The Nonlinear Library

The Nonlinear Library

By The Nonlinear FundEducation
Download on the App Store

All The Nonlinear Library episodes

  • EA - Ask (Everyone) Anything — “EA 101” by Lizka
    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: Ask (Everyone) Anything — “EA 101”, published by Lizka on October 5, 2022 on The Effective Altruism Forum.
    I invite you to ask anything you’re wondering about that’s remotely related to effective altruism. There’s no such thing as a question too basic.
    Try to ask your first batch of questions by Monday, October 17 (so that people who want to answer questions can know to make some time around then).
    Everyone is encouraged to answer (see more on this below). There’s a small prize for questions and answers.
    This is a test thread — we might try variations on it later.
    How to ask questions
    Ask anything you’re wondering about that has anything to do with effective altruism.
    More guidelines:
    Try to post each question as a separate "Answer"-style comment on the post.
    There’s no such thing as a question too basic (or too niche!).
    Follow the Forum norms.
    I encourage everyone to view asking questions that you think might be “too basic” as a public service; if you’re wondering about something, others might, too.
    Example questions
    I’m confused about Bayesianism; does anyone have a good explainer?
    Is everyone in EA a utilitarian?
    Why would we care about neglectedness?
    Why do people work on farmed animal welfare specifically vs just working on animal welfare?
    Is EA an organization?
    How do people justify working on things that will happen in the future when there’s suffering happening today?
    Why do people think that forecasting or prediction markets work? (Or, do they?)
    How to answer questions
    Anyone can answer questions, and there can (and should) be multiple answers to many of the questions. I encourage you to point people to relevant resources — you don’t have to write everything from scratch!
    Norms and guides:
    Be generous and welcoming (no patronizing).
    Honestly share your uncertainty about your answer.
    Feel free to give partial answers or point people to relevant resources if you can’t or don’t have time to give a full answer.
    Don’t represent your answer as an official answer on behalf of effective altruism.
    Keep to the Forum norms.
    You should feel free and welcome to vote on the answers (upvote the ones you like!). You can also give answers to questions that already have an answer, or reply to existing answers, especially if you disagree.
    The (small) prize
    This isn’t a competition, but just to help kick-start this thing (and to celebrate excellent discussion at the end), the Forum team will award $100 each to my 5 favorite questions, and $100 each to my 5 favorite answers (questions posted before Monday, October 17, answers posted before October 24).
    I’ll post a comment on this post with the results, and edit the post itself to list the winners.
    Your feedback is very welcome! We’re considering trying out themed versions in the future; e.g. “Ask anything about cause prioritization” or “Ask anything about AI safety.”
    We’re hoping this thread will help get clarity and good answers, counter some impostor syndrome that exists in the community (see 1 and 2), potentially rediscover some good resources, and generally make us collectively more willing to ask about things that confuse us.
    If I think something is rude or otherwise norm-breaking, I’ll delete it.
    Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org.
    4 min
  • LW - Quick notes on “mirror neurons” by Steven Byrnes
    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: Quick notes on “mirror neurons”, published by Steven Byrnes on October 4, 2022 on LessWrong.
    1. Two definitions of “mirror neurons”
    I have noticed the term “mirror neurons” being used in two ways:
    (A) Any set of neurons that activates both when I am doing / feeling / relating to X and when I am watching (or imagining) someone else doing / feeling / relating to that same X.
    (B) Same as (A), but where X is more specifically an action concept (e.g. “grasp”), and where the neurons in question are more specifically in certain specific frontal-lobe brain areas including (parts of) premotor cortex, primary motor cortex, and Broca’s area.
    The (B) usage is the original one, and much more common in the technical neuroscience literature. By contrast, (A) seems quite common among non-neuroscientists, and shows up at least occasionally even in the technical neuroscientist literature.
    Anyway, regardless of what we call it, I strongly believe that (A) is a real neuroscience phenomenon. If I watch you juggle, versus if I juggle myself, some of the same juggling-related neural circuits are active in both cases. That seems perfectly obvious to me. After all, how else could I possibly generalize from one to the other? How could I even know that they have anything to do with each other??
    By contrast, my sense is that (B) is a giant dumpster fire of nonsense and badly-interpreted experiments. I recommend the excellent book The Myth of Mirror Neurons by Gregory Hickok for that whole story.
