The Nonlinear Library

The Nonlinear Library

By The Nonlinear FundEducation
Download on the App Store

All The Nonlinear Library episodes

  • LW - How Risky Is Trick-or-Treating? 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: How Risky Is Trick-or-Treating?, published by jefftk on October 27, 2022 on LessWrong.
    content warning: discussion of child fatalities
    I've seen this chart going around:
    Sperling and Smith created it from NHTSA
    FARS data, and there have been lots of news stories about it or the underlying data. Some people have a "lets make costumes brighter" angle, some see this as ammunition in the fight against car-centric culture (which I support), some people want to move Halloween earlier in the day. But I think everyone is missing the point: Halloween is actually one of the safest days to be out walking as a kid.
    Let's look at the data. I started by replicating the chart to verify
    I was doing it correctly (code):
    The dip on February 29th is mostly just that this day only occurs once every four years, so it has about a quarter the deaths you'd expect. But what about the spike on 4/30? On April 30th 1992 someone drove their car into a large group of 3rd graders on a field trip, and one of them died. It turns out the FARS records can include cases where someone else died in the same accident. Filtering to just the records representing deaths I get:
    The data available actually goes back to 1982 and now runs to 2020, so we can expand it a bit:
    We can also break down by age, since "child" can mean a range of things and most people are likely thinking mostly about younger children:
    All this is to confirm the original chart: more children do die on Halloween than on a typical day. But I still don't think that's a reason to keep your kids home on Halloween, even when you set aside the consideration that Halloween is an especially fun time to be a child pedestrian.
    The issue is, there are a lot more kids out on Halloween than the typical day. If you learned that more people died in car crashes in the US (42k/y) than in Paraguay (1.6k/y) you wouldn't conclude that
    Paraguay's roads are safer: the US has 50x more people than Paraguay! We would ideally scale the initial chart by the number of children out walking on any given day, but I don't think anyone has numbers on this. Very roughly, though, maybe it's something like 5x as many people (kids, parents, grandparents) out on foot on Halloween than the typical day?
    Every pedestrian is rolling some dice. The dice are pretty favorable: of the ~70M people under age 18 in the US only 320 died in 2020 by being hit by a car. But you still might want to avoid days where the dice were less favorable than usual. Looking at the original chart you might conclude that Halloween was one of those days, but if there are a lot more kids walking around on Halloween it's actually one of the safer days.
    Comment via: facebook
    Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org.
    3 min
  • EA - New tool for exploring EA Forum and LessWrong - Tree of Tags by Filip Sondej
    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: New tool for exploring EA Forum and LessWrong - Tree of Tags, published by Filip Sondej on October 27, 2022 on The Effective Altruism Forum.
    Explore the hidden treasures of the forums
    With this tool you can zoom in on your favorite topics from EA Forum, LessWrong, and Alignment Forum. Or you can just wander around and see what you find.
    You start by seeing the whole forum split into two main topics. Choose the one that you like more, and that topic will be split again into two subtopics. Choose a subtopic, and again, you will see it split into two subsubtopics (and so on).
    It's like climbing a tree, where you start at the trunk and then choose which branches to go into as you go higher.
    Choose a forum to climb:
    EA Forum
    LessWrong
    Alignment Forum
    Some tips:
    it's easier to choose a branch by looking at tags rather than posts
    on mobile, horizontal view is more practical
    Features
    Each topic has a unique and unchanging URL. So if you find a place you like, just bookmark it! The posts inside will be updated, but the theme stays the same.
    The bar at the right of each post is the reading time indicator. Full bar means 30 min, half bar means 15 min, and so on.
    At the top right, you can choose how to rank the posts:
    hot - new and upvoted - the default forum ranking
    top - most upvotes
    alive - has recent comments
    meritocratic - votes from high karma users are exaggerated
    regular - default scoring
    democratic - posts are scored as if everyone has the same voting power
    My hopes
    The amount of content is overwhelming. My problem is not that there's nothing good to read, but that there is so much to plow through to find what's best for me.
