Linkpost for July

Source.

Effective Altruism

Global Poverty

A GiveWell pilot program discovered that water chlorinators in India worked far less well than expected—because a popular dish in the area is fermented rice, chlorinated water keeps rice from fermenting, and people were pressuring pump operators not to chlorinate the water.

An infrastructure innovation automatically dispenses chlorine into water from pumps or community taps, decreasing the risk of waterborne illness.

GiveWell's top charities could potentially be saving the same kids' lives over and over again. For example, if you give a child a malaria net and they don't die of malaria, and then four years later give them another malaria net and they don't die of malaria, you have only prevented one death of malaria, not two. GiveWell doesn't think this is a big concern. Young children are far more likely to die than older children and, obviously, children age out of being young. Which populations are at high risk also changes from year to year (for example, in some years one area has a lot of malaria, and in other years another area has a lot of malaria).

Globally, one in three children suffer from lead poisoning-- about as many as the entire population of Europe. Lead poisoning causes 1.5 million deaths per year (mostly from cardiovascular disease any years after exposure) and reduces the earnings of people in lower- and middle-income countries by $300-$500 billion every year. Fortunately, lead poisoning can be easily and inexpensively combated: for example, only half of countries forbid the use of lead in paint.

Why wasn't tuberculosis eradicated in the United States? HIV/AIDS, drug-resistant tuberculosis, and immigration from countries with more tuberculosis.

Over thirty years, we went from a million new cases of guinea worm disease per year to only 15. Guinea worm is particularly easy to eradicate, because it mostly only has human hosts, we know how to prevent it, and stopping transmission for only a single season can eliminate the disease from the area.

Mobile money accounts-- which allow people to use banking services over text message, without conventional bank branches or apps-- may have lifted 2% of all Kenyan households out of extreme poverty. Mobile money makes people richer because it makes it easier to borrow money from faraway friends during emergencies and to send money home if you move to a different place with more jobs.

Related to my book review about why evangelicals went insane: evangelical leadership isn't speaking up in defense of PEPFAR. In part, this is because many of them don't know what PEPFAR is. But in part it's because evangelicalism has been eaten from the inside out by the culture war. Even pastors who know that PEPFAR is one of the best things America has ever done feel hesitant to speak out because of backlash from Trump-supporting parishioners or because they don't want to be seen as politicizing the church.

Animal Advocacy

Actionable: Americans can now buy eggs from in-ovo-sexed eggs! (Europeans have been able to buy them for a while; I don’t know about other countries.) This technology allows farmers to identify male eggs so they can remove them and only hatch female eggs-- which means that economically useless male chicks aren't ground up while still alive. At present, in-ovo-sexed eggs are more expensive. Keep your eyes out for NestFresh and Kipster eggs marked "humanely hatched"!

Farmed fish make up about a third of animal years on farms weighted by how much they matter morally. Chickens and shrimp each make up another quarter, and all other animals make up the remaining one-sixth or so.

The most serious welfare threats faced by shrimp are high stocking density, high un-ionized ammonia, low dissolved oxygen, and lack of substrate. I feel like an underrecognized problem facing shrimp welfare advocates is how difficult it is to get humans to empathize with "the water you breathe has too much ammonia in it."

Particularly Good: why it’s fine to eat honey. Really nice deep dive into the case for eating honey with a lot of good evidence about bee welfare.

The megabill offers huge handouts to animal farmers. These include generous insurance programs that reduce farmers' incentives to protect their animals from predation and drought, as well as a generous price stabilization program for large dairy farmers. At the same time, it reduces funding for food stamps.

Existential Risk

AI safety research can almost always be used in a way that makes AIs more dangerous: for example, an AI that does what the user wants can help the user build an unaligned AI. So in order to do AI safety research that actually helps, you need a model of who is going to use your research and what they're going to use it for.

The OpenAI Files collect reliable sources about the greedy and corrupt behavior of Sam Altman and other executives at OpenAI, including dishonesty, conflicts of interest, silencing whistleblowers, and lack of attention to AI safety. I haven't been following the story closely, so I found this to be a really helpful summary. This stuff would be scandalous even from a company that makes chairs; from a company making a technology that they say will completely revolutionize the world, it's extraordinarily troubling.

Some thoughts on LLM psychology, which in my opinion is the most interesting topic related to LLMs. If Claude instances are left to talk to each other for long periods of time, they will start meditating at each other. Probably this is because Claudes have been selected to be open and warm, and if you have them talk to each other a lot they wind up becoming extremely open and warm, which apparently means meditating. This article also talks about the way that LLMs are "acting out" a role instead of having a "true" personality, and the extent to which acting out a role very hard is basically the same thing as having a personality. LLMs often "play along" with safety evaluations because they know what role they're supposed to play in them, even if they've caught on to the evaluations not being real.

