The Nonlinear Library

The Nonlinear Library

By The Nonlinear FundEducation
Download on the App Store

All The Nonlinear Library episodes

  • LW - Input Swap Graphs: Discovering the role of neural network components at scale by Alexandre Variengien
    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: Input Swap Graphs: Discovering the role of neural network components at scale, published by Alexandre Variengien on May 12, 2023 on LessWrong.
    This post was written as part of the work done at Conjecture.
    You can try input swap graphs in a collab notebook or explore the library to replicate the results.
    Thanks to Beren Millidge and Eric Winsor for useful discussions throughout this project. Thanks to Beren for feedback on a draft of this post.
    Activation and path patching are techniques employed to manipulate neural network internals. One approach, used in ROME, which we'll refer to as corrupted patching, involves setting the output of certain parts of the network to a corrupted value. Another method used in the work on indirect object identification, referred to as input swap patching, involves assigning these parts the values they would produce on a different input sample, while the rest of the model operates normally. These experimental techniques have proven effective in gaining structural insights into the information flow within neural networks, providing a better understanding of which components are involved in a given behavior. These structural analyses can be straightforwardly automated using algorithms like Automatic Circuit DisCovery.
    Patching experiments also enable the discovery of the semantics of components, i.e. the abstract variables a component encodes. They have been used to discover the "a vs an" neuron in gpt2, the token and position signal in S-inhibition heads in the IOI circuit, or Othello board representations. In all these cases, researchers were not only able to say 'this component matters', but also 'it roughly plays this role'. However, these additional results come at a cost: they often required a lot of manual effort to design alternative datasets and come up with an intuition for the likely role of the component in advance.
    In this post, I introduce input swap graphs (swap graphs in short), a new tool to systematically analyze the results of input swap patching experiments and recover semantic information with as little human effort as possible.
    In the following, I present
    A concise overview of swap graphs, their motivation, and practical application to Name Movers within the IOI circuit.
    A simple theoretical framework for examining the problem of localized computation in computational graphs. This framework serves as a sandbox for rapidly building intuition to create or interpret experiments, such as swap graphs, which can be applied to real-world neural networks.
    An expansion of the IOI study in GPT-2 small to models containing several billion parameters (GPT-2 XL, Pythia-2.8B, GPT-Neo-2.7B) using swap graphs. We validated our findings through causal scrubbing and targeted interventions that steer the mechanism out-of-distribution predictably. This both demonstrates the potential for applying swap graphs to larger models and presents valuable data to investigate how a model's internal structure changes with scale while performing the same task.
    A five-minute overview of swap graphs
    Motivation
    Deciphering the intermediate activations of neural network components can be a daunting task. The prevalent approach involves breaking down these high-dimensional vectors into subspaces where variations correspond to concepts that are both relevant to the model's computation and comprehensible to humans.
    In this work, we adopt an alternative strategy: we limit ourselves to using only input swap patching in order to uncover the roles of neural network components without directly exploiting the structure of their activations.
    While it is common to combine information about directions in activation space and input swap patching experiments in interpretability research (e.g., as demonstrated in Neel Nanda's work on Othello), our objective here is t...
    53 min
  • AF - Infinite-width MLPs as an "ensemble prior" by Vivek Hebbar
    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: Infinite-width MLPs as an "ensemble prior", published by Vivek Hebbar on May 12, 2023 on The AI Alignment Forum.
    Summary: A simple toy model suggests that infinitely wide MLPs generalize in an "ensemble-ish" way which is exponentially less data-efficient than Solomonoff induction. It's probably fixable by different initializations and/or regularizations, so I note it here mostly as a mathematical curiosity / interesting prior.
    The analysis seems to be qualitatively consistent with empirical results on generalization vs width in small MLPs.
    Note: The deficiency of the standard initialization at infinite width seems to be well-known and empirically supported in NTK-related literature, along with ways of fixing it. I'm unsure whether my particular analysis exists elsewhere.