    Also, it's sad to see great posts receive so little attention, the moment after they are pushed off the frontpage. But they are still valuable, and for each forgotten post, there is someone who should read it. So I want to make the right topics find their way to the right people.
    Let's make the forums evergreen!
    Tag similarity
    For each tag, you can also see what tags are most related to it:Explore EA Forum tagsExplore LessWrong tagsExplore Alignment Forum tags
    Tip: you can go directly to some tag, by finding it's URL on the forum, and modifying this site's URL.For example, to go to "Consciousness research", you find go to
    EA Forum:
    LessWrong:
    Alignment Forum:
    How does it work?
    Here I explain all the details.
    Cross-posted to LessWrong.
    Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org.
    3 min
  • EA - Summary of "Technology Favours Tyranny" by Yuval Noah Harari by Madhav Malhotra
    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: Summary of "Technology Favours Tyranny" by Yuval Noah Harari, published by Madhav Malhotra on October 26, 2022 on The Effective Altruism Forum.
    Link: Technology Favours Tyrrany – Yuval Noah Harari
    Summary:
    Economic incentives favour making powerful AI.
    Making powerful AI is easier where you don’t respect liberty / privacy.
    We’re developing a new useless class. People that can’t do anything better than technology. And we don’t know what to do with them.
    Raw Notes:
    Democracy has been the exception, not the norm throughout history.
    He thinks that there are certain technological conditions that promote democracy. And that we're growing further away from those conditions with the technology we're developing.
    I'm not entirely sure what he means. But he gives the example of the "ordinary man" being praised in the 1940s. Everyday steel, coal, farm, military workers would be on propaganda posters in the US, USSR, and Europe. But today, "ordinary workers" are no longer celebrated. They no longer have a future.
    Says ordinary people have political power (in democracies) but are less and less relevant economically. "Perhaps in the 21st century, populist revolts will be staged not against an economic elite that exploits people but against an economic elite that does not need them anymore."
    We can't keep people happy by just getting more and more economic growth. Not if this growth is based on inventing more and more disruptive technologies.
    As technology continues to develop, we'll have new jobs emerge quickly. But the new jobs will also disappear quickly. Ex: Human annotators for AI models. People will need to retrain many many times.
    How will this affect people's stress given how many people are already stressed?
    He thinks that we'll have technology that can tell what people are thinking (roughly). Ex: He says if you look at a picture and get higher blood pressure and activity in your amygdala, then you're getting angry from the contents of that picture. And he says this type of technology could be used in autocratic states to suppress dissent.
    He also says that we already have technology to manipulate our emotions better than our own families can manipulate our emotions. He says that this might be used to convince people to buy into X product, Y politician, Z movement, etc.
    He says that in past times, democracies were better able to use information than autocracies. Because information was available to everyone and more minds could process it. But with AI, the group with the most powerful AI is the best at processing information.
    Ex: If an autocratic government orders all its citizens to take health tests and share their medical data to the government, that government would be much better at biomedical research than governments that respect individual privacy.
    He uses this to point out how the centralised way that autocracies make decisions lend themselves to autocracies being the best at processing information.
    He mentions how we've already started to erode skills based on technological reliance. Ex: Many more people don't know how to get informationw ithout search engines. Ex: Many more people don't know how to navigate long distance without Google Maps.
    For next steps, he suggests that to control the development of technology like AI, we need to control the 'means of production.' In agrarian societies, the means of production was land, so we created property rights. In the 1900s, the means of production were factories/machines, so we protected those most in wartime. In the 2000s, the means of production might be computer chips, data, etc. We currently suck at controlling/regulating this.
    Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org.
    4 min
  • EA - Podcast: The Left and Effective Altruism with Habiba Islam by Garrison
    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: Podcast: The Left and Effective Altruism with Habiba Islam, published by Garrison on October 27, 2022 on The Effective Altruism Forum.
    I recently rebooted my interview podcast, The Most Interesting People I Know (found wherever you find podcasts). I focus on EA and left-wing guests, and have been pretty involved in both communities for the last 5 years. Some example guests: Rutger Bregman, Leah Garcés, Lewis Bollard, Spencer Greenberg, Nathan Robinson, Malaika Jabali, Emily Bazelon, David Shor, and Eric Levitz.
