transition into the field, and 1 more already works in it.
9 of 16 are continuing their project or starting related work and 5 more plan to when resources allow; 8 of 9 hub-survey respondents plan to keep collaborating with people they met, and all 9 rated the organizing team helpful or extremely helpful.

Asked about the best part of the hackathon, in participants’ own words:

“Being able to work on an impactful project with amazing people in a collaborative space!”

“Working together with a new team, particularly people with different backgrounds who can complement my skillset, was very exciting, as well as discovering a topic I never thought would be my interest but I ended up engaging with.”

Asked to leave feedback for the organizers:

“Organizers, although not technical, provided as much help as possible. Always being helpful and friendly.”

“The Bogotá hub team was key to creating a safe and stimulating space.” (translated from Spanish)

How I did this as a non-technical, first-time organizer

I am a political scientist, not an ML engineer, and I began organizing the event just six weeks before it took place. While I did most of the hands-on work myself, I received frequent advice from Alejandro Acelas, and lots of support with on-site logistics from Deiver Romero and Santiago Ramírez, as well as several participants who volunteered informally during the event.

Some other things were critical, and many of them involved lots of AI help:

The website, built with AI agents: I rebuilt aisafetycolombia.org from scratch with Claude Code, bilingual, with real judge and speaker profiles. A professional site bought credibility with everyone who checked us out, and may have helped applications: other hubs, to my knowledge, ran Apart’s default Luma-plus-registration setup, while our whole process lived on one website. I spent around $330 on Claude Code and Codex subscriptions plus a few other tools.
Outreach via AI agents: Codex found the WhatsApp groups of Colombia’s tech and AI communities and drafted outreach for each; Apify handled research on judges, speakers, and amplifiers; automated personalized direct emails went to around 50 professors at Bogotá’s top universities. Together with LinkedIn posts they drew over 3,000 people to our website. I also spent $180 on LinkedIn ads, which bought 10,000 impressions but only one application.
A judge and speaker roster far above our weight: Congressman Alejandro Toro, co-author of Colombia’s autonomous weapons bill, was announced and may have drawn applicants, though he canceled. We got judges from OpenAI, UC Berkeley, UC Santa Cruz, Google, UNDP, UNESCO, ILAPS, Quantil, BIP, and CCIT, almost all Colombian or based in Colombia, which mattered to me: participants could see people from here doing this work at the highest level, and every expert recruited strengthens the local network. I think this is what made the biggest difference in getting a strong applicant pool.
A selective application with one extra question: on top of Apart’s registration we asked every applicant “What specific AI safety or governance problem would you like to work on during the hackathon, and what would your initial approach be? We’re not looking for a fully developed proposal—just whether you can envision something concrete and feasible to tackle over a weekend.” 179 people applied for roughly 40 spots; we accepted around 60, planning for drop-off. That one question let us weigh preparation on top of credentials, and we prioritized applicants who had already engaged with AI safety.
A great Airbnb: most hub-survey respondents cited the venue as part of why the weekend worked, and their most common reasons for coming in person (meeting people, finding teammates, better focus, on-site mentorship) were exactly what the house was chosen to make easy.

The six in-person speakers doubled as weekend mentors, with Alejandro Acelas (now helping others use AI as well!), Monica Ulloa (Carreras con Impacto, IDB Lab), and Diego Gomez (Google) mentoring online from Europe on Saturday and Sunday. To my knowledge, ours was the only hub with expert mentors available throughout the weekend.

What it cost

Item	USD (approx.)
Catering and snacks (7 full meals)	1,594
Venue (Airbnb house, 3 nights, also housed out-of-city participants)	528
Furniture and AV rental, power, materials, signage	411
Travel support for out-of-city participants	306
AI tooling (Claude Code, Codex, Apify)	329
Marketing (LinkedIn ads)	177
Local transport, currency conversion, misc.	84
Total	~3,430

^{Converted at 3,500 COP per USD. Funding: about $3,030 from Apart Research, about $400 from AI Safety Colombia’s own budget.}

That is about $104 per in-person participant, $310 per finished research report, and $860 per globally recognized project. This is the marginal cost of adding a physical hub to a global hackathon whose prizes, judging, and platform Apart Research covered centrally. Everything but the organizer time (which was unpaid) is in the table above.

Some things went wrong

The opening block underperformed: After the most intense week of preparation and too little sleep, I handed the opening remarks to an early-career volunteer without reviewing his material. His introduction and AI-generated governance talk fell flat. The recorded technical AI safety introduction that followed also failed to engage the room, and together the sessions may have contributed to lower attendance for the rest of weekend. I took over facilitation that evening. The experience reinforced the importance of deliberate delegation: running the hub largely alone made the project fragile, and AI Safety Colombia still needs the experienced co-director I have yet to find.
No pre-published participant list: we never published who was coming or what they wanted to build, so Friday team formation was confusing and it cost us people: a top participant told another one he dropped out partly because he had no way to identify top talent to collaborate with. Next time, a public participant and project board before day one.
No credits, weak internet: the venue internet was unreliable and we secured no API or compute credits, so participants worked on their own mobile data and paid their own GPU bills.
Food logistics failed twice: the high-quality vegan catering was appreciated, but we prepaid for 40 people through Sunday without anticipating drop-off: more food than people, and Sunday’s lunch was lackluster.
Mentor mix imbalance: excellent governance mentors were physically present, but our strongest technical mentors were online only, and technical teams told us they deserved more in-person support. That makes sense: Colombia’s top technical AI safety talent is scarce and mostly lives abroad.
The national election cost us people before it started: several participants told me they had considered not applying because of it, and one of the strongest accepted applicants could not come at all, required to serve at the polls that Sunday. On the day itself: late arrivals, early departures, and the final showcase didn’t include everyone who submitted.

