The Nonlinear Library

The Nonlinear Library

By The Nonlinear FundEducation
Download on the App Store

All The Nonlinear Library episodes

  • LW - Dependency Tree For The Development Of Plate Tectonics by Elizabeth
    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: Dependency Tree For The Development Of Plate Tectonics, published by Elizabeth on October 5, 2022 on LessWrong.
    This post is really rough and mostly meant to refer back to when I’ve produced more work on the subject. Proceed at your own risk.
    Introduction
    As I mentioned a few weeks ago I am working on a project on how scientific paradigms are developed. I generated a long list of questions and picked plate tectonics as my first case study. I immediately lost interest in the original questions and wanted to make a dependency graph/tech tree for the development of the paradigm, and this is just a personal project so I did that instead.
    I didn’t reach a stopping point with this graph other than “I felt done and wanted to start on my second case study”. I’m inconsistent about the level of detail or how far back I go. I tried to go back and mark whether data collection was motivated by theory or practical issues but didn’t fill it in for every node, even when it was knowable. Working on a second case study felt more useful than refining this one further so I’m shipping this version.
    “Screw it I’m shipping” is kind of the theme of this blog post, but that’s partially because I’m not sure which things are most valuable. Questions, suggestions, or additions are extremely welcome as they help me narrow in on the important parts. But heads up the answer might be “I don’t remember and don’t think it’s important enough to look up”. My current intention is to circle back after 1 or 2 more case studies and do some useful compare and contrast, but maybe I’ll find something better.
    And if you’re really masochistic, here’s the yEd file to play with.
    Scattered Thoughts
    Why I chose plate tectonics
    It’s recent enough to have relatively good documentation, but not so recent the major players are alive and active in field politics.
    It’s not a sexy topic, so while there isn’t much work on it what exists is pretty high quality.
    It is the only accepted paradigm in its field (for the implicit definition of paradigm in my head).
    Most paradigms are credited to one person on Wikipedia, even though that one person needed many other people’s work and the idea was refined by many people after they created it. Plate tectonics is the first I’ve found that didn’t do that. Continental drift is attributed to Alfred Wegener, but continental drift is not plate tectonics. Plate tectonics is acknowledged as so much of a group effort wikipedia doesn’t give anyone’s name.
    Content notes
    This graph is based primarily on Plate Tectonics: An Insider’s History of the Modern Theory of the Earth, edited by Naomi Oreskes. It also includes parts from this lecture by Christopher White, and Oxford’s Very Short Introduction to Plate Tectonics.
    Sources vary on how much Alfred Wegener knew when he proposed continental drift. Some say he only had the fossil and continental shape data, but the White video says he had also used synchronous geological layers and evidence of glacial travel.
    I tried to resolve this by reading Wegener’s original paper (translated into English) but it only left me more confused. He predicted cracks in plates being filled in by magma, but only mentions fossils once. Moreover he only brings them up to point to fossils of plants that are clearly maladapted to the climate of their current location, not the transcontinental weirdnesses. He does casually mention “Concerning South America and Africa, biologists and geologists are in close agreement that a Brazilian–African continent existed in the Mesozoic”, but clearly he’s not the first one to make that argument.
    I alas ran out of steam before trying Wegener’s book.
    I was stymied in attempts to check his references by the fact that they’re in German. If you really love reading historic academic German and would like to pair on this, please let me know....
    7 min
  • EA - Tracking Compute Stocks and Flows: Case Studies? by Cullen OKeefe
    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: Tracking Compute Stocks and Flows: Case Studies?, published by Cullen OKeefe on October 5, 2022 on The Effective Altruism Forum.
    Posted in my personal capacity
    The AGI governance community has recently converged on compute governance[1] as a promising lever for reducing existential risks from AI.
    One likely building block for any maximally secure compute governance regime is stock and flow accounting of (some kinds of) compute: i.e., requiring realtime accurate declaration to regulators of who possesses which uniquely numbered regulated chips, with penalties for undeclared or unauthorized[2] transfers.
    To understand the optimal design and feasibility of such a regime, we seek historical analogies for similar regimes. One that we are already familiar with include:
    Fissile nuclear material and other nuclear weapons components
    Firearms
    Some financial instruments
    Automobiles
    Real estate
    What are other good existing or historical analogies for compute stock and flow accounting? An ideal analogy will have many of the following traits:[3]
    The thing being tracked is a physical object
    The thing being tracked is economically important
    The thing being tracked is dual-use
    The tracking regime requires registration of current ownership and any transfers
    The tracking regime imposes penalties for failing to register ownership or transfer
    More and better examples of stock and flow accounting mechanisms could be extremely informative in improving compute governance proposals.
    Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org.
    2 min
  • EA - Philosophers speaking against the mistreatment of animals: The Montreal Declaration On Animal Exploitation by Fai
    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: Philosophers speaking against the mistreatment of animals: The Montreal Declaration On Animal Exploitation, published by Fai on October 5, 2022 on The Effective Altruism Forum.
    This week, 473 (so far) philosophers signed to speak against the human exploitation of nonhuman animals, in the Montreal Declaration On Animal Exploitation, it seems to me that it is pretty rare for philosophers to sign and publish such a co-statement.
    "We are researchers in the field of moral and political philosophy. Our work is rooted in different philosophical traditions, and we rarely find ourselves in agreement with one another. We do agree, however, on the need for a profound transformation of our relationships with other animals. We condemn the practices that involve treating animals as objects or commodities."
    My guess with not-so-much thought is that the declaration could be pretty influential to students and academics in the future. It might even feature in books on the history of philosophy in the future!
    To-do: The only EA-related philosophers on the list of signatories are Peter Singer (and he isn't formally linked with EA) and Jeff Sebo. I think it is pretty important (we have an animal welfare cause area but no EA philosopher on this list) and beneficial (strengthen this list and make it more powerful) for major EA philosophers to sign on this. If you are a philosopher, consider becoming a signatory by clicking this link.
    Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org.
    2 min
  • LW - Sleep Training 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: Sleep Training, published by jefftk on October 5, 2022 on LessWrong.
    Let's say you're a baby. You reach the end of a sleep cycle and become partially alert. What do you do? Ideally, if nothing were wrong you would settle back in for another sleep cycle. Chances are, however, you don't know how to do this yet, so you wake up more fully and start to cry. This seems to reliably trigger some combination of cuddles, shushing, motion, and/or nursing—comforted, you fall back asleep. While this is a fine outcome for you, your parents would probably be
    doing a lot better if you didn't rely on them so heavily at night.
    With a very little baby this is unavoidable, though it can be partially automated, but as babies get older they need to learn how to handle this themselves. Sleep training is a collection of strategies for teaching babies (a) the skill of falling asleep on their own
    ("self soothing") and (b) when they should apply this new skill.
    One of the things that makes discussion of sleep training messy is that people often aren't clear on which of (a) or (b) they're talking about. For example, a strategy of putting your kid down when they're very drowsy but not quite asleep yet is about teaching (a), while applying consistent patterns about when or how often you'll feed or settle them is about (b). If you think someone is trying to do (a) when they're actually doing (b), their approach can often sound excessive, uncaring, or unfair.
    It can also be hard to tell which of (a) or (b) your child needs. A child who can self-soothe but prefers to be cuddled doesn't appear that different from one who doesn't know how do it yet. Gradually providing less and less comfort can work well in these cases, because it's a good way to teach (a) and a so-so way of teaching (b). Alternatively, cry-it-out is a decent option for (b) and can also be a way of teaching (a) if other options haven't worked.
    With our youngest we found that tapering with the Snoo taught (a) reasonably well: while there was some crying she mostly was able to gradually build up the skill. With our oldest, pre-Snoo, we tried various methods of trying to teach (a), but it wasn't until we switched to cry-it-out that she really mastered the skill, and started taking longer naps.
    Our youngest still tests the boundaries of (b) sometimes. For example, two months ago (age 13m) we were on vacation (28 people in a 5br). She quickly figured out that every time she cried at night she quickly got to nurse: we didn't want her to wake others up. Over the course of the week she started crying more and more often during the night, correctly (and unfortunately) learning that the adult-implemented pattern for when she needed to go back to sleep on her own had shifted. After we got back home we had about a week of gradually re-teaching her the normal pattern.
    Another consideration I think is often missing in discussions of sleep training is that while crying is unpleasant for the baby, better-slept adults are able to be better parents in other ways. I think the baby typically is better off, even in the short term, with a couple weeks of cry-it-out vs months of zombie parents.
    Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org.
    4 min
  • EA - Upcoming EA conferences in 2023 (and 2022) 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: Upcoming EA conferences in 2023 (and 2022), published by Eli Nathan on October 5, 2022 on The Effective Altruism Forum.
