The Nonlinear Library

The Nonlinear Library

By The Nonlinear FundEducation
Download on the App Store

All The Nonlinear Library episodes

  • EA - FAQ on the relationship between 80,000 Hours and the effective altruism community by Ardenlk
    Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: FAQ on the relationship between 80,000 Hours and the effective altruism community, published by Ardenlk on June 24, 2023 on The Effective Altruism Forum.
    As part of 'strategy fortnight' (and in part inspired by this post) I decided to write this short post clarifying the relationship, as I see it, between 80,000 Hours and the EA community. I chose these questions because I thought there might be some people who care about the answers and who would want to know what (someone on the leadership team at) 80,000 Hours would say.
    Is 80,000 Hours's mission to build the EA community?
    No — our aim is to help people have careers with a lot of social impact. If the EA community didn't exist, we could still pursue our mission.
    However, we count ourselves as part of the EA community in part because we think it's pretty great. It has flaws, and we don't blanket recommend getting involved to our readers (a recommendation we feel better about making widely is to get involved in some kind of community that shares your aims). But we think the EA community does do a lot to help people (including us) figure out how to have a big positive impact, think through options carefully, and work together to make projects happen.
    For that reason, we do say we think learning about and getting involved in the EA community could be a great idea for many of our readers.
    And we think building the EA community can be a way to have a high-impact career, so we list articles focused on it high up on our problem profiles page and among our career reviews.
    Both of these are ways in which we do contribute substantially to building the effective altruism community.
    We think this is one of the ways we've had a positive impact over the years, so we do continue to put energy into this route to value (more on this below). But doing so is ultimately about helping the individuals we are writing for to increase their ability to have a positive impact by getting involved, rather than to benefit the community per se.
    In other words, helping grow the ea community is part of our strategy for pursuing our mission of helping people have high-impact careers.
    Does 80,000 Hours seek to provide "career advice for effective altruists"?
    Somewhat, but not mostly, and it would feel misleading to put it that way.
    80,000 Hours focuses on helping a group much larger than the (current) EA community have higher impact careers. For example, we estimate the size of the group we are trying to reach with the website to be ~100k people — which is around 10x larger than the EA community. (For context, we currently get in the range of 4M visitors to the website a year, and have 300k newsletter subscribers.)
    Some of the people in our audience are part of the EA community already, but they're a minority.
    One reason we focus so broadly is that we are trying to optimise the marginal counterfactual impact of our efforts. This often translates into trying to focus on people who aren't already heavily involved and so don't have other EA resources to draw on. For someone who hasn't heard of EA or who has heard of it but doesn't know much about it, there is much lower hanging fruit for counterfactually helping them improve the impact of their careers. For example, we can introduce them to well-known-within-EA ideas like the ITN framework and cause selection, or particularly pressing issues like AI safety and biosecurity, as well as the EA community itself. Once someone is involved in EA, they are also more likely and able to take advantage of resources that are less optimised for newer people.
    This is not an absolute rule, and it varies by programme – for example, the website tends to focus more (though not exclusively) on 'introductory' materials than the podcast which aims to go more in-depth, and one-on-one advising tries to tailor their discussio...
    9 min
  • EA - On focusing resources more on particular fields vs. EA per se - considerations and takes by Ardenlk
    Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: On focusing resources more on particular fields vs. EA per se - considerations and takes, published by Ardenlk on June 24, 2023 on The Effective Altruism Forum.
    Epistemic status: This post is an edited version of an informal memo I wrote several months ago. I adapted it for the forum at the prompting of EA strategy fortnight. At the time of writing I conceived of its value as mostly in laying out considerations / trying to structure a conversation that felt a bit messy to me at the time, though I do give some of my personal takes too.
    I went back and forth a decent amount about whether to post this - I'm not sure about a lot of it. But some people I showed it to thought it would be good to post, and it feels like it's in the spirit of EA strategy fortnight to have a lower bar for posting, so I'm going for it.
