The Nonlinear Library

The Nonlinear Library

By The Nonlinear FundEducation
Download on the App Store

All The Nonlinear Library episodes

  • LW - Maze-solving agents: Add a top-right vector, make the agent go to the top-right by TurnTrout
    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: Maze-solving agents: Add a top-right vector, make the agent go to the top-right, published by TurnTrout on March 31, 2023 on LessWrong.
    Overview: We modify the goal-directed behavior of a trained network, without any gradients or finetuning. We simply add or subtract "motivational vectors" which we compute in a straightforward fashion.
    In the original post, we defined a "cheese vector" to be "the difference in activations when the cheese is present in a maze, and when the cheese is not present in the same maze." By subtracting the cheese vector from all forward passes in a maze, the network ignored cheese.
    I (Alex Turner) present a "top right vector" which, when added to forward passes in a range of mazes, attracts the agent to the top-right corner of each maze. Furthermore, the cheese and top-right vectors compose with each other, allowing (limited but substantial) mix-and-match modification of the network's runtime goals.
    I provide further speculation about the algebraic value editing conjecture:
    It's possible to deeply modify a range of alignment-relevant model properties, without retraining the model, via techniques as simple as "run forward passes on prompts which e.g. prompt the model to offer nice- and not-nice completions, and then take a 'niceness vector', and then add the niceness vector to future forward passes."
    I close by asking the reader to make predictions about our upcoming experimental results on language models.
    This post presents some of the results in this top-right vector Google Colab, and then offers speculation and interpretation. I produced the results in this post, but the vector was derived using a crucial observation from Peli Grietzer. A lot of the content and infrastructure was made possible by my MATS 3.0 team: Ulisse Mini, Peli Grietzer, and Monte MacDiarmid. Thanks also to Lisa Thiergart, Aryan Bhatt, Tamera Lanham, and David Udell for feedback and thoughts.
    Background
    This post is straightforward, as long as you remember a few concepts:
    Vector fields, vector field diffs, and modifying a forward pass. AKA you know what this figure represents:
    How to derive activation-space vectors (like the "cheese vector") by diffing two forward passes, and add / subtract these vectors from future forward passes
    AKA you can understand the following: "We took the cheese vector from maze 7. ~Halfway through the forward passes, we subtract it with coefficient 5, and the agent avoided the cheese."
    If you don't know what these mean, read this section. If you understand, then skip.
    Langosco et al. trained a range of maze-solving nets. We decided to analyze one which we thought would be interesting. The network we chose has 3.5M parameters and 15 convolutional layers.
    Sampling rollouts from the trained policy adds a lot of noise. A nicer way to view episodes is with a vector field view, which overlays a vector field representing the agent policy for a given maze.
    To compute the cheese vector, we
    Generate two observations—one with cheese, and one without. The observations are otherwise the same.
    Run a forward pass on each observation, recording the activations at each layer.
    For a given layer, define the cheese vector to be CheeseActivations - NoCheeseActivations. The cheese vector is a vector in the vector space of activations at that layer.
    Let's walk through an example, where for simplicity the network has a single hidden layer, taking each observation (shape (3, 64, 64) for the 64x64 RGB image) to a two-dimensional hidden state (shape (2,)) to a logit vector (shape (15,)).
    We run a forward pass on a batch of two observations, one with cheese (note the glint of yellow in the image on the left!) and one without (on the right).
    We record the activations during each forward pass. In this hypothetical,
    CheeseActivations := (1, 3)
    NoCheeseActivations := (0, 2)...
    18 min
  • AF - Maze-solving agents: Add a top-right vector, make the agent go to the top-right by Alex Turner
    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: Maze-solving agents: Add a top-right vector, make the agent go to the top-right, published by Alex Turner on March 31, 2023 on The AI Alignment Forum.
    Overview: We modify the goal-directed behavior of a trained network, without any gradients or finetuning. We simply add or subtract "motivational vectors" which we compute in a straightforward fashion.
