The Nonlinear Library

The Nonlinear Library

By The Nonlinear FundEducation
Download on the App Store

All The Nonlinear Library episodes

  • EA - It was probably hard to hedge financial risk to EA by Stan van Wingerden
    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: It was probably hard to hedge financial risk to EA, published by Stan van Wingerden on January 19, 2023 on The Effective Altruism Forum.
    Summary
    Hedging essentially means reducing the risk of your assets. [Link]
    Deciding whether or not to hedge (future) cashflow requires making a risk-reward tradeoff. [Link]
    Both the risks and the rewards of your hedges are uncertain & hard to guess for a variety of reasons. [Link]
    I think it was probably hard / expensive to safeguard EA crypto funds against the big crypto downswing, but this depends on the circumstances (which I know little about). [Link]
    Intro
    The financial situation of EA has significantly worsened over the last few months. This is not only due to the FTX debacle, but also due to a broad downswing in cryptocurrency and tech stock prices. It’s worth investigating whether the financial risk of these events to EA organisations could have been mitigated somehow. This risk has many components, but the main financial risk issue to EA organisations I’m discussing here is the combination of multi-year $-denominated spending plans with future donations without fixed $ values. These cases show up all the time, be it from expected donations from companies or from broad financial exposure to different sectors, so one might wonder if EA organisations could have hedged some part of these future donations.
    This post deals with what it means to reduce risk through hedging (crypto) exposure, why one would want to, and why doing so might be undesirable or difficult in practice. I’ve tried to keep it non-technical, as it is intended to help a discussion on whether or not EA charities should have hedged their crypto exposure. I think most of this might still apply to stocks, but I’m no expert in that so I’m not discussing that. Definitely do not take any of this as financial or legal advice.
    What is hedging?
    Hedging essentially means reducing the risk of your assets. This can be done in a variety of ways, the most common of which use derivatives to place a bet that pays off if the price of the underlying crypto token goes down. For example, if you have 1 Bitcoin worth ~17000$ today but possibly less in 2024, you can agree with someone you trust to sell this bitcoin to them in 2024 for 17000$. This agreement to sell in the future functions as a hedge against price decreases of bitcoin: even if bitcoin drops in value, you’d still get 17k in 2024. (On the flip side, if bitcoin were to increase in price you would miss out on that as well.) One way to think about a hedge is as an insurance policy against bad market situations.
    You could also of course sell the bitcoin now in order to get rid of the risk in 2024. I’m going to ignore this possibility and focus instead on hedging future income, as I feel that those situations are most relevant to EA. In the above example, this would mean the bitcoin donation is made in 2024, you don’t have access to it beforehand, but you want to get rid of the bitcoin price risk in the meantime.
    A more uncertain but potentially more valuable hedge would use assets that are correlated with the amount you’re expected to be donated. If you both 1. expect to be donated large amounts of money by some company deeply involved in cryptocurrency and 2. expect the health of this company to be strongly related to the price of Bitcoin, then you can ‘hedge’ your expected future donations by taking on a bitcoin short. After all, if the company goes under, then you expect the price of bitcoin to drop, and therefore your hedge should pay off in that scenario, allowing you to still receive some money in the place of your missed donations. Compared to the above example of hedging a to-be-donated bitcoin, this example is not a strictly risk-lowering hedge, as your correlation assumptions can turn out to be wrong. Still, this ...
    15 min
  • LW - [Cross-post] Is the Fermi Paradox due to the Flaw of Averages? by Aryeh Englander
    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: [Cross-post] Is the Fermi Paradox due to the Flaw of Averages?, published by Aryeh Englander on January 18, 2023 on LessWrong.
    [This article is copy-pasted from the Lumina blog, very lightly edited for LessWrong.]
    Where is everybody?— Enrico Fermi
    The omnipresence of uncertainty is part of why making predictions and decisions is so hard. We at Lumina advocate treating uncertainty explicitly in our models using probability distributions. Sadly this is not yet as common as it should be. A recent paper “Dissolving the Fermi Paradox” (2018) is a powerful illustration of how including uncertainty can transform conclusions on the fascinating question of whether our Earth is the only place in the Universe harboring intelligent life. The authors, Anders Sandberg, Eric Drexler and Toby Ord (whom we shall refer to as SDO), show elegantly that the apparent paradox is simply the result of the mistake of ignoring uncertainty, what Sam L. Savage calls the Flaw of Averages. In this article, we review their article and link to an Analytica version of their model that you can explore.