    Overall take
    Some people argue that the effective altruism community should focus more of its resources on building cause-specific fields (such as AI safety, biosecurity, global health, and farmed animal welfare), and less on effective altruism community building per se. I take the latter to mean something like: community building around the basic ideas/principles, and which invests in particular causes always with a more tentative attitude of "we're doing this only insofar as/while we're convinced this is actually the way to do the most good." (I'll call this "EA per se" for the rest of the post.)
    I think there are reasons for some shift in this direction. But I also have some resistance to some of the arguments I think people have for it.
    My guess is that
    Allocating some resources from "EA per se" to field-specific development will be an overall good thing, but
    My best guess (not confident) is that a modest reallocation is warranted, and
    I worry some reasons for reallocation are overrated.
    In this post I'll
    Articulate the reasons I think people have for favouring shifting resources in this way (just below), and give my takes on them (this will doubtless miss some reasons).
    Explain some reasons in favour of continuing (substantial) support for EA per se.
    Reasons I think people might have for a shift away from EA per se, and my quick takes on them
    1. The reason: The EA brand is (maybe) heavily damaged post FTX — making building EA per se less tractable and less valuable because getting involved in EA per se now has bigger costs.
    My take: I think how strong this is basically depends on how people perceive EA now post-FTX, and I'm not convinced that the public feels as badly about it as some other people seem to think. I think it's hard to infer how people think about EA just by looking at headlines or Twitter coverage about it over the course of a few months. My impression is that lots of people are still learning about EA and finding it intuitively appealing, and I think it's unclear how much this has changed on net post-FTX.
    Also, I think EA per se has a lot to contribute to the conversation about AI risk — and was talking about it before AI concern became mainstream — so it's not clear it makes sense to pull back from the label and community now.
    I'd want someone to look at and aggregate systematic measures like subscribers to blogs, advising applications at 80,000 Hours, applications to EA globals, people interested in joining local EA groups, etc. (As far as I know as of quickly revising this in June, these systematic measures are actually going fairly strong, but I have not really tried to assess this. These survey responses seem like a mild positive update on public perceptions.)
    Overall, I think this is probably some reason in favour of a shift but not a strong one.
    2. The reason: maybe building EA per se is dangerous because it attracts/boosts actors like SBF. (See: Holden's last bullet here)
    My take: My guess is that this is a weak-ish reason – though I...
    13 min
  • EA - EA’s success no one cares about by Jakub Stencel
    Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: EA’s success no one cares about, published by Jakub Stencel on June 24, 2023 on The Effective Altruism Forum.
    Status
    While this is part of EA Strategy Fortnight, my intention is to focus more on what effective altruism did well, rather than what it should do. At the same time, my hope would be that the post provides the community with at least some valuable context to the discussions about the path forward.
    This post will be very subjective and a lot less thought-through than what I am usually comfortable sharing, although built on ~13 years of experience in the field. Some details may be off, for example by memory distortion or indirect testimonies. Nevertheless, it’s truthful to my internal models – when I walk and talk to my dog about this kind of stuff, you can expect my mind to go in the same direction as this post.
    Context
    Recently, there has been a lot of attention directed at effective altruism. Some was external, but, from my perspective, most of it came from within the movement. My interpretation was that at least a portion of it was built on feelings of anxiety, doubt, and maybe some anger or fear. Of course, a lot of concerns seemed to me legitimized by what was happening or what we were discovering.
    In some way, I was worried about the community I identify as part of, but at the same time, there was this feeling of appreciation that we can go together through a crisis. It’s a lesson for a young movement, and experience is invaluable. Just like it’s better to learn hard lessons about life as a teenager than an adult, of course ideally with not much harm involved.
    The energy spent on inward focus felt encouraging, even though I disagreed with a chunk of the opinions. After all, some of the values of effective altruism I’m the most optimistic about are openness to criticism, intellectual humility, and truth-seeking. But the more external and internal takes I was reading, the more something seemed off. Something was missing.
    It felt one-sided. There was almost no mention of successes and wins – some appreciation of what this very young and weird movement managed to achieve in such a short period of time. Maybe I shouldn’t expect this in adversarial pieces about EA, and maybe it was implied when people were making criticism internally, but it still didn’t feel fully right to me.
