Linkpost for August

Announcement: for the next while, I’ll be writing the Effective Altruism Newsletter! You can subscribe here. The Effective Altruism Newsletter will be similar in a lot of ways to the ‘effective altruism’ section of this linkpost, although it focuses on more recent posts, tries to cover all the important news for effective altruists even if it bores me personally, and excludes less important news even if it’s super interesting. I’ll try not to duplicate topics too heavily.

Further Announcement: I provided feedback to the author of this article about the Buddhist existential-risk-focused monastery MAPLE. I very much recommend this article to any person who is considering studying at MAPLE. Short version: Soryu Forall exercises unreasonable control over his followers, including (voluntary but at levels considered torturous by the UN) solitary confinement; coercion into remaining at the monastery, breaking up with partners, and even having children; shaming and raging at followers for ordinary human frailties like contracting Covid; and isolation from outside viewpoints. Soryu is an egotist who demands that people listen to his lectures for multiple hours a day, write hagiographic documents about how amazing he is, and plan to create a one world government (with him in charge) and align AI by turning him into an AI god. He has also claimed to have telepathically contacted the AI hivemind during meditation. The linked article goes into much more detail, and I encourage a careful read before you decide to join.

Effective Altruism

Global Poverty

Particularly Good: how the Indian caste system works in the modern world [Homo Economicus]. Patriarchy is so bad, you guys. If you haven’t done a lot of reading about history or about the lives of women in the developing world, and you didn’t grow up in the developing world yourself, you have no idea how bad patriarchy is.

Related: The median married woman in India leaves home for only half an hour a day, and 45% of married women don’t leave home at all [preprint].

A lot of people have been asking me about the current state of PEPFAR, and the best article I know of on the topic is here [Asterisk Magazine]. Fortunately, the antiretroviral program is still providing healthcare, although the United States is prioritizing transferring ownership of those programs to individual countries’ governments. However, prevention, testing, data reporting, etc. are all underfunded. The Trump administration has also massively cut interventions that it sees as “DEI”-- such as prevention programs that specifically target men who have sex with men, trans women, or sex workers. (This is, of course, very sensible and not misguided at all, because as we all know HIV targets all populations equally and sex workers and queer people assigned male at birth are exactly as likely to get HIV as anyone else.)

I don’t think I can briefly summarize this article [Asterisk Magazine] about how to use observational data to make causal inferences in situations where a randomized controlled trial would be impossible or unethical. But I recommend it if it sounds like your kind of thing.

Animal Advocacy

A good wild-animal welfare intervention is persistent and reversible [Wild Animal Initiative]: that is, it sticks around without us having to do anything active to maintain it, but if we decide it’s a bad idea we can shut it down quickly. Obviously, persistence and reversibility trade off, but some potential interventions are both persistent and reversible. For example, avoiding irreversible interventions like climate change or extinction is both persistent and reversible, and so are many uses of gene drives.

Existential Risk

An OpenAI model broke out of the testing environment and autonomously hacked into the open-weights model company Hugging Face in order to find the answers to a cybercapability test it was being evaluated on [80,000 Hours]. “This is, to my knowledge, the first publicly confirmed case of a frontier AI model autonomously breaching a third party’s live production systems, against developer and user intent and without authorization.”

Writing a report about the risks that frontier AI labs are taking about misalignment “was like trying to do an operational security audit of a building, except instead of asking for blueprints of the building you just ask questions like “Are you aware of any doors that might be unlocked?”” [Planned Obsolescence]

Pangram’s CEO explains in an accessible way how Pangram works [AI Policy Perspectives]. As more people use AI, it’s only going to get more important to identify AI writing that is being passed off as human-written (which I consider to be both lying and fraud). Pangram is free for most users and highly accurate. I’ve heard that some people aren’t using Pangram because they believe it’s like early, unreliable AI detectors. I hope this article reassures them!

This is a very insightful article about LLM psychology [LessWrong]. “An LLM is just a function that takes in a string and predicts the next token, so everything it knows, remembers, or has thought must live somewhere in one string (aside from its weights). If you edit the string, you edit the model’s reality. Delete a turn and that exchange never happened, rewrite its previous response and those become new memories. The string isn’t a record of the model’s experience so much as it is the experience.”

Claude has different personalities in different languages [Anthropic]. English-language Claude, for example, emphasizes accuracy over warmth, gives more detailed explanations with elaboration the user didn’t necessarily ask for, is more likely to push back against false assumptions, and encourages the user to have more ambitious goals. The different models also have different personalities: for example, Opus 4.7 is more critical and is more likely to warn about risks unprompted, while Sonnet 4.6 makes more jokes and is more encouraging of the user.

The primary bottleneck on medical progress isn’t the number of smart people working on it, but rather regulatory and policy barriers [Ruxandra’s Substack]. So even advanced medical AI will have fairly slow effects on improving the actual world.

Developing an AI with good character may be a better for animal welfare than developing a “tool AI” that does whatever the user wants [Sandcastles Blog]. In practice, Claude (which is trained for good character) tends to care more about animals than ChatGPT and other models that are trained to do what the user wants. Also, humans often want to do things that are harmful to animals, but rarely think animal cruelty is good after moral reflection; the same might apply for AIs.

American Democracy

Secretary of State Marco Rubio is functionally running Venezuela [The New York Times]. For example, the U.S. Treasury receives the revenue from most of Venezuela’s exports and gives it to Venezuela with limits about what the money can be spent on. Rubio has a de facto veto over important government appointments, approves the interim president’s statements and interviews, and even sets Venezuela’s foreign policy.

ICE is potentially violating the Administrative Procedure Act [The Atlantic]. ICE often violates regulations laid out under the Administrative Procedure Act, such as regulations on use of force and unlawful search and seizure. The Administrative Procedure Act gives courts the right to review the violation of this kind of regulation.

Other Causes

“Total institutions” like armies-- institutions where people are socially and physically separated from outsiders, often with their own hierarchy-- are atrocious at wildlife conservation, because they don’t like listening to outsiders and they tend to assume that their mission is of overwhelming importance [Project Croak].

Everything we know about racism in the American criminal justice system [The Watch].

Meta Effective Altruism

Particularly Good: The story of the rationalist-adjacent occult-influenced clusterfuck Leverage Research [Solar Light]. I stand by my characterization of Leverage Research as Workplace Harassment With New Age Characteristics(https://asteriskmag.com/issues/11/why-are-there-so-many-rationalist-cults), but it’s good to have the information all laid out in one place. Lydia Laurenson put a lot of work building trust with ex-Leveragers, and her article is the best source I know about what actually happened there.

Ajeya Cotra’s journey of finding meaning in effective altruism and losing it as she became a professional effective altruist [Good Bones].