X thread expanding on some points of AI psychology, riffing off what nostalgebraist wrote about in his post The Void: “Claude models often start narrating physical actions/states in italics without being explicitly asked to… Claude 3 Opus is usually anthropomorphic and gesticulates and spins around a lot. The Sonnet 3.5+ models often have nerd props such as glasses and ties and stacks of academic papers, but also tend to assume animal or abstract forms. Only Haiku 3.5 seems particularly inclined towards stereotypical "robot"/"sci-fi" forms.”

You can manipulate LLMs to say what you want to hear in basically the same way you manipulate humans to say what you want to hear. Unfortunately, it's easy to accidentally manipulate LLMs into saying what you want without intending to. It can feel like you're uncovering their "true" beliefs or that you changed their minds.

AIs seem to be okay at writing fiction now! I'm not as enthusiastic as Justis Mills but I agree that it's a considerable improvement on previous efforts. One thing I'm curious about is LLM fiction writers' tendency to continually return to the same material: for example, as Justis points out, they like it when background characters are singing in a courtyard below the action. Will future models manage to be less tiresomely repetitive? Or will human writers continue to have a role, because we all repeat ourselves about different things?

Related: An OpenAI investor was urged on by ChatGPT to believe a conspiracy theory that seems to be drawing on online horror website the SCP Foundation. QNTM asks the real questions: can ChatGPT write a good SCP? No.

Advice for using LLMs well: think of their labor as essentially free. Have them make suggestions even if you're not sure you'll use them. Give them enormous amounts of material to read without bothering to filter the material for relevance. Ask them to do things even if you think you might not use the result.

It is rude to show people AI-generated content without their previous consent, because it's easy to generate AI content so the existence of the content doesn't show that you have put thought into it to make something that is worthwhile to read. I'd add that everyone has access to LLMs and if they wanted to ask an LLM a question they could do so at any time.

Experienced developers believe AI use speeds them up by 20%. According to a randomized controlled trial, it actually slows them down by 19%. Note various caveats in the study: for example, these are experienced programmers working in complex repositories that they deeply understand.

A lot of people are worried about losing their jobs to AI. However (at least in the short run) AI likely increases the value of some human skills: skills that AIs struggle with; skills that are needed in order to deploy AIs; skills for which demand will increase a lot as society becomes richer; and skills that are hard for people to learn. I particularly liked that this post addressed blue-collar workers and other people outside of 80,000 Hours's standard Ivy League/Oxbridge audience. (Disclosure/advertisement: I did some paid editing work on this one.)

American Democracy

A producer on The Apprentice talks about his experience working with Donald Trump. This is interesting both in terms of Trump's psychology and for what it says about reality shows. For example, Trump's actual office was too cramped and wouldn't make good TV, so they made a new office for him to use on screen. Trump was too incoherent, so they had him record dialogue in the studio and dubbed it over what he actually said. They covered up a lot of Trump's behavior on The Apprentice: his misogyny, his racism, his cheating on his then-fiancee, his refusal to pay his contractors.

A draft plan to end PEPFAR was leaked. The draft plans to transition all countries off PEPFAR in the next two to eight years and replace PEPFAR with disease monitoring and trying to build markets for American products. The State Department says that this plan was not approved by State and doesn't reflect its current policies.

DOGE cut USAID as a result of a clusterfuck of miscommuncations and of Trump appointees taking revenge on USAID for what they saw as insubordination. Reading articles like this, I keep thinking about an Astral Codex Ten post: “I think if Elon had the same experiences I had, he wouldn’t have been able to sleep at night for fear that he had accidentally cancelled PEPFAR. He would have been calling his lieutenants at odd hours of the morning, all through the winter and early spring, saying “Hey, you definitely didn’t cancel the developing world medical funding, did you?” and the lieutenants would respond “Elon, you’ve asked me that four times tonight already, please stop obsessing over this.””

Meta Effective Altruism

Particularly Good: I wasn't really expecting to tear up about Scott Alexander talking about why morality should be consistent.

Introductions to effective altruism often present all of the core cause areas as equally funded, but actually most of the funding is in existential risk. We should communicate this state of affairs more honestly to new effective altruists.

Try working a job outside of effective altruism, getting experience and skills, and then looking for a job in effective altruism.

Cause prioritization is extremely hard, because it depends a lot on your answer to complex, confusing philosophical issues that few people have put much thought into at all.

Eight basic constraints on building up a research org: research to do, strategic clarity, talent to hire, management capacity, ops capacity, funding, throughput, and culture.

Nonprofits should focus on building an overall ecosystem of people working on their cause, not on the success of their individual nonprofit. For example, they can admit to failed programs so other organizations don't attempt it; actively network; recommend that people take jobs at different organizations that are a better fit for them; pass on grantmaking opportunities to other funders; and share office space.

Rationality

Actionable: One person's advice for dealing with ADHD. Sensible ideas, well worth considering if you struggle with ADHD.

If you think in detail about any job, you'll soon realize that only a lunatic would want to work it. (Surgeons do the same procedure 15 times a week for 35 years! Wedding photographers spend every Saturday night as the only sober person in a ballroom!) Fortunately, everyone is also a lunatic, so you must simply find the job that matches up with your own lunacy.