    In the original post, we defined a "cheese vector" to be "the difference in activations when the cheese is present in a maze, and when the cheese is not present in the same maze." By subtracting the cheese vector from all forward passes in a maze, the network ignored cheese.
    I (Alex Turner) present a "top right vector" which, when added to forward passes in a range of mazes, attracts the agent to the top-right corner of each maze. Furthermore, the cheese and top-right vectors compose with each other, allowing (limited but substantial) mix-and-match modification of the network's runtime goals.
    I provide further speculation about the algebraic value editing conjecture:
    It's possible to deeply modify a range of alignment-relevant model properties, without retraining the model, via techniques as simple as "run forward passes on prompts which e.g. prompt the model to offer nice- and not-nice completions, and then take a 'niceness vector', and then add the niceness vector to future forward passes."
    I close by asking the reader to make predictions about our upcoming experimental results on language models.
    This post presents some of the results in this top-right vector Google Colab, and then offers speculation and interpretation.
    I produced the results in this post, but the vector was derived using a crucial observation from Peli Grietzer. Lisa Thiergart independently analyzed top-right-seeking tendencies, and had previously searched for a top-right vector. A lot of the content and infrastructure was made possible by my MATS 3.0 team: Ulisse Mini, Peli Grietzer, and Monte MacDiarmid. Thanks also to Lisa Thiergart, Aryan Bhatt, Tamera Lanham, and David Udell for feedback and thoughts.
    Background
    This post is straightforward, as long as you remember a few concepts:
    Vector fields, vector field diffs, and modifying a forward pass. AKA you know what this figure represents:
    How to derive activation-space vectors (like the "cheese vector") by diffing two forward passes, and add / subtract these vectors from future forward passes
    AKA you can understand the following: "We took the cheese vector from maze 7. ~Halfway through the forward passes, we subtract it with coefficient 5, and the agent avoided the cheese."
    If you don't know what these mean, read this section. If you understand, then skip.
    Langosco et al. trained a range of maze-solving nets. We decided to analyze one which we thought would be interesting. The network we chose has 3.5M parameters and 15 convolutional layers.
    Sampling rollouts from the trained policy adds a lot of noise. A nicer way to view episodes is with a vector field view, which overlays a vector field representing the agent policy for a given maze.
    To compute the cheese vector, we
    Generate two observations—one with cheese, and one without. The observations are otherwise the same.
    Run a forward pass on each observation, recording the activations at each layer.
    For a given layer, define the cheese vector to be CheeseActivations - NoCheeseActivations. The cheese vector is a vector in the vector space of activations at that layer.
    Let's walk through an example, where for simplicity the network has a single hidden layer, taking each observation (shape (3, 64, 64) for the 64x64 RGB image) to a two-dimensional hidden state (shape (2,)) to a logit vector (shape (15,)).
    We run a forward pass on a batch of two observations, one with cheese (note the glint of yellow in the image on the left!) and one without (on the righ...
    18 min
  • LW - Stop pushing the bus by cousin it
    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: Stop pushing the bus, published by cousin it on March 31, 2023 on LessWrong.
    With the recent proposals about moratoriums and regulation, should we also start thinking about a strike by AI researchers and developers?
    The reasoning I imagine as follows. AI capability is now growing really fast, and toward levels that will strongly affect the world. And AI safety lags behind. (A minute ago I used a ChatGPT jailbreak to get instructions for torturing a pregnant woman, that's the market leader performance for you.) And finally, I want to make the argument that working on AI capability while it is ahead of AI safety, is "pushing the bus".
    Here's the metaphor, a bunch of people including you are pushing a bus full of children toward a precipice, and you're paid for each step. In this situation would you really say "oh I have to keep pushing, otherwise others will get all the money"? It's not like they'll profit from it! Their children will die, along with everyone else's! So there's no game theoretic angle, you can just make the decision alone, to stop pushing the frigging bus.
    To clarify, working on AI isn't always bad. It could lead to a wonderful future for humanity. But when AI safety is behind, as now, then working on AI capability is pushing the bus. There's no good justification for it.