    The Fermi Paradox
    Enrico Fermi. From Wikimedia commons.
    One day in 1950, Enrico Fermi, the Nobel prize-winning builder of the first nuclear reactor, was having lunch with a few friends in Los Alamos. They were looking at a New Yorker cartoon of cheerful aliens emerging from a flying saucer. Fermi famously asked “Where is everybody?”. Given the vast number of stars in the Milky Way Galaxy and the likely development of life and extraterrestrial intelligence, how come no ETs have come to visit or at least been detected? This question came to be called the “Fermi Paradox”. Ever since, it has bothered those interested in the existence of extraterrestrial intelligence and whether we are alone in the Universe.
    The Flaw of Averages on Steroids
    Dr. Sam Savage who coined the term “Flaw of Averages”
    Sam L. Savage, in his book, The Flaw of Averages, shows how ignoring uncertainty and just working with a single mean or “most likely” value for each uncertain quantity can lead to misleading results. To illustrate how dramatically this approach can distort your conclusions, SDO offer a toy example. Suppose there are nine factors that multiplied together give the probability of extraterrestrial intelligence (ETI) arising on any given star. If you use a point estimate of 0.1 for each factors, you could infer that there is a 10−9probability of any given star harboring ETI. There are about 1011 stars in the Milky Way, so the probability that no star other than our own has a planet harboring intelligent life would be extremely small, (1−10−9)100B≈3.7×10−44. On the other hand, suppose that, based on what we know, each factor could be anywhere between 0 and 0.2, and assign a uniform uncertainty over this interval, using the probability distribution, Uniform(0, 0.2).
    If you combine these distributions probabilistically, using Monte Carlo simulation for example, the mean of the result is 0.21 – over 5,000,000,000,000,000,000,000,000,000,000,000,000,000,000 times more likely!
    The Drake Equation
    Frank Drake, a radio astronomer who worked on the search for extraterrestrial intelligence (SETI), tried to formalize Fermi’s estimate of the number of ETIs. He suggested that we can estimate N, the number of detectable, intelligent civilizations in the Milky Way galaxy from what is now called the “Drake equation”. It is sometimes referred to as the “second most-famous equation in science (after E= mc2)”:
    Frank Drake (1930-2022).
    N=R∗×fp×ne×fl×fi×fc×L
    Where
    R∗ is the average rate of formation of stars in our galaxy,fp is the fraction of stars with planets.ne is the average number of those planets that could potentially support life.fl is the fraction of those on which life had actually developed;fi is the fraction of those with life that ...
    28 min
  • EA - Evaluating StrongMinds: how strong is the evidence? by JoelMcGuire
    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: Evaluating StrongMinds: how strong is the evidence?, published by JoelMcGuire on January 19, 2023 on The Effective Altruism Forum.
    A recent post by Simon_M argued that StrongMinds should not be a top recommended charity (yet), and many people seemed to agree. While I think Simon raised several useful points regarding StrongMinds, he didn't engage with the cost-effectiveness analysis of StrongMinds that I conducted for the Happier Lives Institute (HLI) in 2021 and justified this decision on the following grounds:
    “Whilst I think they have some of the deepest analysis of StrongMinds, I am still confused by some of their methodology, it’s not clear to me what their relationship to StrongMinds is.”.
    By failing to discuss HLI’s analysis, Simon’s post presented an incomplete and potentially misleading picture of the evidence base for StrongMinds. In addition, some of the comments seemed to call into question the independence of HLI’s research. I’m publishing this post to clarify the strength of the evidence for StrongMinds, HLI’s independence, and to acknowledge what we’ve learned from this discussion.
    I raise concerns with several of Simon’s specific points in a comment on the original post. In the rest of this post, I’ll respond to four general questions raised by Simon’s post that were too long to include in my comment. I briefly summarise the issues below and then discuss them in more detail in the rest of the post
    1. Should StrongMinds be a top-rated charity? In my view, yes. Simon claims the conclusion is not warranted because StrongMinds’ specific evidence is weak and implies implausibly large results. I agree these results are overly optimistic, so my analysis doesn’t rely on StrongMind’s evidence alone. Instead, the analysis is based mainly on evidence synthesised from 39 RCTs of primarily group psychotherapy deployed in low-income countries.