    I just released a long conversation with Habiba Islam, an 80K career advisor and lefty, about the relationship between EA and the left.
    This is not an attempt to paper over differences between the two communities, or pretend that EA is more left-wing than it is. Instead, I tried to give an accurate description of both communities, where they are in hidden agreement, where they actually disagree, and what each can learn from the other.
    Habiba is so sharp and thoughtful throughout the conversation. We're very lucky to have her!
    I hope this could be a good reference text as well as an onboarding ramp for leftists who might be open to EA.
    I think there's a real gap in the EA media-verse on the intersection of left-wing politics and EA, and we're almost certainly missing out on some great people and perspectives who would be into EA if they were presented with the right arguments and framing.
    I have no delusions that all leftists would be into EA if they only understood it better, but I think there are tons of bad-faith criticisms and genuine misunderstandings that we could better address. I think we can have a healthier and more productive relationship with the left.
    If you'd like to support the show, here are some things you can do:
    Personally recommend the show/particular episodes to friends. Apparently, this is how podcasts best grow their audiences.
    Share the podcast/episode on social media (I'm on Twitter @garrisonlovely)
    Rate and review the show on Apple Podcasts.
    Give me feedback (anonymous form here). You can also email me at [email protected]
    Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org.
    3 min
  • EA - GiveWell should use shorter TAI timelines by Oscar Delaney
    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: GiveWell should use shorter TAI timelines, published by Oscar Delaney on October 27, 2022 on The Effective Altruism Forum.
    Summary
    GiveWell’s discount rate of 4% includes a 1.4% contribution from ‘temporal uncertainty’ arising from the possibility of major events radically changing the world.
    This is incompatible with the transformative artificial intelligence (TAI) timelines of many AI safety researchers.
    I argue that GiveWell should increase its discount rate, or at least provide a justification for differing significantly from the commonly-held (in EA) view that TAI could come soon.
    Epistemic Status: timelines are hard, and I don’t have novel takes, but I worry perhaps GiveWell doesn’t either, and they are dissenting unintentionally.
    In my accompanying post I argued GiveWell should use a probability distribution over discount rates, I will ignore that here though, and just consider whether their point estimate is appropriate.
    GiveWell’s current discount rate of 4% is calculated as the sum of three factors. Quoting their explanations from this document:
    Improving circumstances over time: 1.7%. “Increases in consumption over time meaning marginal increases in consumption in the future are less valuable.”
    Compounding non-monetary benefits: 0.9%. “There are non-monetary returns not captured in our cost-effectiveness analysis which likely compound over time and are causally intertwined with consumption. These include reduced stress and improved nutrition.”
    Temporal uncertainty: 1.4%. “Uncertainty increases with projections into the future, meaning the projected benefits may fail to materialize. James [a GiveWell researcher] recommended a rate of 1.4% based on judgement on the annual likelihood of an unforeseen event or longer term change causing the expected benefits to not be realized. Examples of such events are major changes in economic structure, catastrophe, or political instability.”
    I do not have a good understanding of how these numbers were derived, and have no reason to think the first two are unfair estimates. I think the third is a significant underestimate.
    TAI is precisely the sort of “major change” meant to be captured by the temporal uncertainty factor. I have no insights to add on the question of TAI timelines, but I think absent GiveWell providing justification to the contrary, they should default towards using the timelines of people who have thought about this a lot. One such person is Ajeya Cotra, who in August reported a 50% credence in TAI being developed by 2040. I do not claim, and nor does Ajeya, that this is authoritative, however it seems a reasonable starting point for GiveWell to use, given they have not and (I think rightly) probably will not put significant work into forming independent timelines. Also in August, a broader survey of 738 experts by AI Impacts resulted in a median year for TAI of 2059. This source has the advantage of including many more people, but conversely most of them will have not spent much time thinking carefully about timelines.