    It felt like we all take effective altruism for granted. There was not much gratitude in the air.
    Maybe one can argue that EA hasn’t done much. While I have my strong intuitions on the counterfactual impact of EA in many areas, in the end I don’t feel fully qualified here, so I would prefer to defer. Yet, I’m confident that there is at least one success we should celebrate, and it’s very much absent from the discourse – making historical progress for animals.
    Animal advocate’s lens on effective altruism
    This is my take on the short path of effective altruism’s impact on animals. Please note that I came from the part of animal advocacy that is closely aligned with effective altruism, so I’ll be biased, and it’s good to expect that reasonable people will disagree with me in at least some parts. Additionally, Open Philanthropy, which is crucial here, is the main funder of my organization. But if bias is present in this summary, I’m very confident it’s rather due to the alignment with the core premises of effective altruism rather than a conflict of interest.
    Funding
    While the animal advocacy movement has never been big, it has always had very active and dedicated activists. People have always been willing to sacrifice a lot to make a difference for animals. Breaking in or getting hired at farms to document the sickening reality of animal suffering or rescue animals, being abused by police, working for years for free or minimal pay, and using any personal funds available to produce necess...
    15 min
  • LW - How tall is the Shard, really? by philh
    Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: How tall is the Shard, really?, published by philh on June 23, 2023 on LessWrong.
    According to Wikipedia, the Shard (the tallest building in the UK) stands "309.6 meters (1,016 feet) high". I put this in my Anki deck as "Height of the Shard / 310m", but I was saying "height" to mean "tallness" (because I don't much like that word) and I had assumed Wikipedia was using it the same way. So I thought the Shard was 310 m tall.
    But according to Bron Maher in Londonist,
    most sources put the building’s tallness - that is, its length base-to-tip - at about 306 metres. The last three metres come from the height of the ground on which the Shard is built. So while its height is indeed 309 metres above sea level, the Shard is only 306 metres tall.
    Is that so? I was curious enough to investigate1, but it turns out I don't really know how to.
    (Also: Maher says the Shard is 306 m, not 309 m, not "not 310 m". And he gets 309 from Wikipedia, quoting it on the list of highest points in London, which talks about "the 309 m (1,014 ft) tall Shard". Where did those extra 60 cm go? The article on the Shard itself also has that number, in the "records" section near the bottom, but the infobox on the right has 309.6/1,016. So is it 306 m, 309 m, or 309.6 m?)
    The boring investigation
    The first thing to do2 is to look at Wikipedia's sources. It doesn't have one on the list of highest points, and it doesn't have one in the "records" section. The infobox does have a source for the height, which is on skyscraperpage.com. It lists the spire at 1,016 ft or 309.7 m, yet another number! There are various drawings, all giving the roof at 304.2 m and the spire at 309.7 m. (1,016 ft is 309.68 m, but 309.6 m is 1015.75 ft and 309.7 m is 1,016.08 ft. So if you convert 309.6 m to feet and back to meters, you could get 309.7 m. But wikipedia shouldn't be taking 309.6 m from that page.)
    Skyscraperpage lists the source for the height as http://www.the-shard.com, and the about page on that site says: "How tall is The Shard? The Shard is 309.6 metres, or 1,016 feet, high and is Western Europe's tallest building." You'd certainly expect the Shard's official webpage to be right about this, but also it says "high" not "tall" so it could be a misdirect.
    Next3, does Maher have any sources? For 306 m tallness he just says "most sources" give that figure. For 3 m above sea level, I think that comes from its height at the top. He quotes someone at Ordnance Survey saying the top of the Shard is 308.9 m above sea level, which gives around 3 m for the base.
    Most of my research was just googling things. Search terms included "how tall is the shard", "height above sea level of the shard", "the shard architectural drawings", "uk contour lines" and "ctbuh the shard" (inspired by one of the earlier results). Here are some of the things I found.
    The London Pass: "The Shard is 306 metres tall, however if you measure all the way up to the tip, it's 310 metres". I don't know what point below the tip it's talking about. It could be roof versus spire, but the skyscraperpage drawings give a different roof height.