    Hence the strike. Not by leadership but by AI researchers and developers themselves. I imagine a desk plaque saying "while AI safety lags behind AI capability, I refuse to work on AI capability". That's the start condition, and it also tells you when to stop (when safety catches up with current capability, which means not just stopping saying bad things, but for stronger AIs also safe and benevolent behavior more generally). And it's also the restart condition if safety starts lagging behind again.
    Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org.
    2 min
  • EA - Introducing the Maternal Health Initiative by Ben Williamson
    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: Introducing the Maternal Health Initiative, published by Ben Williamson on March 31, 2023 on The Effective Altruism Forum.
    The Maternal Health Initiative (MHI) was launched through the 2022 Charity Entrepreneurship Incubation Programme. Our mission is to improve the lives of women and children through a light-touch, scalable training programme on contraceptive counselling for healthcare workers in sub-Saharan Africa.
    Six months after we started, we’re excited to share more about our work so far, our mission, and the impact we’re aiming for. This post presents a short case for the value of MHI’s work, details our goals for the rest of 2023, and explains how you can support us.
    Introduction
    MHI is an evidence-based non-profit dedicated to improving the lives of women and children across the world by training health workers to provide high-quality family planning counselling in the postpartum period.
    We are currently conducting preliminary piloting work in northern Ghana through local partner organisations. At present, we estimate that our programming could avert a disability-adjusted life year (DALY) for $166, and we foresee promising pathways to significant scale.
    This post provides a brief overview of the rationale for MHI’s work, what we’ve accomplished so far, where we’re heading, and how you can help. For more detailed information about our work, please visit our website, sign up to our mailing list, or reach out to us directly.
    Why MHI was founded
    Every day, 830 women die from pregnancy-related causes. These deaths overwhelmingly occur in sub-Saharan Africa, where maternal mortality rates are more than ten times higher than the next worst region. The Maternal Health Initiative was founded to change this.
    Increasing access to contraception is a tractable intervention with extensive evidence of its effectiveness in reducing maternal mortality at scale. More than this, access to family planning services can produce a wealth of additional impact, providing potentially transformative benefits to women’s autonomy, income, and wellbeing. The autonomy benefits are particularly exciting: family planning appears to be one of the most cost-effective ways of improving agency.
    218 million women in lower income countries have an unmet need for family planning. This resulted in more than 85 million unintended pregnancies in 2019 alone. More than one third of these unintended pregnancies in sub-Saharan Africa end in an unsafe abortion. These often take place outside of a health facility in countries where abortions are dangerous, heavily stigmatised, and in some cases illegal.
    While there is substantial work in this space by non-EA actors, there remain significantly neglected opportunities. One of these is the specific provision of postpartum (post-birth) family planning services. Short-spaced pregnancies - births that occur within 2 years over each other - significantly increase the risks of both maternal and infant mortality.
    A synthesis of the literature evaluating the impact of pregnancies that occur within 2 years of each other suggests a 18% higher risk of infant mortality and 16% higher risk of maternal mortality. Despite this, contraceptive use drops to less than half of the national average in the first year after giving birth, with contraceptive counselling after pregnancy rarely taking place currently.
    MHI is developing a programme of training to ensure high-quality contraceptive counselling consistently takes place in the postpartum period. Several randomised control trials have demonstrated the effectiveness of increased quality of care during this period, and we currently estimate that our program can avert a disability-adjusted life year (DALY) for just $166, competitive with GiveWell’s top recommended charities.
    Our work
    What we’ve done so far
    Since September, we...
    10 min
  • EA - Critiques of prominent AI safety labs: Redwood Research by Omega
    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: Critiques of prominent AI safety labs: Redwood Research, published by Omega on March 31, 2023 on The Effective Altruism Forum.
    In this series, we evaluate AI safety organizations that have received more than $10 million per year in funding. We do not critique MIRI and OpenAI as there have been several conversations and critiques of these organizations (1,2,3).