    2. Mirror neurons and kindness
    There’s a pop science trope that mirror neurons explain love and kindness and so on. I don’t think that (A) in itself explains human prosocial behavior. I do think that (A) is probably an ingredient in the explanation of human prosocial behavior. And I also think that (A) is probably an ingredient in the explanation of human antisocial behavior! For more on that, see Symbol grounding & human social instincts, Section 13.5 on empathetic simulation.
    3. Why I myself don’t use the term “mirror neurons”
    Why don’t I personally use the term “mirror neurons” for (A)? First, because it would be arguably an incorrect usage, or at least ambiguous—see above. Second, because I find “mirror neuron” a somewhat misleading term anyway.
    For example, as above, take some set of temporal-lobe neurons which collectively represent the concept “juggle”. These neurons are (I claim) active both when I’m juggling and when I’m watching you juggle. Should we call those neurons “mirror neurons”? I say no. It’s much more helpful and sensible to call them “neurons that represent the concept ‘juggle’”, right??
    Likewise, consider neurons in the anterior cingulate cortex which activate both when I’m disgusted and when I see a disgusted face. Should we call those “mirror neurons”? I don’t think so; in my mind, it’s much clearer to call them “disgust-reaction neurons, and oh by the way, fun fact, those neurons also tend to activate during empathetic simulation of someone else experiencing a disgust-reaction”.
    So my policy is to avoid the term “mirror neuron” altogether when trying to describe how things actually work in the brain.
    Examples of people using the word “mirror neuron” to refer to (A) are Meyza et al. 2017 and van der Gaag et al. 2007. I can also report personal communication with an eminent neuroscientist who was under the impression that the definition of “mirror neuron” was (A).
    Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org.
    4 min
  • EA - Things I did in my job search I recommend 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: Things I did in my job search I recommend, published by ChanaMessinger on October 4, 2022 on The Effective Altruism Forum.
    Made a big google doc of ALL the things I could do. Not just a or b but literally everything I could think of, shared it around a bunch.
    Asked other people what I could do
    Tried to think radically more ambitiously
    Listed skills I thought I had and then thought of jobs that would use those
    Applied to 80k advising
    Which caused me to write up my thoughts on lost of things and answer lots of questions
    And I had a nice chat, which ended up leading to the job I ended up taking
    Talked to, on the margin, probably too many people, but SO many people
    exploring new spaces (institutional decision making, distillation, etc)
    Asking what they thought was needed in those spaces
    Talked to my friends about my job search, which led to more opportunities
    Asked "If I was going to work in x [AI Safety, biosecurity, etc] what would I do?" and set 5 minute timers to generate ideas. (Thank you to Michelle Hutchinson for this)
    Tried to ask a variety of questions on how I might approach this search
    What skills do I have that are helpful for the world?
    What skills will I want to have in 5-10 years?
    What might cause me to burn out? How can I mitigate that?
    What are my current bottlenecks? What's stopping me?
    What's missing in the space? What's the low hanging fruit?
    How can I be ambitious?
    What job will give the most mentorship / the manager I work best with?
    Was super honest in all work trials and interviews about my level of interest and skill and what other jobs I was applying for
    Had conversations about why this job over that one was more impactful
    Used work trials to try to assess my own interest in the work (this didn't work for me, I got excited about everything)
    Wrote another google doc when I had honed it down to three possibilities, shared it with my friends and family, got comments, thought hard.
    Made the call.
    So far, enjoying my job!
    Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org.
    3 min
  • AF - How are you dealing with ontology identification? by Erik Jenner
    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 are you dealing with ontology identification?, published by Erik Jenner on October 4, 2022 on The AI Alignment Forum.
    Not really novel, but I've made this point quite often in recent conversations, so I decided to make a short write-up.
    I think a wide range of alignment proposals will somehow have to deal with ontology identification, i.e. mapping an AI's beliefs about the world to a representations humans can correctly understand. This could happen explicitly (e.g. via an ELK solution or strong interpretability), or be more hidden, or maybe the proposal avoids the problem in some clever way. But I think alignment proposals should have some answer to the question "How are you dealing with ontology identification?"
    A class of examples
    One high-level approach you could try to use to align AI goes as follows:
    Somehow get a model of human values. This could be a utility function, but might also look more like a preference ordering etc.
    Build an AI that optimizes according to this value model.