    The Skyscraper Center: height ("measured from the level of the lowest, significant, open-air, pedestrian entrance") is "306 m / 1,004 ft", both "to tip" and "architectural". (The difference being things like flagpoles and antennae, which the Shard doesn't have. The highest occupied floor is "244.3 m / 802 ft".) Seems like the kind of site that knows what it's talking about. Doesn't cite sources, but the drawing they have is labeled a CTBUH drawing, and that also sounds like the kind of organization that knows what it's talking about, giving me a new google search term.
    SkyscraperCity: says both "309m" and "310m" with no apparent shame. But also, that link points to page 833 of 1385 of a discussion thread4, and at this partic...
    15 min
  • EA - Crisis Boot Camp: lessons learned and implications for EA by Nicole Ross
    Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Crisis Boot Camp: lessons learned and implications for EA, published by Nicole Ross on June 24, 2023 on The Effective Altruism Forum.
    Over the last 7+ months at work, I've needed to handle or support several crises (I'm on the US EV board and, in normal times, head of the community health and special projects team). It's been a crisis-handling boot camp, so I want to share my lessons learned.
    I expect to learn more in the coming months, and it's plausible that longer-term ramifications could change my lessons learned, but I wanted to share my reflections at this point.
    My reason for writing this: I generally think that at least some of us (maybe many of us) may go through many more crises and that the world has a decent chance of getting even crazier due to AI. I want us to learn from crises and make updates to be better prepared for next time. I'm worried about people returning to normal without making relevant updates on many levels. Hopefully, some of my lessons learned will contribute to people making updates and folks being more prepared next time!
    Handling yourself in a crisis:
    Expand your thinking
    1. Hold multiple hypotheses at once
    Generally, people struggle to hold more than one or two hypotheses simultaneously, and this struggle seems even stronger during a crisis.
    The world is complicated, and being confident about what's true is tough. When people make plans, though, they often focus only on the hypothesis that they think is most likely, or at best, their top two hypotheses.
    For a hypothetical example, imagine the following scenario:
    You're working on getting life-sustaining and valuable resources to your allies in a place with a lot of organized crime. Pretty frequently, your supplies are stolen. You suspect Person X is behind it. One day he disappears, and a lot of supplies are missing. You need to move to the next location to get the supplies in the right hands. What should you do?
    Create multiple hypotheses! You are assuming X is behind the theft. In that world, you might want to move on without him and think through things like "Does he know where we're going to be next? Is he going to steal more? How should we update our security measures?".
    Other hypotheses you should consider:
    Maybe he was kidnapped, in which case it would be pretty shitty to leave him.
    Perhaps he was working with someone else. Are they on your team too?
    It could be a coincidence, and he didn't show up for another reason (e.g., incompetence, sickness).
    Each of these hypotheses has different likelihoods; sometimes, hypotheses are not mutually exclusive. E.g., X might be behind the theft and working with others on your team; X may be incompetent and indirectly connected to the theft.
    In my experience, people focus on only 1-2 hypotheses and hold those too tightly, even if they give lip service to multiple hypotheses. You should be looking for evidence for and against many different hypotheses and endeavor to track updates to many of them. Two particularly salient implications of this are:
    If you don't hold multiple hypotheses at once, you might jump to a conclusion, which makes it less likely you make the best decision.
    If you don't consider what you'd do in different worlds, you miss cheap and/or important ways to mitigate harm or realize the upside.
    How to do it: I've found it helpful to literally write down several hypotheses. Then I take a step back and ask myself:
    Would I be surprised if none of these are true?
    What worlds are these hypotheses neglecting to consider?
    What's missing?
    Usually, there are more relevant hypotheses to write down and track.
    2. Think beyond black-and-white binaries
    It's easy to default to thinking in black-and-white binaries when the array of paths available to you is way more expansive.
    For example, imagine that you have to navigate a disagreement bet...
    24 min
  • EA - Correctly Calibrated Trust by ChanaMessinger
    Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Correctly Calibrated Trust, published by ChanaMessinger on June 24, 2023 on The Effective Altruism Forum.
