Linkpost for June
Recent fiction publications: four sex stories; I don’t love you in New York City.
Effective Altruism
Global Poverty
Development Innovation Ventures was a USAID program providing flexible capital for innovative global development projects. Like much of USAID, it closed in early 2025. Fortunately, a group of funders led by Coefficient Giving stepped in to help DIV spin out into a nonprofit [Inside Philanthropy].
GiveWell recently stopped funding Dispensers for Safe Water upon discovering that the charity reached 40-70% fewer people than its own monitoring suggested. I really liked this deep dive into what went wrong and how we can avoid monitoring problems like this in the future [Effective Altruism Forum]. In essence, while charity evaluators like GiveWell are very rigorous about determining whether a program works in principle, they can be pretty vibes-y about monitoring how well it’s being implemented. Check out GiveWell’s response in the comments, which I found very reassuring.
Related: Several anonymous people in global poverty, including several GiveWell grantees, discuss the Evidence Action situation [Effective Altruism Forum]. They agree that GiveWell needs to independently verify results, and that we should spend as much energy monitoring implemention as we do figuring out whether the program works in the first place. They add that GiveWell needs to pay more attention to the cost side of a cost-effectiveness analysis-- variance in the cost of a program can often swamp variance in its effects. Furthermore, organizations like Evidence Action that implement many programs may well be worse at implementation than organizations which focus on getting really good at a single intervention.
Corporations can be more important for lifting the global poor out of poverty than foreign aid [In Development]. Every rich country today became rich through capitalism, not aid. A paycheck is a sustainable “cash transfer” that happens every month and is funded by customers, not donors. Export-focused firms aren’t limited by the global poor’s inability to buy things. By selling goods on the global market, corporations are a permanent solution to global poverty.
Long lives and falling fertility rates mean that developed countries will have a very bad ratio of working-age adults to dependent elders. Immigration from Africa, which has higher fertility rates, could solve this problem [The Economist]. Although Africans are enthusiastic about emigrating, most countries are reluctant to take many African immigrants. Nevertheless, the numberof African immigrants has been increasing dramatically.
Animal Advocacy
Making it easier to build factory farms can actually improve animal welfare, if it means that more animals are raised in developed countries with stricter animal welfare regulations [Social Problems Are Like Maths].
Existential Risk
Longtermists have become less focused on human extinction and more focused on ensuring a good future [Effective Altruism Forum]. Partially, this is because many people believe human extinction is less likely than they used to think (for example, because AI alignment is less hard than they thought). Partially, this is for sociological reasons (for example, because people are building AI-safety coalitions with people who don’t care much about the long-term future).
The public health system’s reaction to hantavirus shows us that we learned nothing from the covid pandemic [The Argument].
Some thoughts on catastrophic biological risk [Effective Altruism Forum]. Most biological threats that might cause human extinction are bioengineered and target humans (rather than agriculture or the environment). Since there are only a small number of ways humans can get sick, we can vastly reduce biorisk by deploying countermeasures to block those pathways (e.g. PPE, air filters). Most risk is from non-state actors doing risky actions deliberately, such as terrorists, so biosecurity should focus on keeping it hard for actors without many resources to make biological weapons (e.g. DNA synthesis screening).
Techniques for ensuring that AIs don’t help bad actors build bioweapons [Lennart Justen]: training AIs to refuse to help make bioweapons; detecting and investigating potential attempts to make a bioweapon; only allowing trusted customers to use cutting-edge biological models; removing bioweapons knowledge from the AI’s knowledge base; and evaluating AIs’ biological capabilities and the associated risks.
Dylan Matthews reflects on his 2015 piece criticizing effective altruism for its focus on AI—a piece which, he admits, aged very poorly [Dylan Matthews]. He identifies his mistakes as too much trust in credentialed institutions and a bias against things that sounded too “science fictional.”
We don’t know how to measure whether a job is being automated by AI [The Argument]. Usage data is limited and hard to interpret. Interviewing people about how they use AI is reliable but slow. Scraping job postings gives us information about how employers are thinking about the future, but can be noisy. Price data helps us analyze consumer behavior, but we don’t track the price of most of the services (like software or the law) most affected by AI.
The first papal encyclical about AI was substantially written by an AI, probably Claude [The Linchpin].
Data centers aren’t a serious cause of land use problems: by 2028, they will use up 3.5 times the amount of land used by Christmas trees [Andy Masley]. Worry about ethanol requirements.
Automation of AI R&D may not lead to superintelligent AI [Effective Altruism Forum]. To train an AI to do tasks, you need data. For most tasks, real data in large quantities isn’t available, and simulated data will be insufficient-- you have to know what good performance is to create simulated data! We should expect AI superintelligence to occur relatively slowly, as AI is integrated throughout the economy.
The Extropians were a highly influential group of futurists in the 1990s whose members include Marvin Minsky, Eliezer Yudkowsky and Nick Bostrom. How well did they do in 1995 on predicting the world in 2023? They ranged from 50% correct to ~4% correct [Maximum Progress]. They were massively overoptimistic about genetic engineering, cryonics, and nanotech, although correct about the Internet.
Key questions for strategy around digital minds [Outpaced]. A nice summary of all the things we don’t know about this emerging field.
Meta Effective Altruism
Rethink Priorities has created its Cross-Cause Fund, which asks questions about your values and empirical beliefs and allocates your donation between effective funds based on the results [RP Research Digest]. If we get any more meta the effective altruism movement will wind up in Inception.
The charity incubator AIM has proposed a new taxonomy for charities [Effective Altruism Forum]. Charities differ by both target outcome and mechanism (e.g. direct service provision, lobbying the government). Two charities that both lobby the government may well have more in common than two charities that both work on helping wild animals, even if they lobby the government about different issues. Some charities are more execution-dominant (i.e. player vs. environment, such as providing bednets) and some charities are more persuasion-dominant (i.e. player vs. player, such as lobbying for tobacco regulation). Some charities involve implementing a concrete intervention that we know works, while other charities involve trying to figure out some approach that might work to an important problem.
Effective altruism teaches that you should maximize the good, but maximizing the good is dangerous; it’s usually better to be a reasonable, virtuous person with pluralistic values [Effective Altruism Forum].
How to not procrastinate on giving away your millions [Effective Altruism Forum]. I don’t know how to summarize this in a way that doesn’t sound like a joke. It is advice for the problem of failing to give away millions of dollars because you keep putting it off, which my secret sources inform me is weirdly common.
People often decide to prioritize particular causes and then give to the most cost-effective opportunities within each cause. But this is a bad idea because (a) cause boundaries are very arbitrary (b) sometimes the best intervention within a particular cause area is more cost-effective than the best intervention within a different cause area, even if the first cause area is less cost-effective on net [Marcus A. Davis].