    2. When should a charity be classed as “top-rated”? I think that a charity could be considered top-rated when there is strong general evidence OR charity-specific evidence that the intervention is more cost-effective than cash transfers. StrongMinds clears this bar, despite the uncertainties in the data.
    3. Is HLI an independent research institute? Yes. HLI’s mission is to find the most cost-effective giving opportunities to increase wellbeing. Our research has found that treating depression is very cost-effective, but we’re not committed to it as a matter of principle. Our work has just begun, and we plan to publish reports on lead regulation, pain relief, and immigration reform in the coming months. Our giving recommendations will follow the evidence.
    4. What can HLI do better in the future? Communicate better and update our analyses. We didn’t explicitly discuss the implausibility of StrongMinds’ data in our work. Nor did we push StrongMinds to make more reasonable claims when we could have done so. We acknowledge that we could have done better, and we will try to do better in the future. We also plan to revise and update our analysis of StrongMinds before Giving Season 2023.
    1. Should StrongMinds be a top-rated charity?
    I agree that StrongMinds’ claims of curing 90+% of depression are overly optimistic, and I don’t rely on them in my analysis. This figure mainly comes from StrongMinds’ pre-post data rather than a comparison between a treatment group and a control. These data will overstate the effect because depression scores tend to decline over time due to a natural recovery rate. If you monitored a group of depressed people and provided no treatment, some would recover anyway.
    My analysis of StrongMinds is based on a meta-analysis of 39 RCTS of group psychotherapy in low-income countries. I didn’t rely solely on StrongMinds’ own evidence alone, I incorporated the broader evidence base from other similar interventions t...
    10 min
  • AF - 200 COP in MI: Studying Learned Features in Language Models by Neel Nanda
    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: 200 COP in MI: Studying Learned Features in Language Models, published by Neel Nanda on January 19, 2023 on The AI Alignment Forum.
    This is the final post in a sequence called 200 Concrete Open Problems in Mechanistic Interpretability. Start here, then read in any order. If you want to learn the basics before you think about open problems, check out my post on getting started. Look up jargon in my Mechanistic Interpretability Explainer
    Motivation
    Motivating Paper: Softmax Linear Units (SoLU), Multimodal Neurons
    To accompany this post, I’ve created a website called Neuroscope that displays the text that most activates each neuron in some language models - check it out!
    This section contains a lot of details and thoughts on thinking about neurons and learned features and relationship to the surrounding literature. If you get bored, feel free to just skip to exploring and looking for interesting neurons in Neuroscope
    MLPs represent ⅔ of the parameters in a transformer, yet we really don’t understand them very well. Based on our knowledge of image models, my best guess is that models learn to represent features, properties of the input, with different directions corresponding to different features. Early layers learn simple features that are basic functions of the input, like edges and corners, and these are iteratively refined and built up into more sophisticated features, like angles and curves and circles. Language models are messier than those image models, since they have attention layers and a residual stream, but our guess is that analogous features are generally computed in MLP layers.
    But there’s a lot of holes and confusions in this framework, and I’d love to have these filled in! How true actually is this in practice? Do features correspond to neurons vs arbitrary directions? What kinds of features get learned, and where do they occur in a model? What features do we see in small models vs large ones vs enormous ones? What kinds of things are natural for a language model to express vs extremely hard and convoluted? What are the ways our intuitions will trip us up here?
    Issues like polysemanticity and superposition make it difficult to actually reverse engineer specific neurons, and it seems clear that the model can learn features that do not correspond to specific neurons. But even if we relax our standards of rigour and just want to understand what features are present at all, we don’t know much about what role these layers actually serve in the model! There’s been a fair amount of work studying this for BERT, but very little on studying this in generative language models, especially large ones!
    There’s a bunch of angles to make progress on this, and I’m excited to see a diverse range of approaches! But here I’m going to focus on the somewhat unreliable but useful technique of looking at max activating dataset examples: running the model across a bunch of data and for each neuron tracking the texts that most activated it. If we see a consistent pattern (or patterns) in these texts we can infer that the model is detecting the feature represented by that pattern. To emphasise, a consistent pattern in a neuron’s examples does not mean the neuron only represents that feature. But it can be pretty good evidence that the feature is represented in the model, and that that neuron responds strongly to it!