    This post comes from finding out that Asya Bergal was having thoughts about this and was maybe going to write a post, thoughts I was having along similar lines, and a decision to combine energy and use the strategy fortnight as an excuse to get something out the door. A lot of this is written out of notes I took from a call with her, so she get credit for a lot of the concrete examples and the impetus for writing a post shaped like this.
    Interested in whether this resonates with people's experience!
    Short version:
    [Just read the bold to get a really short version]
    There’s a lot of “social sense of trust” in EA, in my experience. There’s a feeling that people, organizations and projects are broadly good and reasonable (often true!) that’s based on a combination of general vibes, EA branding and a few other specific signals of approval, as well as an absence of negative signals. I think that it’s likely common to overweight those signals of approval and the absence of disapproval.
    Especially post-FTX, I’d like us to be well calibrated on what the vague intuition we download from the social web is telling us, and place trust wisely.
    [“Trust” here is a fuzzy and under-defined thing that I’m not going to nail down - I mean here something like a general sense that things are fine and going well]
    Things like getting funding, being highly upvoted on the forum, being on podcasts, being high status and being EA-branded are fuzzy and often poor proxies for trustworthiness and of relevant people’s views on the people, projects and organizations in question.
    Negative opinions (anywhere from “that person not so great” to “that organization potentially quite sketch, but I don't have any details”) are not necessarily that likely to find their way to any given person for a bunch of reasons, and we don’t have great solutions to collecting and acting on character evidence that doesn't come along with specific bad actions. It’s easy to overestimate what you would know if there’s a bad thing to know.
    If it’s decision relevant or otherwise important to know how much to trust a person or organization, I think it’s a mistake to rely heavily on the above indicators, or on the “general feeling” in EA. Instead, get data if you can, and ask relevant people their actual thoughts - you might find them surprisingly out of step with what the vibe would indicate.
    I’m pretty unsure what we can or should do as a community about this, but I have a few thoughts at the bottom, and having a post about it as something to point to might help.
    Longer version:
    I think you'll get plenty out of this if you read the headings and read more under each heading if something piques your curiosity
    Part 1: What fuzzy proxies are people using and why would they be systematically overweighted?
    (I don’t know how common these mistakes are, or that they apply to you, the specific reader of the post. I expect them to bite harder if you’re newer or less connected, but I also expect that it’s easy to be somewhat biased in the same directions even if you have a lot of context. I’m hoping this serves as contextualization for the former and a reminder / nudge for the latter.)
    Getting funding from OP and LTFF
    Seems easy to expect that if someone got funding from Open Phil or the Long Term Future Fund, that’s a reasonable signal about the value of their work or the competence or trustworthiness or other virtues of the person running it. It obviously is Bayesian evidence, but I expect this to be extremely noisy.
    These organisations engage in hits-based philanthropy - as I understand it, they don’t expect most of the grants they make to be especially valuable (but the amount and way this is true varies by funder - Linch describes...
    16 min
  • LW - Automatic Rate Limiting on LessWrong by Raemon
    Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Automatic Rate Limiting on LessWrong, published by Raemon on June 23, 2023 on LessWrong.
    The LessWrong team recently began rolling out automatic rate limits. The general idea is if a user gets heavily downvoted, the site automatically restricts their posting and commenting privileges. (See Well-Kept Gardens Die By Pacifism for some background on LessWrong's overall moderation philosophy)
    After thinking awhile about various side-effects (see previous discussion), I designed some rate limits based on the following user metadata:
    Total karma. i.e. all net upvotes/downvotes the user has ever received
    Recent karma: The net karma the user received from their last 20 comments and/or posts.
    Last month's karma: The karma the user received for their last 20 comments/posts within the last month (i.e. a strict subset of the "recent karma.")
    Unique downvoter count: The number of individuals who've downvoted the user's recent content. This ensures that a single angry individual or small clique can't wreck your day.
    The exact karma thresholds and rate-limit-strengths chosen are based on the LW team's experience with daily moderation maintenance. Every day, moderators look over ~20 posts and comments from new-ish users. We've built up some intuitions for what user-karma and comment/post-karma tends to correspond with stuff we want to see more of on LessWrong. And for the past month or so I've been checking how these four metrics apply to the current distribution of LW users.