    (If the value model is just a preference ordering, you might have to deal with the fact that it can be inconsistent, but we're ignoring those issues here.)
    To be clear, this is not the only approach for building aligned AI (maybe not even the most promising one). It's a special case of separately solving outer alignment and inner alignment, which itself is a special case of solving alignment more broadly. But this still covers an important class of approaches.
    I think any approach along those lines has to deal with ontology identification, for a pretty simple reason. The model of human values we get from step 1. will be defined in terms of the human ontology. For example, if we're learning a utility function, it will have type signature Uhuman:{states in human ontology}R. But for step 2., we need something the AI can optimize over, so we need a utility function of type signature UAI:{states in AI ontology}R. This means that as a bridge between steps 1. and 2., we need a translation from the AI ontology to the human ontology, τ:{states in AI ontology}{states in human ontology}. Then we can use UAI:=Uhuman∘τ.
    Note that in practice, we will often collapse learning Uhuman and specifying τ into a single step. For example, when doing RLHF on videos of trajectories, we're learning a reward function defined directly in terms of videos, rather than a human representation of what's happening in those videos. These videos also map easily into the RL agent's ontology (the agent's initial layers perform that mapping). So we don't need to explicitly use an ontology translation τ anywhere.
    Implicitly, we're using the translation τ that satisfies UAI=Uhuman∘τ, where UAI is the reward model, defined as a function on videos. τ maps from videos to human's understanding of what's happening inside those videos, and Uhuman are the "true" human values (which aren't explicitly represented anywhere in this setup).
    This implicit translation τ corresponds to the human simulator from the ELK report: it maps videos to what humans think is happening when they see those videos. This leads to outcomes that look good to humans on video, rather than outcomes that are actually good. So when I say that alignment proposals have to "deal with ontology identification" somehow, I mean that they need a better map τ than this implicit one.
    So far, we've focused on RLHF as an example, in which case this perspective doesn't add much. But I think the same argument applies much more broadly. For example, one response to the issues with RLHF is "Ah, we need a better model of human irrationality". In other words, we should do a lot of cognitive science to figure out precisely in which ways humans give feedback that doesn't reflect their true preferences. Then we can back out true human preferences from irrational human feedba...
    6 min
  • EA - Overreacting to current events can be very costly by Kelsey Piper
    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: Overreacting to current events can be very costly, published by Kelsey Piper on October 4, 2022 on The Effective Altruism Forum.
    epistemic status: I am fairly confident that the overall point is underrated right now, but am writing quickly and think it's reasonably likely the comments will identify a factual error somewhere in the post.
    Risk seems unusually elevated right now of a serious nuclear incident, as a result of Russia badly losing the war in Ukraine. Various markets put the risk at about 5-10%, and various forecasters seem to estimate something similar. The general consensus is that Russia, if they used a nuclear weapon, would probably deploy a tactical nuclear weapon on the battlefield in Ukraine, probably in a way with a small number of direct casualties but profoundly destabilizing effects. A lot of effective altruists have made plans to leave major cities if Russia uses a nuclear weapon, at least until it becomes clear whether the situation is destabilizing.
    I think if that happens we'll be in a scary situation, but based on how we as a community collectively reacted to Covid, I predict an overreaction -- that is, I predict that if there's a nuclear use in Ukraine, EAs will incur more costs in avoiding the risk of dying in a nuclear war than the actual expected costs of dying in a nuclear war, more costs than necessary to reduce the risks of dying in a nuclear war, and more costs than we'll endorse in hindsight.
    With respect to Covid, I am pretty sure the EA community and related communities incurred more costs in avoiding the risk of dying of Covid than was warranted. In my own social circles, I don't know anyone who died of Covid, but I know of a healthy person in their 20s or 30s who died of failing to seek medical attention because they were scared of Covid. A lot of people incurred hits to their productivity and happiness that were quite large.
    This is especially true for people doing EA work they consider directly important: being 10% less impactful at an EA direct work job has a cost measured in many human or animal or future-digital-mind lives, and I think few people explicitly calculated how that cost measured up against the benefit of reduced risk of Covid.