    The authors of this post include two technical AI safety researchers, and others who have spent significant time in the Bay Area community. One technical AI safety researcher is senior (>4 years experience), the other junior. We would like to make our critiques non-anonymously but unfortunately believe this would be professionally unwise. Further, we believe our criticisms stand on their own. Though we have done our best to remain impartial, readers should not assume that we are completely unbiased or don’t have anything to personally or professionally gain from publishing these critiques. We take the benefits and drawbacks of the anonymous nature of our post seriously, and are open to feedback on anything we might have done better.
    The first post in this series will cover Redwood Research (Redwood). Redwood is a non-profit started in 2021 working on technical AI safety (TAIS) alignment research. Their approach is heavily informed by the work of Paul Christiano, who runs the Alignment Research Center (ARC), and previously ran the language model alignment team at OpenAI. Paul originally proposed one of Redwood's original projects and is on Redwood’s board. Redwood has strong connections with central EA leadership and funders, has received significant funding since its inception, recruits almost exclusively from the EA movement, and partly acts as a gatekeeper to central EA institutions.
    We shared a draft of this document with Redwood prior to publication and are grateful for their feedback and corrections (we recommend others also reach out similarly). We’ve also invited them to share their views in the comments of this post.
    We would like to also invite others to share their thoughts in the comments openly if you feel comfortable, or contribute anonymously via this form. We will add inputs from there to the comments section of this post, but will likely not be updating the main body of the post as a result (unless comments catch errors in our writing).
    Summary of our views
    We believe that Redwood has some serious flaws as an org, yet has received a significant amount of funding from a central EA grantmaker (Open Philanthropy). Inadequately kept in check conflicts of interest (COIs) might be partly responsible for funders giving a relatively immature org lots of money and causing some negative effects on the field and EA community. We will share our critiques of Constellation (and Open Philanthropy) in a follow-up post. We also have some suggestions for Redwood that we believe might help them achieve their goals.
    Redwood is a young organization that has room to improve. While there may be flaws in their current approach, it is possible for them to learn and adapt in order to produce more accurate and reliable results in the future. Many successful organizations made significant pivots while at a similar scale to Redwood, and we remain cautiously optimistic about Redwood's future potential.
    An Overview of Redwood Research
    Grants: Redwood has received just over $21 million dollars in funding that we are aware of, for their own operations (2/3, or $14 million) and running Constellation (1/3 or $7 million) Redwood received $20 million from Open Philanthropy (OP) (grant 1 & 2) and $1.27 million from the Survival and Flourishing Fund. They also were granted (but never received) $6.6 million from FTX Future Fund.
    Output:
    Research: Redwood lists six research projects on their website: causal scrubbing, interpretability ...
    35 min
  • EA - Deference on AI timelines: survey results by Sam Clarke
    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: Deference on AI timelines: survey results, published by Sam Clarke on March 30, 2023 on The Effective Altruism Forum.
    Crossposted to LessWrong.
    In October 2022, 91 EA Forum/LessWrong users answered the AI timelines deference survey. This post summarises the results.
    Context
    The survey was advertised in this forum post, and anyone could respond. Respondents were asked to whom they defer most, second-most and third-most, on AI timelines. You can see the survey here.
    Results
    This spreadsheet has the raw anonymised survey results. Here are some plots which try to summarise them.
    Simply tallying up the number of times that each person is deferred to:
    The plot only features people who were deferred to by at least two respondents.
    Some basic observations:
    Overall, respondents defer most frequently to themselves—i.e. their “inside view” or independent impression—and Ajeya Cotra. These two responses were each at least twice as frequent as any other response.
    Then there’s a kind of “middle cluster”—featuring Daniel Kokotajlo, Paul Christiano, Eliezer Yudkowsky and Holden Karnofsky—where, again, each of these responses were ~at least twice as frequent as any other response.
    Then comes everyone else. There’s probably something more fine-grained to be said here, but it doesn’t seem crucial to understanding the overall picture.
    What happens if you redo the plot with a different metric? How sensitive are the results to that?