    The karma system isn't perfect. Sometimes you'll get randomly downvoted by someone for idiosyncratic or petty reasons. Sometimes it's hard for newer user's content to get noticed and upvoted. But people tend to be pretty hesitant to downvote stuff – most comments just don't get voted on. Downvoted content from new users tends to be some combination of poorly argued, difficult to read, rude-without-actually-saying-much, or rehashing topics that have been discussed extensively on LessWrong without introducing new considerations.
    I suspect we'll iterate and fine-tune these rules over time.
    Later in this post I have advice on writing posts/comments that tend to be well received on LessWrong. If you've been downvoted and/or rate-limited, don't take it too hard. LessWrong has fairly particular standards. My recommendation is to read some of the advice at the end here and try again.
    Current AutoRateLimits (as of June 22)
    We'll probably experiment with these a bit. The exact implementation here involves a lot of different numbers, but here's a rough overview of the current rate limit philosophy:
    Default rate limits for users with less than 5 karma (aimed at new users).
    Moderately strong (but non-escalating) rate limits for users with negative karma.
    Escalating rate limits for users with negative recent karma.
    Stronger rate limits require more unique downvoters, so that a single angry voter or small clique can't have too large an impact.
    The hope is for users to mostly show up, get recognized if they write good content, and then get full posting permissions. If a user is getting net-downvoted, then when they go to post they'd they get a message looking something like:
    Default Rate Limits
    Users whose total karma is < 5 are limited to 3 comments a day, and 2 posts per week.
    The hope here is that these numbers are: a) high enough that most users can get started commenting/posting without having to deal with a rate limit, and b) low enough that a user writing low-quality content can't go on too big a commenting spree, before having to slow down a bit and learn some site norms.
    The default commenting rate limits apply to writing comments on your own posts.
    Negative Karma Rate Limits
    If a user has -1 or less total karma, they can only write one comment per day and one post per two-weeks.
    This doesn't require multiple downvoter...
    12 min
  • EA - Growth and engagement in EA groups: 2022 Groups Census results by BrianTan
    Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Growth and engagement in EA groups: 2022 Groups Census results, published by BrianTan on June 23, 2023 on The Effective Altruism Forum.
    Summary
    In December 2022, CEA sent out a survey to all groups. 309 EA groups reported that they were active. However, based on our records, we estimate that ~63 active EA groups didn’t fill out the census. Therefore, it’s important to note that the data in this post are not representative of all active groups.
    Almost half of all active groups (309) are university groups (150). After that, most groups are location-based (53 city groups, 25 city-university hybrid groups, and 40 national groups). There were 40 groups of other types (e.g. workplace).
    The number of groups has been growing quickly in recent years. The net growth in the number of groups was 19% in 2020–2021 and 31% in 2021–2022. Most of the groups that were founded in recent years are university groups: the number of university groups roughly doubled between 2020 and 2022.
    We hypothesize that the two main reasons for the fast growth rate in general are CEA’s University Group Accelerator Program (UGAP) and the increased focus on and funding for EA community building in 2021–2022.
    We’d like to note that we don’t aim to just maximize the number of groups or members — we care about the quality and fidelity of groups, and we’re taking steps to measure and help improve group quality.
    There are active groups in 56 countries. Most of the currently active groups are in Europe and North America (39% and 30% respectively). However, the number of groups in Asia, South America, and Africa significantly grew in 2022, ranging from 65-86% growth per continent that year.
    We think the high growth in these continents is partially due to UGAP, EA Virtual Programs, and other community building efforts (e.g. by the Spanish-speaking community).
    We also asked groups for attendance and engagement data.
    The data we have shows that the median group size remained constant between 2020 and 2022, although due to changes in the questions we asked between the 2020 and 2022 surveys, we can only meaningfully compare the attendance and engagement data of groups on two measures (total event attendees and engaged EAs).