    If Russia uses a nuclear weapon in Ukraine, here is what I expect to happen: a lot of people will be terrified (correctly assessing this as a significant change in the equilibrium around nuclear weapon use which makes a further nuclear exchange much more likely.) Many people will flee major cities in the US and Europe. They will spend a lot of money, take a large productivity hit from being somewhere with worse living conditions and worse internet, and spend a ton of their time obsessively monitoring the nuclear situation. A bunch of very talented ops people will work incredibly hard to get reliable fast internet in remote parts of Northern California or northern Britain. There won't be much EAs not already in nuclear policy and national security can do, but there'll be a lot of discussion and a lot of people trying to get up to speed on the situation/feeling a lot of need to know what's going on constantly. The stuff we do is important, and much less of it will get done.
    It will take a long time for it to become obvious if the situation is stable, but eventually people will mostly go back to cities (possibly leaving again if there are further destabilizing events).
    The recent Samotsvety forecast estimates that a person staying in London will lose 3-100 hours to nuclear risk in expectation (edit: which goes up by a factor of 6 in the case of actual tactical nuke use in Ukraine.) I think it is really easy for that person to waste more than 3-100 hours by being panicked, and possible to waste more than 20 - 600 hours on extreme response measures. And that's the life-hour costs of never fleein...
    6 min
  • AF - Smoke without fire is scary by Adam Jermyn
    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: Smoke without fire is scary, published by Adam Jermyn on October 4, 2022 on The AI Alignment Forum.
    This post was written in response to Evan Hubinger’s shortform prompt below and benefited from discussions with him. Thanks to Justis Mills for feedback on this post.
    If we see precursors to deception (e.g. non-myopia, self-awareness, etc.) but suspiciously don’t see deception itself, that’s evidence for deception.
    Setup
    Suppose an AI lab tries to look for the capabilities that enable deceptive alignment. This definitely seems like a good thing for them to do. Say they search for evidence of models optimizing past the current training epoch, for manipulation during human feedback sessions, for situational awareness, and so on. Suppose that when all is said and done they see the following:
    There are at least two ways to interpret this:
    “Phew, bigger models are less likely to be deceptively aligned than small ones. Scale away!”
    “Bigger models are more capable at hiding their deceptive alignment. Run away!”
    These are very different interpretations, with very different recommendations, so we’d really like to know which to put more credence in.
    Priors
    Going in, I think we should have a prior that capabilities getting worse with increasing scale is unlikely. We know of specific examples thanks to the inverse scaling prize, but these are rare exceptions to the general rule of increasing capabilities with increasing scale. As a result, a decreasing deceptive capabilities score should be suspicious.
    The flip side is that it’s conceivable that some alignment problems get easier with scale. Larger language models understand more of what we say and value (e.g. Figure 8 of Bai+22), and so in some sense outer-aligning them may be easier than outer-aligning smaller models. It’s not inconceivable that the same happens for inner alignment, though I wouldn’t want to rely on this happening.
    Tests
    We don’t have to rely on priors alone. Here are a few ways we might distinguish between interpretations:
    Use transparency tools and figure out what the model is doing.
    Impair bigger models (e.g. with dropout or random or adversarial perturbations) and see if they get better at some deceptive capabilities. Impairing models shouldn’t improve capabilities, so this would be good evidence that the bigger models are capable of deception but choosing to hide that capability.This experiment is a little tricky to implement, but the core idea is to hope that interventions impair something critical in the model's ability to execute on deception but don't impair all of the deceptive capabilities.Even if the largest models can continue hiding their capabilities despite the impairment, models that are just barely able to hide their capabilities may fail to do so once impaired, which should increase the model size where deceptive capabilities peak.
    Study the models on a very different data distribution. Models that are hiding relevant capabilities may believe they’ve been deployed and reveal their real goals.Once more, while the largest models may not be fooled by this, models just on the cusp of deceptive alignment are more likely to be tricked and so deceptive alignment may manifest as a shift in the peak of deceptive capabilities towards larger models.
    Study which specific capabilities get worse with scale. It’s hard to understand how some capabilities could get worse with scale (e.g. situational awareness, planning ability, etc.) so if those are getting worse that’s really worrying, whereas it’s less obvious if e.g. models should/shouldn’t get more epoch-level myopic with scale.
    Scaling Helps
    Having access to models of many different sizes is helpful here: we can experiment with small models that are bad at deceptive capabilities, use those results as a baseline, and then try to understand what happens in ...