    One thing we tried was computing a “weighted” score for each person, by giving them:
    3 points for each respondent who defers to them the most
    2 points for each respondent who defers to them second-most
    1 point for each respondent who defers to them third-most.
    If you redo the plot with that score, you get this plot. The ordering changes a bit, but I don’t think it really changes the high-level picture. In particular, the basic observations in the previous section still hold.
    We think the weighted score (described in this section) and unweighted score (described in the previous section) are the two most natural metrics, so we didn’t try out any others.
    Don’t some people have highly correlated views? What happens if you cluster those together?
    Yeah, we do think some people have highly correlated views, in the sense that their views depend on similar assumptions or arguments. We tried plotting the results using the following basic clusters:
    Open Philanthropy cluster = {Ajeya Cotra, Holden Karnofsky, Paul Christiano, Bioanchors}
    MIRI cluster = {MIRI, Eliezer Yudkowsky}
    Daniel Kokotajlo gets his own cluster
    Inside view = deferring to yourself, i.e. your independent impression
    Everyone else = all responses not in one of the above categories
    Here’s what you get if you simply tally up the number of times each cluster is deferred to:
    This plot gives a breakdown of two of the clusters (there’s no additional information that isn’t contained in the above two plots, it just gives a different view).
    This is just one way of clustering the responses, which seemed reasonable to us. There are other clusters you could make.
    Limitations of the survey
    Selection effects. This probably isn’t a representative sample of forum users, let alone of people who engage in discourse about AI timelines, or make decisions influenced by AI timelines.
    The survey didn’t elicit much detail about the weight that respondents gave to different views. We simply asked who respondents deferred most, second-most and third-most to. This misses a lot of information.
    The boundary between [deferring] and [having an independent impression] is vague. Consider: how much effort do you need to spend examining some assumption/argument for yourself, before considering it an independent impression, rather than deference? This is a limitation of the survey, because different respondents may have been using different b...
    5 min
  • EA - GWWC's 2020–2022 Impact evaluation (executive summary) by Michael Townsend
    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: GWWC's 2020–2022 Impact evaluation (executive summary), published by Michael Townsend on March 31, 2023 on The Effective Altruism Forum.
    Giving What We Can (GWWC) is on a mission to create a world in which giving effectively and significantly is a cultural norm. Our research recommendations and donation platform help people find and donate to effective charities, and our community — in particular, our pledgers — help foster a culture that inspires others to give.
    In this impact evaluation, we examine GWWC's cost-effectiveness from 2020 to 2022 in terms of how much money is directed to highly effective charities due to our work.
    We have several reasons for doing this:
    To provide potential donors with information about our past cost-effectiveness.
    To hold ourselves accountable and ensure that our activities are providing enough value to others.
    To determine which of our activities are most successful, so we can make more informed strategic decisions about where we should focus our efforts.
    To provide an example impact evaluation framework which other effective giving organisations can draw from for their own evaluations.
    This evaluation reflects two months of work by the GWWC research team, including conducting multiple surveys and analysing the data in our existing database. There are several limitations to our approach — some of which we discuss below. We did not aim for a comprehensive or “academically” correct answer to the question of “What is Giving What We Can’s impact?” Rather, in our analyses we are aiming for usefulness, justifiability, and transparency: we aim to practise what we preach and for this evaluation to meet the same standards of cost-effectiveness as we have for our other activities.
    Below, we share our key results, some guidance and caveats on how to interpret them, and our own takeaways from this evaluation. GWWC has historically derived a lot of value from our community’s feedback and input, so we invite readers to share any comments or takeaways they may have on the basis of reviewing this evaluation and its results, either by directly commenting or by reaching out to [email protected].
    Key results
    Our primary goal was to identify our overall cost-effectiveness as a giving multiplier — the ratio of our net benefits (additional money directed to highly effective charities, accounting for the opportunity costs of GWWC staff) compared to our operating costs.
    We estimate our giving multiplier for 2020–2022 is 30x, and that we counterfactually generated $62 million of value for highly effective charities.