    While the total number of attendees per group stayed the same, the number of members who engaged with EA deeply — defined as 100 hours of engagement with EA content and taking some significant action — increased significantly. That is, the mean number of “engaged EAs” per group doubled (from 6 to 12) and the median even tripled (from 2 to 6). We think it’s plausible that this suggests that it has become easier to deeply engage with EA for people around the world (e.g. through accessible fellowships, more conferences, and higher Forum activity).
    As expected, engagement data across all measures we asked for were skewed: a small share of the groups produced the majority of outcomes in terms of number of events organized, EAG(x) attendees, engaged EAs, and total attendees. The least skewed measures were the number of regular attendees and intro fellowship participants per group. We discuss potential explanations for this in this post.
    On average, organizers rated their satisfaction with CEA’s support as 7.6/10. This is roughly the same as in 2020 — though it’s worth noting that the number of groups that CEA supports, and thus the number that rated their satisfaction with CEA’s support, has increased by ~50%.
    Compared to two years ago, there was an increase both in groups that are dissatisfied and groups that are very satisfied with CEA support. We think that one of the reasons may be that CEA’s support has become more targeted at specific groups (e.g. city/national groups in key locations and new uni groups in UGAP).
    Profession-based and virtual groups were on average less satisfied with support fro...
    36 min
  • LW - When is correlation transitive? by Ege Erdil
    Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: When is correlation transitive?, published by Ege Erdil on June 23, 2023 on LessWrong.
    It's a well-known property of correlation that it's not transitive in general. If X,Y,Z are three real-valued random variables such that ρ(X,Y)>0 and ρ(Y,Z)>0, it doesn't have to be the case that ρ(X,Z)>0.
    Nevertheless, there are some circumstances under which correlation is transitive. I will focus on two such cases in this post.
    Primer: correlation as an inner product
    For what follows, some background knowledge is necessary that we can regard correlations of real-valued random variables with finite second moments as inner products in an appropriate Hilbert space.
    Specifically, if X,Y are two such random variables with zero mean and unit standard deviation, which is a simplification we can always make as correlation is invariant under translation and scalar multiplication, then we can compute
    ρ(X,Y)=cov(X,Y)σXσY=E[XY]−E[X]E[Y]=E[XY]
    The pairing (A,B)E[AB] defines an inner product on the space of random variables with finite second moments where two random variables are considered equivalent if they are equal with probability 1 (almost surely). The properties that we expect out of an inner product are easy to check: the pairing is obviously bilinear and positive definite.
    Furthermore, it turns out this inner product turns the space of random variables with finite second moments into a Hilbert space: the vector space turns out to be complete under the induced norm ∥X∥=√E[X2]. Roughly speaking, this means that we can take orthogonal projections onto closed subspaces with impunity.
    Now that we have this framework, we can move on to the main results of this post.
    Correlation is transitive when the correlations are sufficiently strong
    I'll first prove the following:
    Claim 1: If ρ(X,Y)=a and ρ(Y,Z)=b, then
    ab−√(1−a2)(1−b2)≤ρ(X,Z)≤ab+√(1−a2)(1−b2)
    Moreover, these bounds are tight: for any a,b, there is a combination X,Y,Z for which we can make either the right or the left inequality into an equality.
    Proof
    We can assume X,Y,Z have mean zero and unit variance without loss of generality. Taking orthogonal projections of X,Z onto the one-dimensional subspace spanned by Y, we can write
    X=aY+√1−a2HXYZ=bY+√1−b2HZY
    where E[YHXY]=E[YHZY]=0 and the random variables HXY,HZY have mean zero and variance 1. Taking inner products gives
    E[XZ]=ab+√(1−a2)(1−b2)E[HXYHZY]
    Using the Cauchy-Schwarz inequality for our inner product finishes the proof: |E[HXYHZY]|≤∥HXY∥∥HZY∥=1. For the existence proof, let Y be an arbitrary random variable with mean zero and unit variance and pick HXY,HYZ to be perfectly correlated or perfectly anti-correlated standard Gaussians that are uncorrelated with Y.
    Interpretation
    When a,b are large and positive, the lower bound ab−√(1−a2)(1−b2) is also positive, and so we have a guaranteed positive correlation between X and Z.