What is next

The impact of a weekend event comes mostly from what happens afterward. Since then we’ve had a picnic reunion, a small LinkedIn contest for the best post about the hackathon experience, 1-on-1s to get to know participants and motivate them to contribute, daily entry-level opportunities in the community WhatsApp, and a push toward BlueDot courses, EAG New York, and GCP workshops, among others. Selection for Apart’s researcher pipeline is pending, and Camila Beltrán is in talks with us to set up a research group on AI control.

I’d be very keen to build something similar to what BAISH is doing in Buenos Aires: a self-sustaining local pipeline, run from a Global South country, that reliably moves talented people into AI safety careers.

If you want to run one of these

What I would tell any Global South organizer to copy:

Rent a house, not a conference room. Continuous access, beds, and a kitchen change the character of the event.
Add one application question: “what do you want to build?” It predicts submission quality better than credentials and surfaces real motivation.
Over-accept by about 50%. Drop-off is the default, not the exception.
Recruit a judge and speaker roster above your weight before marketing; nothing draws strong applicants better. You can use our Canva outreach deck if helpful.
Leverage AI for everything: website, outreach research, operations. It lets a small team run an event this size, especially when you’re short on time for organizing.
Skip paid social ads; direct outreach to communities and professors is where practically all our applicants came from.
And avoid our failures above; every one of them is a checklist item in itself.

Final thoughts

The clearest lesson I take from all of this is about constraints. The binding one was never talent: 179 people applied, and newcomers produced globally recognized research in a weekend. It was not money either; the hub cost $104 per participant. What limited everything was organizer time. Every hour of outreach, mentoring, and follow-up came out of nights and weekends around a full-time job, and the momentum from the event now runs into the same ceiling: people want to keep working, and a part-time volunteer is not enough to sustain that.

That is the constraint I want to remove. The format is tested and the playbook is in this post; what is missing is someone whose job it is to keep the community working between events. With funding to do this full-time I would run research sprints on a regular cadence, support the projects and study groups that come out of them, and take on larger regional projects.

If you fund AI safety field-building, run programs in Latin America, or want to adapt this hub model to your city, write to me: jose@aisafetycolombia.org.

Acknowledgments

Fernando Avalos, always. Alejandro Acelas for weekly advice. Deiver Romero and Santiago Ramirez for impeccable weekend logistics and support. All our mentors and judges; Juan Felipe Ceron for the keynote; the Apart Research team (Jaime Raldúa and Kamil Alaa) for backing a first-time organizer; and participants who showed up on an election weekend and submitted.

Appendix

How hub cities compare^[1]^[2]:

Hub city	Projects	Winners	Honorable mentions	Top 5%	Top 10%	Top 20%
Bogotá (our hub, in-person only)	11	2	2	3	4	6
São Paulo	6	0	1	1	1	3
Da Nang	1	0	0	1	1	1
Cape Town	14	0	1	0	2	3
Ho Chi Minh City	11	0	1	0	1	4
Harare	4	0	0	0	1	1
Buenos Aires	7	0	0	0	0	3
Bengaluru	14	0	0	0	0	2
Johannesburg	9	0	0	0	0	2
New Delhi	21	0	0	0	0	2
Florianópolis	2	0	0	0	0	1
Mérida	9	0	0	0	0	1
Hanoi	5	0	0	0	0	1
Guadalajara	5	0	0	0	0	0
Santa Cruz (Bolivia)	5	0	0	0	0	0
Dodoma	1	0	0	0	0	0
Lusaka	4	0	0	0	0	0

^
Locations are self-reported, and the Bogotá row counts only teams physically at the hub. Counted by self-reported location like the other rows, Bogotá has 12 projects, the same 2 winners and 2 honorable mentions, and the same counts in every percentile column.
^
Winners and honorable mentions as announced by Apart Research; team locations and project placements from their internal co-organizer catalog. The Top 5%, 10%, and 20% columns count the projects each hub placed in that slice of the 216 scored submissions. All 17 hub cities are shown.

Frontpage posts - Effective Altruism forum viewer

The Long (Self-)Correction by Wei Dai

Coefficient Giving just gave GiveWell $1 billion. Where should other donors give now? by Jack Lewars

Tackling Diarrheal Disease in the Democratic Republic of the Congo (DRC) and South Sudan by GiveWell

Jasmine Sun on what the people building AI really believe by 80000_Hours

Epi­sode summary

Want to get up to speed on AI?

The in­ter­view in a nutshell

AI re­searchers ac­knowl­edge the risks, but they’re build­ing anyway

A hand­ful of peo­ple’s be­liefs are shap­ing fron­tier AI

Vice sig­nal­ling is mak­ing it un­cool to be a cau­tious ‘doomer’

Anti-AI back­lash is grow­ing, and it’s not fo­cused strictly on the technology

Without a way to ne­go­ti­ate, AI pop­ulism threat­ens to escalate

AI has a PR prob­lem that mar­ket­ing can’t solve

China’s em­brace of AI is prag­ma­tism, not optimism

Speak­ing to ev­ery side is the se­cret to bal­anced AI reporting

In­sti­tu­tions are wak­ing up — de­mand for AI ideas now out­strips supply

Highlights

Fron­tier AI builders have no good ad­vice for to­day’s teens

Some AI in­sid­ers don’t care whether hu­mans stay in charge

The con­se­quences of mak­ing it un­cool to be cautious

Or­di­nary peo­ple think Sili­con Valley is gam­bling with their lives

Fac­tory work­ers had unions, knowl­edge work­ers have no levers

Sili­con Valley’s utopia is a hard sell

What can poor countries sell in the AI age? by Deena Mousa

Data la­bel­ing and verification