    7 min
  • LW - Paper+Summary: OMNIGROK: GROKKING BEYOND ALGORITHMIC DATA by Marius Hobbhahn
    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: Paper+Summary: OMNIGROK: GROKKING BEYOND ALGORITHMIC DATA, published by Marius Hobbhahn on October 4, 2022 on LessWrong.
    There is a new paper by Liu et al. that claims to have understood the key mechanism underlying grokking (potentially even generalization more broadly).
    They argue: 1. Grokking can be explained via the norm of the weights. They claim that there is a constant level of the weight norm that is optimal for generalization.2. If there is an optimal level of the weight norm, the weight norm of your model after initialization can be either too low, too high or optimal. They claim that grokking is a phenomenon where we initialize the model with a large weight norm and it then slowly walks toward the optimal weight norm and then generalizes.
    3. They also claim that you can get the same results as grokking but much faster if you set the weight norm correctly at every step.4. They set the norm "correctly" by rescaling the weights after each unconstrained optimization step (so after every weight update loss.backward()?!)Implications:
    I think they have found a very important insight of grokking and finding generalizing circuits more broadly.
    I'm still a bit skeptical of some of the claims and results. On some level "just fix your weight norm and the model generalizes" sounds too simple to be true for all tasks.
    I think this result could have big implications but I'm not yet sure whether they are positive or negative. On the one hand, finding generalizing circuits seems to solve some of the problems associated with bad out-of-distribution generalization. On the other hand, it likely speeds up capabilities.
    I'm very unsure about this paper but intuitively it feels important. Thoughts?
    Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org.
    2 min
  • EA - Announcing Charity Entrepreneurship’s ‘23 ideas. Apply now. by KarolinaSarek
    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: Announcing Charity Entrepreneurship’s ‘23 ideas. Apply now., published by KarolinaSarek on October 4, 2022 on The Effective Altruism Forum.
    Animals, Policy & Biosecurity. Start a high-impact charity with CE’s 2023 Incubation Programs
    Applications are now open to the Charity Entrepreneurship Incubation Programs! As part of our scale up, Charity Entrepreneurship will now run two Incubation Programs in 2023. That means twice as many ideas and twice as many career opportunities.
    The first program will run from February 6 to March 31, 2023. The charities we’re most excited to launch are:
    A policy organization focused on banning the import of low-welfare animal products
    A policy organization focused on banning the use of live-bait fish
    An organization working on milkfish welfare issues in the Philippines
    A policy organization advocating for increasing tobacco taxes to reduce tobacco consumption
    An advocacy organization working to increase road traffic safety
    The second cohort will run throughout July and August 2023. The focus of this incubation program will be launching evidence-based charities in the biosecurity/health security space (more below) as well as highly scalable global health & development ideas.
    Your first thought might be that you lack the relevant experience in the specific fields above. That’s okay, our training and launch model doesn’t require you to be an expert. Lucia and Jack (Lead Exposure Elimination Project) weren’t experts in lead elimination, and Nikita and Brandon didn’t know much about food fortification before they started Fortify Health. Yet both of these organizations, and many more we have launched, are on track to becoming field leaders or GiveWell recommended.
    What we do
    Charity Entrepreneurship launches high-impact nonprofits by connecting potential founders like you with effective ideas, training, and funding. This means we spend thousands of research hours to identify highly-effective interventions in chosen cause areas. We then provide you with a two-month intensive training program (all costs covered) to teach you how to run effective charities. We help you pair with a co-founder that will best complement your skills and personality. You finish the program with a proposal for funders that we deliver to our seed network. They grant up to $200,000 USD per project. You can learn more about the program at our Incubation Program website.
    Top animal welfare charity ideas for the February- March 2023 Incubation Program
    For the February-March 2023 Incubation Program, we currently recommend three animal welfare charity ideas: ban low-welfare imports, ban the use of live-bait fish, and milkfish welfare in the Philippines. A deeper report on each of these will be released in the coming months; this post is intended only to give a sense of what the final ideas will look like for applicants to the Incubation Program.
    Our top three ideas are the result of an intensive nine-month research process designed to identify the interventions most likely to succeed as high-impact start-up charities. The process began by listing nearly 275 ideas and gradually narrowing down, looking at the options in more and more depth. In order to assess how promising interventions would be for future charity entrepreneurs, we use a variety of different decision-making tools such as group consensus, weighted factor models, cost-effectiveness analyses, quality of evidence assessments, case study analyses, and expert interviews.