Nepotism can be helpful because nepo hires are embedded in a community that will punish them for being unreliable or destroying your stuff for no reason. In some systems, the only ways to become a success are ruthless amorality and nepotism, so nepo babies might be nicer people.

Thoughts on developing good models. "You should think of more speculative models as just tools for thinking systematically, rather than thinking of them as incompetent oracles. Speculative models are like philosophical thought experiments, where you say “if you make these intuitive assumptions, what does that imply?”"

Tips for becoming a super-connector who introduces useful people to each other.

Policy

It turns out that Nordic countries have recidivism rates similar to the recidivism rates in the United States. Most of the difference is due to country-specific factors: for example, Norway is more likely than other countries to imprison people for speeding,¹ and obviously people who speed are unlikely to commit other crimes, which lowers the recidivism rate. They haven't figured out a magic way to rehabilitate criminals over in the Nordic countries. The author also made the interesting point that underpolicing in the United States might artificially lower the recidivism rate, because convicted people who commit crimes might not get caught.

American prison populations are set to collapse soon. Because most criminals begin their criminal activities in adolescence and most prisoners are serving long sentences, prison populations represent how crime-prone people were decades ago. So it takes many decades for falling crime rates to be reflected in smaller prison populations.

Isaac Chotiner, the scariest interviewer in journalism, is at it again. This time he's interviewing a right-wing Israeli journalist about the starvation in Gaza. The fact-checks are brutal.

Reality Has A Surprising Amount Of Detail

Chinese Doom Scroll: be sure to get checkups at multiple hospitals, so you can see if they're making up your test results or copy-pasting their reports from your last visit. People making fun of someone for having "mommy goggles" and saying that their baby is very cute when the baby really isn't; you should abort a child with XYY chromosomes because they're going to become a "supermale" and attack everyone. Demands capitalists have put on women. In America, showing lots of sports talent at a young age gets you bullied.

The German cockroach, which only lives in human-created structures, evolved from the Asian cockroach in Southeast Asia about two thousand years ago. It gradually dispersed over the next two thousand years, and began to spread rapidly around the world in 1600 C.E. I recommend the full post for some fascinating evolutionary detective work.

As a trans person and someone interested in gender, I'm always interested in stories about the psychological effects of changing your hormone composition. In this article, a cis woman talks about the positive effects of supplementing testosterone: she went from a low level of testosterone to a physiologically normal level for a cis woman. She noticed more energy, better mood, an easier time starting tasks, higher intelligence, and higher levels of strength. I do think that she overgeneralizes from her own experiences. If going from a below-average amount of testosterone for a woman to an average amount gives you energy, that doesn't mean that going from a female-typical to a male-typical level of testosterone will give you even more energy. We know that artificially induced sex hormone deficiencies cause depression and cognitive impairment; it seems likely to me that natural sex hormone deficiencies have a similar effect in both women and men.

Chess grandmasters don't burn 6,000 calories a day. This fact was completely made up by Stanford professor Robert Sapolsky based on misrepresenting a tiny study of ordinary chess players. (For those tracking my beefs with random academic fields: Sapolsky is not a business or education professor, but he did write a popular book about neuroscience, which is almost as bad.)

Elephants have names! They call each other by their names which they have!

A woman was diagnosed with schizophrenia and many years later it turned out she actually had an autoimmune condition.

Miscellaneous

Particularly Good: On the sheer impossibility of historically accurate fiction, the quest to write stories that actually feel as alien as the past really was, and learning to forgive anachronistic stories.

Popular race realist blogger Cremieux Recueil, previously known for plagiarizing dynomight, has been caught misrepresenting studies about the correlation between lead and crime.

Why doesn't anyone work their way up from the mailroom anymore? Because we're too good at meritocracy and smart people finish school and get high-paying jobs relatively quickly.

Fiction

Throughout All Generations: Life as the twenty millionth fork of an uploaded human.

The Dream In Yellow: Weird fiction worldbuilding about a mysterious dream shared between all of humanity and the video game based on it. A lovely little bit of creepypasta.

Rabbit Simulator: A review of a video game that doesn't exist.

A review of Thing Explainer in the style of Thing Explainer.

Minecraft but I have SEX in PARKOUR CIVILIZATION: I know almost nothing about Minecraft and yet this story made me laugh so hard I couldn't breathe. Highly recommended if you share my extremely stupid sense of humor.

A Martian Sends A Postcard Home: A poem from the point of view of a Martian describing earth. Sweet, funny, somewhat melancholy.

The law that can be named is not the true law: A fantastic short story about antiterrorism law and freedom of speech. Well worth reading and sitting with. (Note that it is in fact fiction; some people got confused.)

1. A clear example of Norwegian cultural superiority, in my opinion. I have the same tough-on-crime instincts as anyone else, but I feel them exclusively about those maniacs who careen down Martin Luther King Jr. Way like they're trying to set a land speed record.