    One way to simplify this is to make it single-dimensional by assuming a=b. In this case, the lower bound is 2a2−1. If we want a guaranteed positive correlation between X and Z, this means the correlations ρ(X,Y)=ρ(Y,Z)=a have to satisfy a>1/√2≈0.7.
    This condition is quite strict, and we might wonder if some transitivity of correlation can be recovered in the absence of such strong correlations between X,Y and Y,Z. It turns out the answer is yes, at least if we assume the random variables are in some sense "generic".
    Correlation is transitive on the average
    It turns out that in a suitable sense, when X,Y and Y,Z are positively correlated, there is a tendency for X,Z to also be positively correlated, even though per Claim 1 we can't deduce that they must be positively correlated.
    The precise version of this claim is as follows:
    Claim 2: Let X,Y,Z be vectors independently and uniformly distributed on the n-dimensional unit sphere Sn⊂Rn+1, and let −1≤a,b≤1 be two re...
    10 min
  • AF - Catastrophic Risks from AI #3: AI Race by Dan H
    Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Catastrophic Risks from AI #3: AI Race, published by Dan H on June 23, 2023 on The AI Alignment Forum.
    This is the third post in a sequence of posts giving an overview of catastrophic AI risks.
    3 AI Race
    The immense potential of AIs has created competitive pressures among global players contending for power and influence. This “AI race” is driven by nations and corporations who feel they must rapidly build and deploy AIs to secure their positions and survive. By failing to properly prioritize global risks, this dynamic makes it more likely that AI development will produce dangerous outcomes. Analogous to the nuclear arms race during the Cold War, participation in an AI race may serve individual short-term interests, but it ultimately results in worse collective outcomes for humanity. Importantly, these risks stem not only from the intrinsic nature of AI technology, but from the competitive pressures that encourage insidious choices in AI development.
    In this section, we first explore the military AI arms race and the corporate AI race, where nation-states and corporations are forced to rapidly develop and adopt AI systems to remain competitive. Moving beyond these specific races, we reconceptualize competitive pressures as part of a broader evolutionary process in which AIs could become increasingly pervasive, powerful, and entrenched in society. Finally, we highlight potential strategies and policy suggestions to mitigate the risks created by an AI race and ensure the safe development of AIs.
    3.1 Military AI Arms Race
    The development of AIs for military applications is swiftly paving the way for a new era in military technology, with potential consequences rivaling those of gunpowder and nuclear arms in what has been described as the “third revolution in warfare." The weaponization of AI presents numerous challenges, such as the potential for more destructive wars, the possibility of accidental usage or loss of control, and the prospect of malicious actors co-opting these technologies for their own purposes. As AIs gain influence over traditional military weaponry and increasingly take on command and control functions, humanity faces a paradigm shift in warfare. In this context, we will discuss the latent risks and implications of this AI arms race on global security, the potential for intensified conflicts, and the dire outcomes that could come as a result, including the possibility of conflicts escalating to a scale that poses an existential threat.
    3.1.1 Lethal Autonomous Weapons (LAWs)
    LAWs are weapons that can identify, target, and kill without human intervention [32]. They offer potential improvements in decision-making speed and precision. Warfare, however, is a high-stakes, safety-critical domain for AIs with significant moral and practical concerns. Though their existence is not necessarily a catastrophe in itself, LAWs may serve as an on-ramp to catastrophes stemming from malicious use, accidents, loss of control, or an increased likelihood of war.
    LAWs may become vastly superior to humans. Driven by rapid developments in AIs, weapons systems that can identify, target, and decide to kill human beings on their own—without an officer directing an attack or a soldier pulling the trigger—are starting to transform the future of conflict. In 2020, an advanced AI agent outperformed experienced F-16 pilots in a series of virtual dogfights, including decisively defeating a human pilot 5-0, showcasing “aggressive and precise maneuvers the human pilot couldn't outmatch" [33]. Just as in the past, superior weapons would allow for more destruction in a shorter period of time, increasing the severity of war.
    Militaries are taking steps toward delegating life-or-death decisions to AIs. Fully autonomous drones were likely first used on the battlefield in Libya in March 20...
    53 min