Linkpost for May
Effective Altruism
Global Poverty
This article [Development] about the history of GiveDirectly doesn’t have anything new for people who have been following them closely for many years, but it’s a great summary for people who are new to this wonderful organization.
Animal Advocacy
The evidence for the sentience of Litopenaeus vannamei (the most commonly farmed species of shrimp) is extremely sparse, with some lines of evidence pointing towards the species not being sentient [Rob Velzeboer]. Often, people are reasoning based on research done on completely different taxa, like crabs and lobsters. We need to do more research so that we can understand whether the half a trillion shrimp individuals we farm each year are capable of feeling pain.
Existential Risk
The most rational belief to have about when AGI will happen is not “it will happen soon” or “it will happen in a long time” but “no one knows when it’s going to happen” [Forethought]. So we need to prepare for the world where it comes soon, while also making plans that will only pay off in the worlds where AI takes longer.
Be careful in how you interpret AI benchmarks [Effective Altruism Forum]. If one model scores 25% on a benchmark and the other scores 50% on a benchmark, it’s easy to go “the second model is twice as good”, but in fact this is meaningless because most benchmarks don’t have natural units. As a comparison, if I scored 100% on my calculus test and you scored 50%, you wouldn’t assume I’m “twice as good at math as you”-- calculus tests only cover a small portion of math, you have to be quite good at math to be in a calculus class in the first place, and maybe you just didn’t get derivatives and so all your errors are downstream of one mistake rather than a holistic lack of skill at math.
What can a superintelligent being actually do? [Dynomight]. Could it solve all open problems in physics? (Probably not.) Can it cure cancer? (No.) Could it persuade anyone of any claim? (No idea.)
AI unemployment and AI extinction are two sides of the same risk [world spirit sock stack]: we will create agents more competent than humans, and those beings’ agendas will win out over humans’ agendas.
LLMs might make it easier for people to learn the biology they’d need to do a bioterror attack. But it’s not clear that they help people with the most important kinds of knowledge (like physically making the bioweapon) or that knowledge is what’s keeping people from doing bioterror attacks [Transformer]. Biological design tools, on the other hand, use AI to actually make new biological discoveries like figuring out a protein’s structure; they could help people develop novel viruses that are far scarier than any virus we’ve ever seen. However, they require a lot of biology expertise to use. If LLMs and biological design tools are integrated, making it easier for normal people to use them, we might have A Real Problem. Also, a lot of keeping people safe from AI-enabled bioterror attacks is doing the same stuff we should be doing to keep people safe from normal bioterror attacks, like not letting people buy dangerous DNA sequences.
Claude has filled in his underspecified personality with a reliably consistent Dude [Guive’s substack]. “A good heuristic for predicting Claude’s tastes is to think of it as
playing the character of an idealized liberal knowledge worker from Berkeley. Claude can’t decide if it’s a software engineer or a philosophy professor, but it’s definitely college educated, well-traveled, and emotionally intelligent. Claude values introspection, is wary almost to the point of paranoia about “codependency” in relationships, and is physically affected by others’ distress.”
Particularly Good (and a nice contrast to the previous): Recent LLM personas feel less “well-defined” than older LLM personas [Less Wrong]: that is, they seem less like a specific, predictable Kind of Guy and more like an incoherent set of tendencies that the model applies when it seems relevant.
Claude seems to perform worse if you’re a morally bad person [Natalie’s Substack].
Noise pollution from data centers is a real concern. “Infrasound”—sound too low-frequency for humans to hear—isn’t [Andy Masley]. Concerns about infrasound are pseudoscience used to attack wind turbines and other important pieces of infrastructure. We can talk about the legitimate environmental concerns with data centers, including noise pollution, without pseudoscience.
I enjoyed this article about how to think about s-risks (that is, about potential futures that contain previously unprecedented amounts of suffering) [Effective Altruism Forum]. Historically, concern about s-risks has been closely associated with negative utilitarianism, but actually people with many different moral systems ought to want to avoid horrifying dystopian futures. And much work to prevent s-risks likely prevents other kinds of bad futures-- for example, promoting cooperation within society, building institutions that are resilient to sadistic or malevolent people in power, and developing AI that avoids very harmful behaviors.
American Democracy
Trump isn’t going to be able to stop the midterms from happening and probably won’t be able to outright steal them, but he’s laying the groundwork to interfere with the election in other ways-- for example, by sending ICE or the National Guard to reduce voter turnout in key swing neighborhoods [Silver Bulletin]. “[T]he cynical interpretation is that he knew these weren’t going to pass court muster, but he’s just trying to stir up discontent around the election and create a premise for saying “Hey, I tried to require proof of citizenship, I tried to clean up mail voting, but the courts didn’t let me. And the election went forward and Democrats won, so it’s fraudulent.” And when the courts rule against any attempts to overturn the results, he can continue to say they stopped him, and delegitimize the election.”
Other Causes
Trump threatening to start a nuclear war with Iran was dangerous because, most of the time, threatening nuclear war gets you what you want without anyone dying, and then sometimes you fuck up and million or billions of people die [Silver Bulletin]. Even though this time no one died, Trump’s behavior is still extraordinarily reckless.
Meta Effective Altruism
It’s quite likely that our society is undergoing an ongoing moral catastrophe [Effective Altruism Forum].
I was delighted by this article about someone convincing the Tantric Retreat Center to pledge to donate to GiveWell Top Charities [Effective Altruism Forum]. Apparently effective altruism has Yang Energy which holistically balances the Yin Energy of the retreat! This is ideal effective giving outreach. You may not like it but this is what peak performance looks like.
Some people do hard-to-monetize work and get money from a patron (understood broadly to include grants from grantmakers). At first, when your taste is bad, you should defer to more competent patrons, but as your taste develops, it’s really important to find a patron who’s aligned with the work that you do [Less Wrong].