    I will not give a sophisticated instantiation of what I propose, but rather gesture at what I think a good approach would be, and give a toy example to improve on the status quo. A naive thing to do would be to imagine that there is a fixed annual probability of developing TAI conditional on not having developed it to date. This method gives an annual probability of 3.8% under Ajeya’s timelines, and 1.9% under the AI Impacts timelines.[1] In reality, more of our probability mass should be placed on the later years between now and 2040 (or 2059), and we should not simply stop at 2040 (or 2059). A proper model would likely need to dispense with a constant discount rate entirely, and instead track the probability the world has not seen a “major change” by each year. A model that accomplishes ...
    5 min
  • EA - Recommend Me EAs To Write About by Stephen Thomas
    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: Recommend Me EAs To Write About, published by Stephen Thomas on October 26, 2022 on The Effective Altruism Forum.
    Project basics
    I’ve been funded by the Future Fund Regranting Program to start a series of magazine-style profiles of people doing interesting EA work, called “Humans of Effective Altruism.”
    [Recommend someone here]
    The idea is to write these for an audience of both EAs and non-EAs, with the idea of giving people tangible examples of interesting and effective career paths and/or life paths.
    I’d like to get into the nitty-gritty of what people do day to day, and also dig into who they are as people, ‘what makes them tick’.
    I’ll be publishing them on this Substack. I anticipate most interviews will be done remotely, but if someone is in NYC, where I am, I’d somewhat prefer an in-person meeting, to give the profile color.
    I’m now looking for recommendations for who to write about.
    You can use this Google form, comment on this post, DM me, or email me at [email protected].
    Please err on the side of suggesting anyone you think would make an interesting profile subject! It’s better for me to have more people to consider, and probably at least some people suggested/recommended won’t want to be profiled, so I’d love to have a very long list of options.
    That said, I basically have 3 criteria:
    3 criteria
    They are doing high-impact activities (probably their job, but not necessarily). One motivation for this project is to give people considering life-options tangible examples of net-positive things to do.
    MOST IMPORTANTLY, they have an especially interesting personal story, especially if it relates to how they think about their work, or they’re just an especially interesting person.
    And, obviously, they must be open to being profiled, and are willing to get at least a little personal.
    I’m also open to the idea of a profile of more than one person (a collective, a charity, a purpose-driven group house), but I don’t expect these to be the majority of the profiles.
    Self-recommendations A-okay.
    Thanks in advance.
    [Recommend someone here]
    Further details about process
    I’d prefer to have a preliminary chat without committing to writing up a full profile
    Different profiles may take on different shapes. Some may be long, some short. Some may involve multiple interviews, some just one. Because of this, it’s possible (though maybe somewhat unlikely) that a ‘preliminary chat’ could be the entire process, which get processed into a short profile.
    Further details about how I'm thinking about the project
    Basically, prioritizing an interesting story is a bid to attract the attention of a wider, non-EA audience
    Slight preference for profile subjects for whom there’s something the reader could do when they finish reading, a specific action they could take if they were inspired by this person, to help their specific cause, over and above ‘get [more] into EA’
    Preference for someone who represents an idea, and/or whose story illustrates a broader point
    Recommend someone
    Recommend someone here.
    Subscribe / Follow
    Subscribe to the Substack here to get the first profile in your inbox.
    Follow on Twitter, Instagram, and Facebook
    Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org.
    4 min
  • EA - GiveWell Misuses Discount Rates by Oscar Delaney
    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: GiveWell Misuses Discount Rates, published by Oscar Delaney on October 27, 2022 on The Effective Altruism Forum.
    Summary
    GiveWell currently uses a time discount rate of 4% for all their cost-effectiveness analyses (CEAs).
    I argue that it is a mathematical mistake to pick any single best guess value to use for the CEAs.
    Instead, GiveWell should use a probability distribution over possible discount rates.
    This is not just an aesthetic judgement for mathematical puritans; it materially changes the CEAs, notably by making all the deworming interventions more attractive relative to other interventions.
    This is because deworming interventions rely on multi-decadal effects, and so a lower discount rate would make them much more valuable.