    To help you explore what’s inside a language model, I’ve created a tool called Neuroscope which displays the max activating dataset examples for a range of language models. Each neuron has its own page, and there are some pretty weird ones in there! For example, take neuron 456 in my one layer SoLU model - this activates on the word " the" after " love", on what seems to be the comments section of cutesy arts and crafts blogs.
    One vision for an accessible, beginner-fr...
    47 min
  • EA - Possible changes to EA, a big upvoted list by Nathan Young
    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: Possible changes to EA, a big upvoted list, published by Nathan Young on January 18, 2023 on The Effective Altruism Forum.
    We should put all possible changes/reforms in a big list, that everyone can upvote/downvote, agree disagree.
    EA is governed but a set of core EAs, so if you want change, I suggest that giving them less to read and a strong signal of community consensus is good.The top-level comments should be a short clear explanation of a possible change. If you want to comment on a change, do it as a reply to the top level comment
    This other post gives a set of reforms, but they are a in a big long list at the bottom. Instead we can have a list that changes by our opinions!
    Note that I do not agree with all comments I post here.
    Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org.
    1 min
  • LW - Flying With Covid by jefftk
    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: Flying With Covid, published by jefftk on January 18, 2023 on LessWrong.
    I flew out to CA for a
    gig this weekend, and two of my kids
    came along. We got into SFO late Thursday night,
    drove to
    San Jose to sleep, and got to the gig on Friday. We took rapid tests, all negative. A late night, and the kids fell asleep backstage in the keyboard case:
    Saturday morning one was mostly fine, but the other was sleepy ("jet lag"). She stayed sleepy all day, napping more backstage, and she wasn't eating much ("different food than she usually eats"). After choosing what food we should get for dinner she wouldn't eat it and was cranky ("seems more 'sick' than 'jet lag'"). Another rapid test [1] immediate positive:
    So I'm 3,000 miles from home with two kids, one of which has the contagious disease everyone's spent the last three years avoiding. We're scheduled to check out by noon tomorrow, drive back to
    SFO, and take a red-eye home. What do I do?
    Two options:
    Isolate in place.
    Fly home anyway.
    Earlier in the pandemic this would have been a clear choice: isolate in place. It was required by public health order, and especially before vaccination traveling would have been too much risk to others and contribution to pandemic spread.
    On the other hand, pre-covid I think almost everyone would have gone home. If you read advice it was mostly about how to decide whether to cancel a vacation or how to keep a sick kid comfortable on a plane. While there's some "if you're sick, don't fly", the examples were whether to begin a trip, not whether to return home. Discussion of "well enough to fly" was about whether you were up to it, not about risk to others. And of course there's no discussion of when you should wear a mask to protect others in the terminal or on the plane.
    At this point, however, we are partway through a transition to treating covid as a normal disease. While the CDC recommendation is still not to fly, this was also what they recommended for people with similar illness pre-covid, and they recommend other things that most people aren't doing like masking when local covid levels are high. Airlines are back to charging change fees even with a positive test.
    Isolating in place is at least four more days in CA, more if someone else gets sick (which did happen). Maybe six days in expectation? With flight change fees ($300), rental car extension ($400), and more days in the hotel ($600) this is $1,300+. Possibly we need a second room ($600) to separate the sick and well. Then there's dealing with food (probably delivery, more money) and work (with my wrist and neck issues this would need a more ergonomic setup than a laptop on a desk, so either more money or not working). It's possible, we could do it if we had to, but it's a lot.
    I decided we'd fly home. We stayed as far away from others as possible in the terminals and kept our (K)N95s on the whole time—normally I let the kids take their masks off on the plane once the HEPA-filtered air is running and I've adjusted their blowers to 'full'. We got home Monday morning, and have been isolating since, including from each other.
    Was this the right choice? I still don't know. I don't want to be a chump who expends inordinate effort to follow guidelines strictly when
    most of society has moved on, but there are also people still working very hard not to get covid who risk serious consequences if they were to get sick. [2] I'm not sure how to balance these, and the decision to come home rather than take on the logistics of isolating with kids in a far-off hotel was in part a selfish one. I'm writing this up not to encourage or discourage similar choices but to walk through how I was thinking about it and humanize a choice I expect some of you think was reckless and inconsiderate.