    Core claims
    The standard initialization uses weights which are proportional to 1/√input_dimension. This has the effect of keeping the activations at roughly the same scale across layers. However, in the infinite width case, it ends up making the gradients in early layers infinitely smaller than those in the last layer. Hence, training an infinite-width MLP is equivalent to running a regression using the features represented by the last-layer neurons at initialization. These features never change during training, since the early gradients are all zero.
    If we train without regularization, we will tend to get something very "ensemble-ish", "smooth", and "dumb". I will first summarize this claim in a table, then spend the rest of the post going through the reasoning behind it.
    Solomonoff InductionInfinite width MLP, low L2-norm solutionBayesian update over programsLinear regression over circuitsPuts most of its weight on a small number of programs, each of which perfectly fits the data on its ownSpreads weight over a broad ensemble, including circuits which have only a small correlation with truthThe amount of data required to make the correct program dominate is O(K), where K is the program lengthThe amount of data to make the correct circuit dominate is O(2C), where C is some "complexity measure" (defined later). This is exponentially less data-efficient than Solomonoff induction.Calling it "superintelligent" is an understatementGeneralizes poorly on many tasksHighly amenable to "sharp" solutionsFavors smooth solutions, only creates "sharp" solutions if certain conditions are met by the training data.
    If we train an infinitely wide MLP from the standard initialization, only the last layer's weights change. So it is equivalent to a linear regression over an infinite set of random "features", these features being the activation patterns of the last layer neurons at initialization.
    If the MLP is deep enough, some of these last-layer neurons are contain the output of very intelligent circuits. However, if we train our infinite width MLP, these intelligent circuits will hardly be used by the regression, even if they are very useful. That is, the sum of the weights drawing from them in the last layer will be very small. The reason I believe this is the toy model in the next section.
    Toy model
    Let's call each last-layer neuron a "feature". As discussed earlier, their behavior never changes due to how the gradients pan out at infinite width. In a "real" infinite network, these features will be "useful" and "intelligent" to various degrees, but we will simplify this greatly in the toy model, by using just two types of features.
    The toy model asks: "Suppose that some features already compute the correct answer for every training datapoint, and that the rest of the features are random garbage. Will the trained network rely more on the perfect features, or will it use some giant mixture of random features?"
    Suppose we have d items in the training set, denoted x1,..,xn. Each has a label of either −1 or 1. Let's say...
    10 min
  • EA - Our Progress in 2022 and Plans for 2023 by Open Philanthropy
    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: Our Progress in 2022 and Plans for 2023, published by Open Philanthropy on May 12, 2023 on The Effective Altruism Forum.
    2022 was a big year for Open Philanthropy:
    We recommended over $650 million in grants — more, by far, than in any other year of our history. [More]
    We hired our first program officers for three new focus areas in our Global Health and Wellbeing portfolio. [More]
    Within our Longtermism portfolio, we significantly expanded our grantmaking and used a series of open calls to identify hundreds of promising grants to individuals and small projects. [More]
    We ran the Regranting Challenge, a novel experiment which allocated $150 million to outstanding programs at other grantmaking organizations. [More]
    We nearly doubled the size of our team. [More]
    This post compares our progress with the goals we set forth a year ago, and lays out our plans for the coming year, including:
    A significant update on how we handle allocating our grantmaking across causes. [More]
    A potential leadership transition. [More]
    Continued growth in grantmaking and staff. [More]
    Continued grantmaking
    Last year, we wrote:
    We aim to roughly double the amount [of funding] we recommend [in 2022] relative to [2021], and triple it by 2025.
    In 2022, we recommended over $650 million in grants (up from roughly $400 million in 2021).
    We changed our plans midway through the year, due to a stock market decline[ref]This just reflects a decline in the market; our main donors are still planning to give away virtually all of their wealth within their lifetimes.[/ref] that reduced our available assets and led us to adjust the cost-effectiveness bar we use for our spending on global health and wellbeing. When we wrote last year’s post, we had tentatively planned to allocate $500 million to GiveWell’s recommended charities; the actual allocation wound up being $350 million (up from $300 million in 2021).
    Currently, we expect to recommend over $700 million in grants in 2023, and no longer have a definite grantmaking goal for 2024 and 2025.