    Epistemic Status
    On the object level, I cannot think of any reasons to justify GiveWell's current modelling choice over my proposal.
    However, I still doubt my conclusion because on the meta level it seems like an obvious thing that would be surprising if no one at GiveWell had ever thought of doing, which is evidence I am missing something important.
    Main
    GiveWell’s CEAs are an impressive attempt to model many different factors in assessing the near-term impacts of various interventions.[1] I will ignore all of this complexity. For my purposes, it is sufficient to note that the CEA for most interventions is well characterised by decomposing impact into several constituents, and multiplying these numbers together. Consider Helen Keller International’s Vitamin A Supplementation program:
    V=M×R×1C [2] where:
    V is cost-effectiveness [deaths/dollar],
    M is baseline mortality [deaths/year/child],
    R is mortality reduction [%], and
    C is treatment cost [dollars/child/year]
    Obviously, all of these terms are uncertain. Treatment costs we can estimate quite accurately, but there may be fluctuations in the price of labour or materials needed in the distribution. Mortality data is generally good, but some deaths may not be reported, and mortality rates will change over time. The mortality reduction is based on a solid-seeming meta-analysis of RCTs, but things change over time, and circumstances differ between the trial and intervention locations.
    GiveWell’s model makes a subtle mathematical assumption, namely that the expectation of the product of these three random variables is equal to the product of their expectations:
    E[V]=E[M×R×1C]=E[M]×E[R]×E[1C]
    This is not, in general, true.[3] However, if the three random variables are independent, it is true. I cannot think of any plausible ways in which these three random variables correlate. Surely learning that the price of vitamin A tablets just doubled (C) does not affect how effective they are (R) or change the baseline of how many kids die (M). Thus, while GiveWell’s method is mathematically unsound, it gives the correct answer in this case. It could well be that GiveWell has considered this, and decided not to explain this in their CEAs because it doesn’t change the answer. I think this would be a mistake in communication, but otherwise benign.
    The one place where I believe this mathematical mistake translates into an incorrect answer is in the use of discount rates. From GiveWell’s explanatory document:
    “The discount rate's primary effect in the cost-effectiveness analyses of our top charities is to represent how much we discount increases in consumption resulting from the long run effects of improved child health for our malaria, deworming and vitamin A charities (which we call "developmental effects"). It also affects the longer-run benefits from cash transfers. We don't discount mortality benefits in our cost-effectiveness analyses.”
    This figure shows the cost-effectiveness of all the charities in the CEA spreadsheet, when varying the discount rate.[4]
    Deworming interventions, shown in dashed lines, v...
    7 min
  • EA - Apply to the Redwood Research Mechanistic Interpretability Experiment (REMIX), a research program in Berkeley by Max Nadeau
    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: Apply to the Redwood Research Mechanistic Interpretability Experiment (REMIX), a research program in Berkeley, published by Max Nadeau on October 27, 2022 on The Effective Altruism Forum.
    This winter, Redwood Research is running a coordinated research effort on mechanistic interpretability of transformer models. We’re excited about recent advances in mechanistic interpretability and now want to try to scale our interpretability methodology to a larger group doing research in parallel.
    REMIX participants will work to provide mechanistic explanations of model behaviors, using our causal scrubbing methodology (forthcoming within a week) to formalize and evaluate interpretability hypotheses. We hope to produce many more explanations of model behaviors akin to our recent work investigating behaviors of GPT-2-small, toy language models, and models trained on algorithmic tasks (also forthcoming). We think this work is a particularly promising research direction for mitigating existential risks from advanced AI systems (more in Goals and FAQ).
    Apply here by November 8th to be a researcher in the program. Apply sooner if you’d like to start early (details below) or receive an earlier response.
    Some key details:
    We expect to accept 30-50 participants.
    We plan to have some researchers arrive early, with some people starting as soon as possible. The majority of researchers will likely participate during the months of December and/or January.
    We expect researchers to participate for a month minimum, and (all else equal) will prefer applicants who are able to come for longer. We’ll pay for housing and travel, and also pay researchers for their time. We’ll clarify the payment structure prior to asking people to commit to the program.