    [1] There was nothing that required us to test here. If you'd t...
    5 min
  • AF - Gradient Filtering by Arun Jose
    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: Gradient Filtering, published by Arun Jose on January 18, 2023 on The AI Alignment Forum.
    Several of the ideas in this post originated in conversations between the authors months ago. I did the work involved in exploring them and thinking of new ideas and framings, as well as writing this post; old writings by Janus were the inspiration for some of the directions of thought here.
    Thanks to Adam Jermyn and Paul Colognese for their feedback and thoughts.
    Generally speaking, gradient hacking can seem pretty hard for an agent to do. But the dangers don’t lie in the galaxy-brained strategies that models with strong deceptive capabilities can implement, it’s in the idea that a model’s gradient can be influenced by an inner agent at all. So examining variants that are seemingly more accessible would be pretty useful.
    This post discusses gradient filtering, one such variant where a model is made to generate outputs that are wrong in ways that lead gradient descent to update the model in a direction desirable by some malicious agent. It aims broadly to make the case for gradient hacking being more plausible than previously discussed - because unlike other imagined variants gradient filtering does not require the agent to know or understand the policy's internal implementation - and also to explore some aspects of simulator theory.
    First, we lay out a scenario where gradient filtering could play a role, and how this benefits the model. Then, we discuss specific conditions necessary in its arisal, specifically in the context of deceptive simulacra (although the points should be generally applicable), as well as arguments both for and against them being fulfilled. Finally, we talk about how simulacra may need to be focused by the model to be capable of coming up with this kind of strategy.
    I think this is a topic of some importance because few of the conditions for some form of gradient hacking arising are very restrictive, and the rest still have plausible pathways in which they could arise. Further, gradient hacking is likely to be pretty dangerous if feasible, reducing the likelihood that deceptive agents would be solved by default, making deceptive agents more powerful, and requiring potential solutions to be even stronger.
    That said, there are definitely arguments against its plausibility and ultimate danger, as we’ll see. I think they’re far from being strong enough to dismiss this as a problem, but they make the question ambiguous enough that I’m not making any strong claims in this post on whether gradient hacking will definitely be a problem - there isn’t a concluding section for the same reason, to allow you to draw your inferences based on the arguments.
    There’s also the question of whether this is really worth thinking about in comparison to other important and more pressing problems, which is really good to ask in general. I think that ignoring likelihood, the dangers merit it a lot more consideration, and the likelihood is far from being definitively low enough to dismiss. The question of gradient hacking is especially relevant to certain classes of alignment proposals (for example, interpretability approaches that inspect models post-training are vulnerable to deceptive agents gradient hacking themselves into cryptographically-sound obfuscation), and is therefore pretty important to keep in mind.
    Introduction
    I think it’s reasonable to assume that powerful AI would be able to do creative things with neural networks that we would find hard to prevent (and in some cases, predict). Gradient hacking is one class of strategies that comes under this, where a misaligned mesa-optimizer can leverage some property of the training process to gain partial control over what direction future updates will lead the model toward - which it could use to prevent the model from updat...
    24 min
  • EA - The EA community does not own its donors' money by Nick Whitaker
    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: The EA community does not own its donors' money, published by Nick Whitaker on January 18, 2023 on The Effective Altruism Forum.
    A number of recent proposals have detailed EA reforms. I have generally been unimpressed with these - they feel highly reactive and too tied to attractive sounding concepts (democratic, transparent, accountable) without well thought through mechanisms. I will try to expand my thoughts on these at a later time.
    Today I focus on one element that seems at best confused and at worst highly destructive: large-scale, democratic control over EA funds.
    This has been mentioned in a few proposals: It originated (to my knowledge) in Carla Zoe Cremer's Structural Reforms proposal:
    Within 5 years: EA funding decisions are made collectively
    First set up experiments for a safe cause area with small funding pots that are distributed according to different collective decision-making mechanisms
    (Note this is classified as a 'List A' proposal - per Cremer: "ideas I’m pretty sure about and thus believe we should now hire someone full time to work out different implementation options and implement one of them")
    It was also reiterated in the recent mega-proposal, Doing EA Better:
    Within 5 years, EA funding decisions should be made collectively
    Furthermore (from the same post):
    Donors should commit a large proportion of their wealth to EA bodies or trusts controlled by EA bodies to provide EA with financial stability and as a costly signal of their support for EA ideas
    And:
    The big funding bodies (OpenPhil, EA Funds, etc.) should be disaggregated into smaller independent funding bodies within 3 years
    (See also the Deciding better together section from the same post)
    How would this happen?