    We were also particularly interested in the average lifetime value that GWWC contributes per pledge, as this can inform our future priorities.
    We estimate we counterfactually generate $22,000 of value for highly effective charities per GWWC Pledge, and $2,000 per Trial Pledge.
    We used these estimates to help inform our answer to the following question: In 2020–2022, did we generate more value through our pledges or through our non-pledge work?
    We estimate that pledgers donated $26 million in 2020–2022 because of GWWC. We also estimate GWWC will have caused $83 million of value from the new pledges taken in 2020–2022.
    We estimate GWWC caused $19 million in donations to highly effective charities from non-pledge donors in 2020–2022.
    These key results are arrived at through dozens of constituent estimates, many of which are independently interesting and inform our takeaways below. We also provide alternative conservative estimates for each of our best-guess estimates.
    How to interpret our results
    This section provides several high-level caveats to help readers better understand what the results of our impact evaluation do and don’t communicate about our impact.
    We generally looked at average rather than marginal cost-effectiveness
    Most of our ...
    15 min
  • LW - "Dangers of AI and the End of Human Civilization" Yudkowsky on Lex Fridman by DragonGod
    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: "Dangers of AI and the End of Human Civilization" Yudkowsky on Lex Fridman, published by DragonGod on March 30, 2023 on LessWrong.
    It's a 3 hours 23 minutes episode.
    [I might update this post with a summary once I'm done listening to it.]
    Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org.
    1 min
  • LW - The 0.2 OOMs/year target by Cleo Nardo
    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 0.2 OOMs/year target, published by Cleo Nardo on March 30, 2023 on LessWrong.
    Paris Climate Accords
    In the early 21st century, the climate movement converged around a "2°C target", shown in Article 2(1)(a) of the Paris Climate Accords:
    The 2°C target helps facilitate coordination between nations, organisations, and individuals.
    It provided a clear, measurable goal.
    It provided a sense of urgency and severity.
    It promoted a sense of shared responsibility.
    It helped to align efforts across different stakeholders.
    It created a shared understanding of what success would look like.
    The AI governance community should converge around a similar target.
    0.2 OOMs/year target
    In this article, I propose a target of 0.2 OOMs/year. OOM stands for "orders of magnitude", and corresponds to a ten-fold increase. This corresponds to a 58% year-on-year growth.
    I do not propose any specific policy for achieving the 0.2 OOMs/year target, because the purpose of the target is to unify stakeholders even if they support different policies.
    I do not propose any specific justification for the 0.2 OOMs/year target, because the purpose of the target is to unify stakeholders even if they have different justifications.
    Here is the statement:
    "Humanity — which includes all nations, organisations, and individuals — should limit the growth rate of machine learning training runs from 2020 until 2050 to below 0.2 OOMs/year."
    The statement is intentionally ambiguous about how to measure "the growth rate of machine learning training runs". I suspect that a good proxy metric would be the effective training footprint (defined below) but I don't think the proxy metric should be included in the statement of the target itself.
    Effective training footprint
    What is the effective training footprint?
    The effective training footprint, measured in FLOPs, is one proxy metric for the growth rate of machine learning training runs. The footprint of a model is defined, with caveats, as the total number of FLOPs used to train the model since initialisation.
    Caveats:
    A randomly initialised model has a footprint of 0 FLOPs.
    If the model is trained from a randomly initialised model using SGD or a variant, then its footprint is the total number of FLOPs used in the training process.
    If a pre-trained base model is used for the initialisation of another training process (such as unsupervised learning, supervised learning, fine-tuning, or reinforcement learning), then the footprint of the resulting model will include the footprint of the pre-trained model.
    If multiple models are composed to form a single cohesive model, then the footprint of the resulting model is the sum of the footprints of each component model.
    If there is a major algorithmic innovation which divides by a factor of r the FLOPs required to train a model to a particular score on downstream tasks, then the footprint of models trained with that innovation is multiplied by the same factor r.
    This list of caveats to the definition of Effective Training Footprint is non-exhaustive. Future consultations may yield additional caveats, or replace Effective Training Footprint with an entirely different proxy metric.