    We’re interested in some participants acting as team leaders who would help on-board and provide research advice to other participants. This would involve arriving early to get experience with our tools and research directions and participating for a longer period (~2 months). You can indicate interest in this role in the application.
    We’re excited about applicants with a range of backgrounds; we’re not expecting applicants to have prior experience in interpretability research. Applicants should be comfortable working with Python, PyTorch/TensorFlow/Numpy (we’ll be using PyTorch), and linear algebra. We’re particularly excited about applicants with experience doing empirical science in any field.
    We’ll allocate the first week to practice using our interpretability tools and methodology; the rest will be researching in small groups. See Schedule.
    Feel free to email [email protected] with questions.
    Goals
    Research output. We hope this program will produce research that is useful in multiple ways:
    We’d like stronger and more grounded characterizations of how language models perform a certain class of behaviors. For example, we currently have a variety of findings about how GPT-2-small implements indirect object identification (“IOI”, see next section for more explanation), but aren’t yet sure how often they apply to other models or other tasks. We’d know a lot more if we had a larger quantity of this research.
    For each behavior investigated, we think there’s some chance of stumbling across something really interesting. Examples of this include induction heads and the “pointer manipulation” result in the IOI paper: not only does the model copy information between attention streams, but it also copies “pointers”, i.e. the position of the residual stream that contains the relevant information.
    We’re interested in learning whether different language models implement the same behaviors in similar ways.
    We’d like a better sense of how good the current library of interpretability techniques is, and we’d like to get ideas for new techniques.
    We’d like to have mo...
    20 min
  • AF - Apply to the Redwood Research Mechanistic Interpretability Experiment (REMIX), a research program in Berkeley by maxnadeau
    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: Apply to the Redwood Research Mechanistic Interpretability Experiment (REMIX), a research program in Berkeley, published by maxnadeau on October 27, 2022 on The AI Alignment Forum.
    This winter, Redwood Research is running a coordinated research effort on mechanistic interpretability of transformer models. We’re excited about recent advances in mechanistic interpretability and now want to try to scale our interpretability methodology to a larger group doing research in parallel.
    REMIX participants will work to provide mechanistic explanations of model behaviors, using our causal scrubbing methodology (forthcoming within a week) to formalize and evaluate interpretability hypotheses. We hope to produce many more explanations of model behaviors akin to our recent work investigating behaviors of GPT-2-small, toy language models, and models trained on algorithmic tasks (also forthcoming). We think this work is a particularly promising research direction for mitigating existential risks from advanced AI systems (more in Goals and FAQ).
    Apply here by November 8th to be a researcher in the program. Apply sooner if you’d like to start early (details below) or receive an earlier response.
    Some key details:
    We expect to accept 30-50 participants.
    We plan to have some researchers arrive early, with some people starting as soon as possible. The majority of researchers will likely participate during the months of December and/or January.
    We expect researchers to participate for a month minimum, and (all else equal) will prefer applicants who are able to come for longer. We’ll pay for housing and travel, and also pay researchers for their time. We’ll clarify the payment structure prior to asking people to commit to the program.
    We’re interested in some participants acting as team leaders who would help on-board and provide research advice to other participants. This would involve arriving early to get experience with our tools and research directions and participating for a longer period (~2 months). You can indicate interest in this role in the application.
    We’re excited about applicants with a range of backgrounds; we’re not expecting applicants to have prior experience in interpretability research. Applicants should be comfortable working with Python, PyTorch/TensorFlow/Numpy (we’ll be using PyTorch), and linear algebra. We’re particularly excited about applicants with experience doing empirical science in any field.
    We’ll allocate the first week to practice using our interpretability tools and methodology; the rest will be researching in small groups. See Schedule.
    Feel free to email [email protected] with questions.
    Goals
    Research output. We hope this program will produce research that is useful in multiple ways:
    We’d like stronger and more grounded characterizations of how language models perform a certain class of behaviors. For example, we currently have a variety of findings about how GPT-2-small implements indirect object identification (“IOI”, see next section for more explanation), but aren’t yet sure how often they apply to other models or other tasks. We’d know a lot more if we had a larger quantity of this research.