    One could try to personally convince Dustin Moskovitz that he should turn OpenPhil funds over to an EA Community panel, that it would help OpenPhil distribute its funds better.
    I suspect this would fail, and proponents would feel very frustrated.
    But, as with other discourse, these proposals assume that because a foundation called Open Philanthropy is interested in the "EA Community" that the "EA Community" has/deserves/should be entitled to a say in how the foundation spends their money. Yet the fact that someone is interested in listening to the advice of some members of a group on some issues does not mean they have to completely surrender to the broader group on all questions. They may be interested in community input for their funding, via regranting for example, or invest in the Community, but does not imply they would want the bulk of their donations governed by the EA community.
    (Also - I'm using scare quotes here because I am very confused who these proposals mean when they say EA community. Is it a matter of having read certain books, or attending EAGs, hanging around for a certain amount of time, working at an org, donating a set amount of money, or being in the right Slacks? These details seem incredibly important when this is the set of people given major control of funding, in lieu of current expert funders)
    So at a basic level, the assumption that EA has some innate claim to the money of its donors is basically incorrect. (I understand that the claim is also normative). But for now, the money possessed by Moskovitz and Tuna, OP, and GoodVentures is not the property of the EA community. So what, then, to do?
    Can you demand ten billion dollars?
    Say you can't convince Moskovitz and OpenPhil leadership to turn over their funds to community deliberation.
    You could try to create a cartel of EA organizations to refuse OpenPhil donations. This seems likely to fail - it would involve asking tens, perhaps hundreds, of people to risk their livelihoods. It would also be an incredibly poor way of managing the relationship between the community and its most generous funder--and...
    8 min
  • EA - Exceptional Research Award by Effective Thesis (ETERA): reflection by Effective Thesis
    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: Exceptional Research Award by Effective Thesis (ETERA): reflection, published by Effective Thesis on January 18, 2023 on The Effective Altruism Forum.
    This post was written by David Janků with contributions from Adéla Novotná (graphs and analyses on the background) and useful edits by Sophie Kirkham.
    Summary
    In autumn 2022, Effective Thesis held the Exceptional Research Award to encourage and recognize students conducting promising research that has the potential to significantly improve the world.
    We received 35 submissions, had 25 of them reviewed and chose 6 main prize winners and 4 commendation prize winners (see the winning submissions and winners’ profiles here).
    We received applications from across a variety of study disciplines and cause areas, though causes that have traditionally been prioritised by EA leaders - such as AI alignment, biosecurity, and animal welfare - were relatively underrepresented in our sample. It is unclear to us why this is the case (e.g. whether people working on those causes already have existing cause-specific prizes to apply to; or have different publication norms - like publishing mainly via blog posts - that our award is not compatible with; or to what extent some of these directions are not bottlenecked by the creation of new research ideas and understanding but rather by execution of existing ideas). We think this represents room for improvement in the future.
    Only 57 % of all applicants and 50 % of all winners had any engagement with Effective Thesis prior to applying for the award, suggesting the award was able to attract students who are doing valuable work but who have not interacted with Effective Thesis before.
    That being said, people previously coached by Effective Thesis scored higher in the reviews on “potentially having more impact with their research” than people who were not coached, suggesting our coaching might help people arrive at better ideas for research topics.
    The review scores indicated that the quality of the application pool was fairly high. The highest scores were usually given for rigour and novelty, and the lower scores were for the ability to make progress in the research field and potential positive impact in the world. This is to be expected since achieving impact is harder than achieving rigour and novelty.
    Both reviewers and applicants (successful and unsuccessful) seemed to be happy about their participation in this award, as reflected by their willingness to be associated with the award publicly, high ratings of the usefulness of the reviews by applicants, and a number of qualitative comments we received.
    What is ETERA and why did we decide to run it?