    Fixing the y-axis
    According to the 0.2 OOMs/year target, there cannot exist an ML model during the year (2022+x) with a footprint exceeding f(x), where f(x+1)=100.2×f(x). That means that log10f(x)=(0.2x+a) FLOPs for some fixed constant a.
    If we consult EpochAI's plot of compute training runs during the large-scale era of ML, we see that footprints have been growing with approximately 0.5 OOMs/year.
    We can use this trend to fix the value of A. In 2022, the footprint was approximately 1.0e+24. Therefore a=24.
    In other words, log10f(x)=0.2x+24.
    I have used 2022 as an anchor. If I had used 2016 instead, then the 0.2 OOMs/yr target would've been stricter. If I...
    8 min
  • EA - What's surprised me as an entry-level generalist at Open Phil & my recommendations to early career professionals by Sam Anschell
    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: What's surprised me as an entry-level generalist at Open Phil & my recommendations to early career professionals, published by Sam Anschell on March 30, 2023 on The Effective Altruism Forum.
    Disclaimer The opinions expressed in this post are my own and do not represent Open Philanthropy.
    Valentine’s day 2022 was my first day of work at Open Phil. As a 24-year-old who had spent two post-grad years as a poker dealer/cardroom union representative, I had little in the way of white collar context or transferable skills. Recently, a few undergraduates and early career professionals have reached out to learn what the job is like & how they can get further involved in EA. In this post I’ll try to provide the advice that I would have benefited from hearing a couple years ago.
    I’m hoping to widen the aperture of possibilities to early career professionals who are excited to use their time and talents to do good. I know how difficult it can be to land an EA job - it took years of on-and-off applying before I got an offer. It’s normal to face a string of rejections and it’s valid to feel frustrated by that, but I think the benefits to individuals and organizations when a hire is made are so great that continuing to apply is worth it. I encourage anyone who is struggling to get their foot in the door to read Aaron’s Epistemic Stories - I found it really motivating.
    TLDR:
    Before starting this job, I underestimated the $ value of person-hours at EA orgs. I may have done this because:
    There’s a disconnect between salary and social value generated (even though salaries at EA orgs are generous). Most for-profit companies value their average staff member’s contributions at about 2x their salary, and I suspect EA orgs value their average staff member’s contributions at more like 8x+ their salary.
    It could be uncomfortable to think that time at an EA org would be very valuable, both because of what it would imply for labor/leisure tradeoffs and because it could lead to imposter syndrome.
    It can be easy to mentally compartmentalize work at EA orgs as creating a similar level of social impact to work at nonprofits in general, despite believing that EA interventions are much more cost-effective than the average nonprofit’s interventions.
    Due to this underestimate, I now think I should have focused on working directly on EA projects and spending more time applying for EA jobs earlier. Here are some of my recommendations to early career professionals:
    Don't feel like you have to put multiple years into a job before leaving to show you’re not a job-hopper. EA orgs understand the desire to contribute to work you find meaningful as soon as you can!
    I suspect people apply to too few jobs given how unpleasant it can be to job hunt, and I strongly encourage you to keep putting yourself out there.
    I applied to a few hundred jobs before landing this one, as did many of my friends who work at EA orgs. Not getting any jobs despite many applications isn't a sign that you're a bad applicant.
    Doing even unsexy work for an organization that you’re strongly mission-aligned with is more motivating than you might expect.
    I write about impactful ways that anyone can spend time at the end of this post.
    A ~Year in the Life
    What I did at work
    It’s hard to look at a job description and get a sense of what the day-to-day looks like (and see whether one might be qualified for the job). Success in my and many other entry-level jobs seems to be a product of enthusiasm, dependability (which I’d define as the independence/organization skills to manage a task so that the person who assigns it doesn’t need to follow up), and good judgment (when to check in, what tone to use in emails, etc.) In my day-to-day as a business operations generalist (assistant level), I've:
    Helped manage the physical office space:
    Purchased, ren...
    20 min