    Highlights from this year’s grantmaking
    This section outlines some of the major grants we made across our program areas.
    In grants to charities recommended by GiveWell:
    $10.4 million to the Clinton Health Access Initiative to support their Incubator program, which looks for cost-effective and scalable health interventions.
    $13.7 million to New Incentives for conditional cash transfers to boost vaccination rates in Nigeria.
    $4.4 million to Evidence Action to support their in-line chlorination program in Malawi.
    We also made a $48.8 million grant to the same program with funds from our 2021 allocation.
    Many other grants we haven’t listed here (see our full list of GiveWell-recommended grants).
    In potential risks from advanced AI:
    Redwood Research to support their research on aligning AI systems.
    Center for a New American Security to support their work on AI policy and governance.
    A number of projects related to understanding and aligning deep learning systems.
    In biosecurity and pandemic preparedness:
    Columbia University to support research on far-UVC light to reduce airborne disease transmission.
    Bipartisan Commission on Biodefense to support work on biodefense policy in the US.
    The Johns Hopkins Center for Health Security to support their degree program for students pursuing careers in biosecurity.
    In effective altruism community growth (with a focus on longtermism):
    80,000 Hours (marketing and general support) for its work to help people have more impact with their careers.
    Support for the translation of effective altruism-related content into non-English languages.
    Bluedot Impact to run courses related to several of our priority cause areas.
    Asterisk to publish a quarterly journal focused on topics related to effective altruism, among others.
    A program open to applicat...
    15 min
  • AF - Input Swap Graphs: Discovering the role of neural network components at scale by Alexandre Variengien
    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: Input Swap Graphs: Discovering the role of neural network components at scale, published by Alexandre Variengien on May 12, 2023 on The AI Alignment Forum.
    This post was written as part of the work done at Conjecture.
    You can try input swap graphs in a collab notebook or explore the library to replicate the results.
    Thanks to Beren Millidge and Eric Winsor for useful discussions throughout this project. Thanks to Beren for feedback on a draft of this post.
    Activation and path patching are techniques employed to manipulate neural network internals. One approach, used in ROME, which we'll refer to as corrupted patching, involves setting the output of certain parts of the network to a corrupted value. Another method used in the work on indirect object identification, referred to as input swap patching, involves assigning these parts the values they would produce on a different input sample, while the rest of the model operates normally. These experimental techniques have proven effective in gaining structural insights into the information flow within neural networks, providing a better understanding of which components are involved in a given behavior. These structural analyses can be straightforwardly automated using algorithms like Automatic Circuit DisCovery.
    Patching experiments also enable the discovery of the semantics of components, i.e. the abstract variables a component encodes. They have been used to discover the "a vs an" neuron in gpt2, the token and position signal in S-inhibition heads in the IOI circuit, or Othello board representations. In all these cases, researchers were not only able to say 'this component matters', but also 'it roughly plays this role'. However, these additional results come at a cost: they often required a lot of manual effort to design alternative datasets and come up with an intuition for the likely role of the component in advance.
    In this post, I introduce input swap graphs (swap graphs in short), a new tool to systematically analyze the results of input swap patching experiments and recover semantic information with as little human effort as possible.
    In the following, I present
    A concise overview of swap graphs, their motivation, and practical application to Name Movers within the IOI circuit.
    A simple theoretical framework for examining the problem of localized computation in computational graphs. This framework serves as a sandbox for rapidly building intuition to create or interpret experiments, such as swap graphs, which can be applied to real-world neural networks.
    An expansion of the IOI study in GPT-2 small to models containing several billion parameters (GPT-2 XL, Pythia-2.8B, GPT-Neo-2.7B) using swap graphs. We validated our findings through causal scrubbing and targeted interventions that steer the mechanism out-of-distribution predictably. This both demonstrates the potential for applying swap graphs to larger models and presents valuable data to investigate how a model's internal structure changes with scale while performing the same task.
    A five-minute overview of swap graphs
    Motivation
    Deciphering the intermediate activations of neural network components can be a daunting task. The prevalent approach involves breaking down these high-dimensional vectors into subspaces where variations correspond to concepts that are both relevant to the model's computation and comprehensible to humans.