    You can see the full initial list of about 275 ideas, along with a summary of our prioritization process here.
    This process was exploratory and rigorous – but it was not comprehensive. We did not research all 275 ideas in depth, nor do we think the initial 275 ideas are the only good ideas Therefore, there are still potentially good interventions tha...
    19 min
  • EA - Open call for EA stories by arushigupta
    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: Open call for EA stories, published by arushigupta on October 4, 2022 on The Effective Altruism Forum.
    TLDR: Tell us your stories! We are making an open call for EA stories that might be a good fit for a variety of media projects, including: a written profile series on EA websites, video interview series, podcast interviews, or coverage in newspapers and magazines.
    Submit your stories at:.
    Every day, people in the effective altruism movement do an incredible amount of impactful work, directly helping to make the world a better place. As part of our communications strategy, we’d like to find those stories—and share them with the world. The goal is to have a wide variety of media that shows lots of examples of what effective altruism looks like in practice, rather than in theory. Especially with the recent launch of “What We Owe The Future”, there has been a lot of mainstream media discussion of the theoretical and philosophical ideas behind effective altruism and longtermism. Our goal is to also encourage more coverage and discussion of the practical work that people in, or inspired by, the effective altruism community are doing in the real world, and the kinds of impact they are creating.
    We’d like your suggestions of stories that we should try to share. We’re interested in a wide range of stories, including personal career journeys, stories of failure, big successes that have been accomplished and potential big successes that are in the works. These could be stories about individuals, or organizations. Ideally, we would love stories across a wide range of cause areas. You can either share your own story, or nominate other people/stories that you think we should consider (but ideally these would be people and projects that are comfortable with being more widely known).
    Some examples of what these stories could be used for include: a written profile series on EA websites, video interview series, interviews on EA and non-EA podcasts, or coverage in newspapers and magazines.
    We will reach out to some of the people/organizations whose stories seem like a good fit and help them move forward with sharing their stories more widely. Examples of how we will help include: connecting them with resources like media training, connecting them with journalists, and/or helping to develop a communications strategy.
    The stories we’re most likely to find exciting probably have some combination of the following qualities:
    A compelling story/protagonist (eg. a clear conflict, or interesting journey)
    Work that illustrates EA values or thinking styles
    Relatively easy for a wide audience to understand why it's important or exciting
    New, interesting data that is well-suited to visualizations
    If you’re unsure about your story, we encourage you to have a low bar for submission!
    Team Involved:
    Shakeel Hashim is the head of communications at CEA, and is focused on communicating EA ideas accurately outside EA. He will be leading this project.
    Mike Levine is a Principal at TSD, a strategic communications firm that has worked with Open Philanthropy for several years.
    Arushi Gupta was a recent EA Communications Fellow, and formerly the Co-Director of Effective Altruism NYC.
    The project is still early right now and we’d appreciate any feedback in the comments on how you think we can best approach this, or what you’d like to see us do with these stories. You can also reach out to [email protected] with any feedback, questions, or story submissions.
    Go forth and send us your stories!
    Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org.
    4 min
  • LW - A review of the Bio-Anchors report by jylin04
    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: A review of the Bio-Anchors report, published by jylin04 on October 3, 2022 on LessWrong.
    This is a linkpost for a review of Ajeya Cotra's Biological Anchors report (see also update here) that I wrote in April 2022. It's since won a prize from the EA criticism and red-teaming contest, so I thought it might be good to share here for further discussion.
    Here's a summary from the judges of the red-teaming contest:
    This is a summary and critical review of Ajeya Cotra’s biological anchors report on AI timelines. It provides an easy-to-understand overview of the main methodology of Cotra’s report. It then examines and challenges central assumptions of the modelling in Cotra’s report. First, the review looks at reasons why we might not expect 2022 architectures to scale to AGI. Second, it raises the point that we don’t know how to specify a space of algorithmic architectures that contains something that could scale to AGI and can be efficiently searched through (inability to specify this could undermine the ability to take the evolutionary anchors from the report as a bound on timelines).
    Note that a link to a summary/review of the book Principles of Deep Learning Theory on page 8 is currently offline. I plan to put it on LW later this week. (Relatedly, I may not be around to engage much with comments right away, but I'll be back later in the week.)
    Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org.
    2 min