    The Effective Thesis Exceptional Research Award (ETERA) has been established to encourage and recognize promising research, conducted by students from undergraduate to PhD level, that has the potential to significantly improve the world. We accepted submissions of theses, dissertations and capstone papers. Entries were required to have been submitted to a university between 1st January 2021 - 1st September 2022 and to relate to one or multiple prioritised research directions on our site. We accepted entries from all countries, though all entries had to be submitted in English.
    A prize of 1000 USD was offered for winning entries (and 100 USD for commendation prizes). In addition, prize-winning entries would be featured on our website and promoted by us.
    We expected the award to have three primary positive impacts:
    Providing a career boost for aspiring researchers already focused on very impactful questions and producing high quality work. Winning this award would give them additional credentials and increase their chances of building a successful research career.
    Motivating additional aspiring researchers to begin exploring questions in some of the ...
    22 min
  • EA - 2022 EA conference talks are now live by Eli Nathan
    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: 2022 EA conference talks are now live, published by Eli Nathan on January 18, 2023 on The Effective Altruism Forum.
    Recordings from various 2022 EA conferences are now live on our YouTube channel, these include talks from London, San Francisco, Washington, D.C., EAGxBoston, EAGxOxford, EAGxBerlin, and EAGxVirtual (alongside many other talks from previous years).
    Listening to talks can be a great way to learn more about EA and stay up to date on EA cause areas, and recording them allows people who couldn’t attend (or who were busy in 1:1 meetings) to watch them in their own time. Recordings from other EA conferences will likely be live on our channel soon, and we recommend subscribing if you’d like to be notified of these.
    Some highlighted talks are displayed below:
    EA Global: London
    Presenting big ideas & complex data to the public | Edouard Mathieu and Hannah Ritchie
    In this talk, Edouard Mathieu discusses the lessons of data communication in the COVID-19 pandemic. Hannah and Edouard then have a fireside chat and Q&A on Our World in Data and how it fits with the EA framework.
    The state of aquatic animal advocacy | Sophika Kostyniuk, Andrés Jiménez Zorrilla, Alex Holst, and Bruce Friedrich
    Addressing aquatic animal welfare is important, as it is highly neglected, and tractable. Estimates vary, but there are approximately 100 billion fin fish and 350–400 billion shrimps farmed annually, which is far more than all of the land animals combined (more than 7x as many at the upper estimation). For the most part, farmed aquatic animals are treated like inanimate objects — their suffering is almost unimaginable.
    This session discusses why aquatic animal welfare is critical to address, and some of the priority interventions that can alleviate vast amounts of suffering.
    EA Global: San Francisco
    Betting on AI is like betting on semiconductors in the 70's | Danny Hernandez
    Danny discusses the three exponentials driving AI progress: hardware, algorithmic, and spending. He considers extrapolating these trends 10–20 years out and translates effective compute progress into GPT-2 to GPT-3 sized jumps and builds an intuition for such jumps. Danny uses the extrapolations and jump intuitions to think about what capabilities normal progress in effective compute are likely to yield. This session is likely to be particularly relevant to people very concerned about AI, considering working in AI, or choosing an agenda within AI.
    Science of scaling | Ahmed Mushfiq Mobarak and Heidi McAnnally-Linz
    Mushfiq Mobarak and Heidi McAnnally-Linz speak to learning from cutting edge research on the science of scaling using examples from scaling interventions targeted at addressing seasonal poverty and from the NORM model targeted at increasing community-level mask wearing during the COVID-19 pandemic. They draw on these experiences to make the case for innovation through research as well as working with existing at-scale partners. They emphasize the importance of using direct evidence from RCTs as well as exploring other complexities of scale such as national impacts, spillovers, etc.
    EA Global: Washington, D.C.
    Safeguarding modern bioscience & biotechnology to prevent catastrophic biological events | Jaime Yassif, Beth Cameron, and Jaspreet Pannu
    Bioscience and biotechnology advances are vital for fighting disease, protecting the environment, and promoting economic development — and they hold incredible promise. However, these innovations can also pose unique challenges — increasing the risks of accidental misuse or deliberate abuse with potentially catastrophic global consequences.
    This session begins with a talk by Dr. Jaime Yassif highlighting these issues and discussing effective strategies for improving bioscience governance and reducing emerging biorisks — including through the establishment of t...
    6 min