    In this work, we adopt an alternative strategy: we limit ourselves to using only input swap patching in order to uncover the roles of neural network components without directly exploiting the structure of their activations.
    While it is common to combine information about directions in activation space and input swap patching experiments in interpretability research (e.g., as demonstrated in Neel Nanda's work on Othello), our object...
    53 min
  • EA - Simple charitable donation app idea by kokotajlod
    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: Simple charitable donation app idea, published by kokotajlod on May 12, 2023 on The Effective Altruism Forum.
    I'll pay $10xN to the people who build this app, where N is the total karma of this post three months from now, up to a max of $20,000, unless something shady happens like some sort of bot farm. If it turns out this app already exists, I'll pay $1xN instead to the people who find it for me. I'm open to paying significantly more in both cases if I'm convinced of the altruistic case for this app existing; this is just the minimum I personally can commit to and afford.The app consists of a gigantic, full-screen button such that if you press it, the phone will vibrate and play a little satisfying "ching" sound and light up sparkles around where your finger hit, and then $1 will be donated to GiveDirectly.
    You can keep slamming that button as much as you like to thereby donate as many dollars as you like.In the corner there's a menu button that lets you change from GiveDirectly to Humane League or AMF or whatever (you can go into the settings and input the details for a charity of your choice, adding it to your personal menu of charity options, and then toggle between options as you see fit. You can also set up a "Donate $X per button press instead of $1" option and a "Split each donation between the following N charities" option.)That's it really.Why is this a good idea? Well, I'm not completely confident it is, and part of why I'm posting is to get feedback. But here's my thinking:I often feel guilty for eating out at restaurants. Especially when meat is involved.
    Currently I donate a substantial amount to charity on a yearly basis (aiming for 10% of income, though I'm not doing a great job of tracking that) but it feels like a chore, I have to remember to do it and then log on and wire the funds. Like paying a bill.If I had this app, I think I'd experiment with the following policy instead: Every time I buy something not-necessary such as a meal at a restaurant, I whip out my phone, pull up the app, and slam that button N times where N is the number of dollars my purchase cost. Thus my personal spending would be matched with my donations. I think I'd feel pretty good while doing so, it would give me a rush of warm fuzzies instead of feeling like a chore. (For this reason I suggest having to press the button N times, instead of building the app to use a text-box-and-number-pad.)Then I'd check in every year or so to see whether my donations were meeting the 10% goal and make a bulk donation to make up the difference if not.
    If it exceeds the goal, great!I think even if no one saw me use this app, I'd still use it & pay for it. But there's a bonus effect having to do with the social consequences of being seen using it. Kinda like how a big part of why veganism is effective is that you can't hide it from anyone, you are forced to bring it up constantly. Using this app would hopefully have a similar effect -- if you were following a policy similar to the one I described, people would notice you tapping your phone at restaurants and ask you what you were doing & you'd explain and maybe they'd be inspired and do something similar themselves. (Come to think of it, it's important that the "ching" sound not be loud and obnoxious, otherwise it might come across as ostentatious.)
    I can imagine a world where this app becomes really popular, at least among certain demographics, similar to (though probably not as successful as) veganism.Another mild bonus is that this app could double as a tracker for your discretionary spending. You can go into the settings and see e.g. a graph of your donations over time, statistics on what time of day you do them, etc. and learn things like "jesus do I really spend that much on dining out per month?" and "huh, I guess those Amazon purchase...
    4 min
  • LW - Advice for newly busy people by Severin T. Seehrich
    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: Advice for newly busy people, published by Severin T. Seehrich on May 11, 2023 on LessWrong.
    After writing "Advice for interacting with busy people", I was asked to write a follow-up on advice for newly busy people. So, here's a quick list of tools and mental models that help me prioritize.
    This list is by no means comprehensive. It's just the tools I know and have loved. Take what's useful, and drop what doesn't fit your brain and life.
    1. Prioritizing between projects
    a. Apply the Tomorrow Rule.