    For each behavior investigated, we think there’s some chance of stumbling across something really interesting. Examples of this include induction heads and the “pointer manipulation” result in the IOI paper: not only does the model copy information between attention streams, but it also copies “pointers”, i.e. the position of the residual stream that contains the relevant information.
    We’re interested in learning whether different language models implement the same behaviors in similar ways.
    We’d like a better sense of how good the current library of interpretability techniques is, and we’d like to get ideas for new techniques.
    We’d like to have more exam...
    20 min
  • LW - Trigger-based rapid checklists by VipulNaik
    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: Trigger-based rapid checklists, published by VipulNaik on October 26, 2022 on LessWrong.
    In this post, I describe an approach I've been following for the past few months, and that I feel has been moderately helpful to me: trigger-based rapid checklists. A trigger-based rapid checklist is a memorized checklist that I execute rapidly in response to a specific trigger. I'll explain with examples, and then talk of some of the general features.
    Although this wasn't directly inspired by any specific life hack that others use, it has some overlap with trigger-action planning as well as with the ideas popularized by Getting Things
    Done.
    This post should not be taken as a documentation of the exact details of what I do -- I've left out some details and made some tweaks for simplicity. The information here should be thought of as mostly illustrative of how somebody might use trigger-based rapid checklists.
    Examples of trigger-based rapid checklists
    Some triggers
    I currently have several trigger-based rapid checklists. Here is a list (not necessarily comprehensive):
    Morning checklist: I execute this after I start using my laptop in the morning.
    Night checklist: I execute this right before I plan to close my laptop in the evening (this is not right before I go to sleep).
    Pre-outside trip checklist: I execute this right before leaving my home for an extended period of time (e.g., for shopping).
    Post-outside trip checklist: I execute this right after returning from an outside trip.
    Context switch checklist: I execute this when I'm switching context between different kinds of work.
    Sample checklists
    Here is a lightly-edited version of my morning checklist:
    [ ] (W) Check the weather for the day
    [ ] (E) Catch up on email: personal and work
    Also check spam, trash, and all mail (I won't check these for each intermediate check during the day) - [ ] (M) Catch up on Facebook messages
    [ ] (S) Catch up on Slack: personal and work
    [ ] (C) Check calendar to plan my day
    [ ] (A) Turn off alarm
    [ ] (L) Run the loose ends script
    And here's a lightly-edited version of my night checklist:
    [ ] (W) Check the weather for the next 1-2 days
    [ ] (E) Catch up on email: personal and work
    [ ] (M) Catch up on Facebook messages
    [ ] (S) Catch up on Slack: personal and work
    [ ] (C) Check calendar for the next 3 days: - For next day, pay particular attention to implications regarding e.g., when to wake up; in particular, try to make it so that I go to sleep at least 9 hours before the time of the first appointment
    For the two days after that, make sure that the calendar makes sense, e.g., I didn't mark an incorrect event; if there are things that need to be in sync across multiple calendars (e.g., across two jobs while I'm juggling them), make sure they are in sync
    [ ] (A) Turn alarm on for tomorrow
    The default alarm setting is 8:30 AM; this is fine for most days (I am usually awake well before this time -- the alarm is a fallback measure) - I should make sure that the alarm time is early enough to be at least 1 hour before any online meeting, and a larger amount of time for other kinds of events
    Also use this occasion to make sure that the phone is plugged in and (charging or charged) - [ ] (W) Make sure bedroom window and hall terrace door are open
    [ ] (L) Run the loose ends script
    [ ] (C) Close lid and cover laptop
    Important features of execution of the checklists
    Trigger-based nature
    These checklists are trigger-based rather than time-based. In general, I've found time-based things hard to adhere to, since there can be many sources of variation such as when I wake up or when I go to sleep. Trigger-based actions are easier to remember, because I just have to build in my mind the association from the trigger to the action. The LessWrong trigger-action planning tag has a bunch of great content explaining why tr...
    14 min