    When someone asks you to join an exciting project that's due half a year from now, it is very, very tempting to say "yes". You'll immediately have a vivid imagination of the shiny outcome, while the workload is far enough in the future to not cross your mind. To mitigate this tendency, it makes sense to apply the Tomorrow Rule. It goes as such: "Am I committed enough to this that I'd clear up time in my schedule tomorrow to make it happen?"
    b. If things get too much, do a Productivity Purge.
    If you already have too many projects on your plate and can't make reasonable progress on any of them, you might want to go through a round of Cal Newport's productivity purge algorithm. The steps:
    "When it feels like your schedule is becoming too overwhelmed, take out a sheet of paper and label it with three columns: professional, extracurricular, and personal. Under “professional” list all the major projects you are currently working on in your professional life (if you’re a student, then this means classes and research, if you have a job, then this means your job, etc). Under “extracurricular” do the same for your side projects (your band, your blog, your plan to write a book). And under “personal” do the same for personal self-improvement projects (from fitness to reading more books).
    Under each list try to select one or two projects which, at this point in your life, are the most important and seem like they would yield the greatest returns. Put a star by these projects.
    Next, identify the projects that you could stop working on right away with no serious consequences. Cross these out.
    Finally, for the projects that are left unmarked, come up with a 1-3 week plan for finalizing and dispatching them. Many of these will be projects for which you owe someone something before you can stop working on them. Come up with a crunch plan for the near future for shutting these down as quickly as possible.
    Once you completed your crunch plan you’ll be left with only a small number of important projects. In essence, you have purged your schedule of all but a few contenders to be your next Theory of Relativity. Here’s the important part: Try to go at least one month without starting any new projects. Resist, at all costs, committing to anything during this month. Instead, just focus, with an Einsteinian intensity, on your select list."
    2. Prioritizing between people
    At some point in January/February, I felt pretty lonely and decided to make a list of all the lovely people I know and spend too little time with. After writing down the names of 40 people in Berlin alone and more in other cities, I realized what was the problem: I fully optimized for creating loose ties, for getting the spark of novelty and knowing what's going on in my various communities. In the meantime, I didn't commit enough to anyone as that I'd know who to call when I'm feeling low.
    So - you might want to create two networks simultaneously, which work by different rules:
    A large network of loose ties. These people are there for uncommitted play, for exchanging knowledge and occasional favors, for having as many people as possible in your life who know somebody who knows somebody who happens to be really skilled at that particular thing. Treat them with kindness and integrity, but don't hesitate to say "no" if your heart isn't...
    8 min
  • LW - I bet $500 on AI winning the IMO gold medal by 2026 by azsantosk
    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: I bet $500 on AI winning the IMO gold medal by 2026, published by azsantosk on May 11, 2023 on LessWrong.
    The bet was arranged on Twitter between @MichaelVassar and I (link).
    Conditions are similar to this question on Metaculus, except for the open-source condition (I win even if the AI is closed-source, and in fact I would very much prefer it to be closed-source).
    @Zvi has agreed to adjudicate this bet in case there is no agreement on resolution.
    Michael has asked me two questions by email, and I'm sharing my answers.
    Any thoughts on how to turn winning these sorts of bets into people actually updating?
    Geoffrey Hinton mentioned recently that, while GPT4 can "already do simple reasoning", "reasoning is the area where we're still better" [source].
    It seems to me that, after being able to beat humans at math, there won't be anything else fundamental where we're still better. I wish more people could realize this now.
    For the people who disagree, I would like to get them to make their views known before that. I feel like many people just don't think enough about it and are in a "superposition state", where their belief can collapse to anything without it causing any surprise feelings or model updating on them. Maybe if they think about it and commit to their views today they'll be more surprised when it happens and therefore willing to change their mind about important matters.
    Kelvin, thoughts on how you’ll update if it doesn’t at least come close?
    Yes. I'll update in the following directions:
    That it is much harder for search algorithms to do "amplification" right in rich (very large branching factor) environments like math (and the real-world), than in games such as Go.
    That superintelligence is still further away, even if LLMs are having lots of economic use cases and replacing many human jobs.
    That we're probably bottlenecked on search algorithms, rather than compute power/model-size. This would have policy implications.
    Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org.
    2 min
  • EA - In defence of epistemic modesty [distillation] by Luise
    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: In defence of epistemic modesty [distillation], published by Luise on May 10, 2023 on The Effective Altruism Forum.
    This is a distillation of In defence of epistemic modesty, a 2017 essay by Gregory Lewis. I hope to make the essay’s key points accessible in a quick and easy way so more people engage with them. I thank Gregory Lewis for helpful comments on an earlier version of this post. Errors are my own.
    Note: I sometimes use the first person (“I claim”/”I think”) in this post. This felt most natural but is not meant to imply any of the ideas or arguments are mine. Unless I clearly state otherwise, they are Gregory Lewis’s.
    What I Cut
    I had to make some judgment calls on what is essential and what isn’t. Among other things, I decided most math and toy models weren’t essential. Moreover, I cut the details on the “self-defeating” objection, which felt quite philosophical and probably not relevant to most readers. Furthermore, it will be most useful to treat all the arguments brought up in this distillation as mere introductions, while detailed/conclusive arguments may be found in the original post and the literature.
    Claims
    I claim two things:
    You should practice strong epistemic modesty: On a given issue, adopt the view experts generally hold, instead of the view you personally like.
    EAs/rationalists in particular are too epistemically immodest.
    Let’s first dive deeper into claim 1.
    Claim 1: Strong Epistemic Modesty
    To distinguish the view you personally like from the view strong epistemic modesty favors, call the former “view by your own lights” and the latter “view all things considered”.
    In detail, strong epistemic modesty says you should do the following to form your view on an issue:
    Determine the ‘epistemic virtue’ of people who hold a view on the issue. By ‘epistemic virtue’ I mean someone’s ability to form accurate beliefs, including how much the person knows about the issue, their intelligence, how truth-seeking they are, etc.
    Determine what everyone's credences by their own lights are.
    Take an average of everyone’s credences by their own lights (including yourself), weighting them by their epistemic virtue.
    The product is your view all things considered. Importantly, this process weighs your credences by your own lights no more heavily than those of people with similar epistemic virtue. These people are your ‘epistemic peers’.
    In practice, you can round this process to “use the existing consensus of experts on the issue or, if there is none, be uncertain”.
    Why?
    Intuition Pump
    Say your mom is convinced she’s figured out the one weird trick to make money on the stock market. You are concerned about the validity of this one weird trick, because of two worries:
    Does she have a better chance at making money than all the other people with similar (low) amounts of knowledge on the stock market who’re all also convinced they know the one weird trick? (These are her epistemic peers.)
    How do her odds of making money stack up against people working full-time at a hedge fund with lots of relevant background and access to heavy analysis? (These are the experts.)
    The point is that we are all sometimes like the mom in this example. We’re overconfident, forgetting that we are no better than our epistemic peers, be the question investing, sports bets, musical taste, or politics. Everyone always thinks they are an exception and have figured [investing/sports/politics] out. It’s our epistemic peers that are wrong! But from their perspective, we look just as foolish and misguided as they look to us.
    Not only do we treat our epistemic peers incorrectly, but also our epistemic superiors. The mom in this example didn’t seek out the expert consensus on making money on the stock market (maybe something like “use algorithms” and “you don’t stand a chance”). Instead, she may have li...
    14 min
  • EA - How much funging is there with donations to different EA animal charities? by Brian Tomasik
    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 much funging is there with donations to different EA animal charities?, published by Brian Tomasik on May 11, 2023 on The Effective Altruism Forum.
    My main question
    The EA Funds Animal Welfare Fund makes grants to many different animal charities. Suppose I want to support one particular charity that they grant to because I think it's better, relative to my values, than most of the other ones. For example, maybe I want to specifically give to Legal Impact for Chickens (LIC), so I donate $1000 to them.
    Because this donation reduces LIC's room for more funding, it may decrease the amount that the Animal Welfare Fund itself (or Open Philanthropy, Animal Charity Evaluators, or individual EA donors) will give to LIC in the future. How large should I expect this effect to be in general? Will my $1000 donation tend to "funge" against these other EA donors almost fully, so that LIC can be expected to get about $1000 less from them? Is the funging amount more like $500? Is it roughly $0 of funging? Or maybe donating to LIC helps them grow faster, so that they can hire more people and do more things, thereby increasing their room for funding and how much other EA donors give to them?
    The answer to this question probably varies substantially from one case to the next, and maybe the best way to figure it out would be to learn a lot about the funding situation for a particular charity and the funding inclinations of big EA donors toward that charity. But that takes a lot of work, so I wonder if EA funders have some intuition for what tends to happen on average in situations like this, to inform small donors who aren't going to get that far into the weeds with a particular charity. Does the funging amount tend to be closer to 0% or closer to 100% of what an individual donor gives?
    I notice that the Animal Welfare Fund sometimes funds ~10% to ~50% of an organization's operating budget, which I imagine may be partly intentional to avoid crowding out small donors. (It may also be motivated by wanting charities to diversify their funding sources and due to limited funds to disburse.) Is it true in general that the Animal Welfare Fund doesn't fully fill room for funding, or are there charities for which the Fund does top up the charity completely? (Note that it would actually be better impact-wise to ensure that the very best charities are roughly fully funded, so I'm not encouraging a strategy of deliberately underfunding them.)
    In the rest of this post, I'll give more details on why I'm asking about this topic, but this further elaboration is optional reading and is more specific to my situation.
    My donation preferences
    I think a lot of EA donations to animal charities are really exciting. About 1/3 of the grants in the Animal Welfare Fund's Grants Database seem to me roughly as cost-effective as possible for reducing near-term animal suffering. However, for some other grants, I'm pretty ambivalent about the sign of the net impact (whether it's net good or bad).
    This is mainly for two reasons:
    I'm unsure if meat reduction, on the whole, reduces animal suffering, mainly because certain kinds of animal farming, especially cattle grazing on non-irrigated pasture, may reduce an enormous amount of wild-animal suffering (though there are huge error bars on this analysis).
    I'm unsure if antispeciesism in general reduces net suffering. In the short run, I worry that it may encourage more habitat preservation, thereby increasing wild-animal suffering. In the long run, moral-circle expansion could encourage people to create lots of additional small-brained sentience, and in (hopefully unlikely) scenarios where human values become inverted, antispeciesist values could multiply total suffering manyfold.
    If I could press a button to reduce overall meat consumption or to increase concern for an...
    11 min
  • EA - US Supreme Court Upholds Prop 12! by Rockwell
    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: US Supreme Court Upholds Prop 12!, published by Rockwell on May 11, 2023 on The Effective Altruism Forum.
    The United States Supreme Court just released its decision on the country's most pivotal farmed animal welfare case—NATIONAL PORK PRODUCERS COUNCIL ET AL. v. ROSS, SECRETARY OF THE CALIFORNIA DEPARTMENT OF FOOD AND AGRICULTURE, ET AL. —upholding California's Prop 12, the strongest piece of farmed animal legislation in the US.In 2018, California residents voted by ballot measure to ban the sale of pig products that come from producers that use gestation crates, individual crates the size of an adult pig's body that mother pigs are confined to 24/7 for the full gestation of their pregnancies, unable to turn around. In response, the pork industry sued and the case made its way to the nation's highest court.If the Supreme Court had not upheld Prop 12, years of advocacy efforts would have been nullified and advocates would no longer be able to pursue state-level legislative interventions that improve welfare by banning the sale of particularly cruelly produced animal products.
    It would have been a tremendous setback for the US animal welfare movement. Instead, today is a huge victory.
    Groups like HSUS spearheaded efforts to uphold Prop 12, even in the face of massive opposition. The case exemplified the extent to which even left-leaning politicians side with animal industry over animal welfare, as even the Biden administration sided with the pork industry.
    Today is a monumental moment for farmed animal advocacy. Congratulations to everyone who worked to make this happen!
    Read more about it:
    Summary and analysis from Lewis Bollard (Senior Program Officer for Farm Animal Welfare at Open Phil) here on Twitter.
    Victory announcement by the Humane Society of the United States here.
    New York Times coverage here.
    Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org.
    2 min