    The Centre for Effective Altruism will be organizing and supporting conferences for the EA community all over the world in 2023. We currently have the following events scheduled:
    EA Global
    EA Global: Bay Area | (February 24–26, 2023)
    EA Global: London | (May 19–21, 2023)
    EA Global: US East Coast | (Fall/Autumn 2023) | Details TBC
    EAGx
    EAGxVirtual | (October 21–23, 2022)
    EAGxRotterdam | (November 4–6, 2022)
    EAGxBerkeley | (December 2–4, 2022)
    EAGxLatinAmerica | (January 7–8, 2023)
    EAGxIndia | (January 13–15, 2023)
    EAGxNordics | (April 21–23, 2023)
    Applications for EAGxVirtual and EAGxRotterdam are open (here), and we expect applications for the other conferences to open approximately 2 months before the event. Applications for EA Global: Bay Area will open in November.
    (If you'd like to add EA events like these directly to your Google Calendar, use this link.)
    Some notes on these conferences:
    EA Globals are run in-house by the CEA events team, whereas EAGx conferences are organized independently with financial support and mentoring from CEA.
    EA Global conferences have a high bar for admission and are for people who are very familiar with EA and are taking significant actions (e.g. full-time work or study) based on EA ideas.
    Admissions for EAGx conferences are processed independently by the EAGx conference organizers. These events are primarily for those who are newer to EA and interested in getting more involved and who are based in the region the conference is taking place in (e.g. EAGxIndia is primarily for people who are interested in EA and are based in India).
    Please apply to all conferences you wish to attend once applications open — we would rather get too many applications for some conferences and recommend that applicants attend a different one, than miss out on potential applicants to a conference.
    Applicants can request financial aid to cover the costs of travel, accommodation and tickets. If you are accepted to one of our conferences, we want you to attend without financial burden or stress, so if you are on the fence about applying please do so. For example, if you feel like you technically could afford the travel, but it would be financially uncomfortable to do so, we’d encourage you to apply for funding.
    The EA Global dates were selected so as to avoid any major public or religious holidays. We also wanted to space the conferences out as much as possible, and as such we’re hosting the Bay Area conference in the winter as it’s our only major location that has good weather around this time. We selected Oakland (rather than San Francisco as per usual) because of a good venue we found there, the Oakland Marriott.
    We also expect that there will be more EAGx events in or near the same locations as our EA Globals, but they may not all be planned out or booked in yet.
    Find more info on our website.
    As always, please feel free to email [email protected] with any questions, or comment below.
    Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org.
    4 min
  • EA - What is the most pressing feature to add to the forum? Upvote for in general, agreevote to agree with reasoning. 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: What is the most pressing feature to add to the forum? Upvote for in general, agreevote to agree with reasoning., published by Nathan Young on October 5, 2022 on The Effective Altruism Forum.
    A snap vote on features that people would like to see.
    Upvote based on some sense of what the community should want, agreevote (the tick) to adjust for you personally.
    There is a feature request thread, but I sense that old requests are getting lost. Hence, this. Edit, I'm shifting my suggested voting rules. Agreevotes should now just be about the reasoning of the comment, as usual. They were basically entirely correlated. It was a trial and I was wrong.
    Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org.
    1 min
  • AF - Tracking Compute Stocks and Flows: Case Studies? by Cullen OKeefe
    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: Tracking Compute Stocks and Flows: Case Studies?, published by Cullen OKeefe on October 5, 2022 on The AI Alignment Forum.
    Posted in my personal capacity
    The AGI governance community has recently converged on compute governance[1] as a promising lever for reducing existential risks from AI.
    One likely building block for any maximally secure compute governance regime is stock and flow accounting of (some kinds of) compute: i.e., requiring realtime accurate declaration to regulators of who possesses which uniquely numbered regulated chips, with penalties for undeclared or unauthorized[2] transfers.
    To understand the optimal design and feasibility of such a regime, we seek historical analogies for similar regimes. One that we are already familiar with include:
    Fissile nuclear material and other nuclear weapons components
    Firearms
    Some financial instruments
    Automobiles
    Real estate
    What are other good existing or historical analogies for compute stock and flow accounting? An ideal analogy will have many of the following traits:[3]
    The thing being tracked is a physical object
    The thing being tracked is economically important
    The thing being tracked is dual-use
    The tracking regime requires registration of current ownership and any transfers
    The tracking regime imposes penalties for failing to register ownership or transfer
    More and better examples of stock and flow accounting mechanisms could be extremely informative in improving compute governance proposals.
    Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org.
    2 min
  • AF - Paper: Discovering novel algorithms with AlphaTensor [Deepmind] by Lawrence Chan
    Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Paper: Discovering novel algorithms with AlphaTensor [Deepmind], published by Lawrence Chan on October 5, 2022 on The AI Alignment Forum.
    The authors apply an AlphaZero-like algorithm to discover new matrix multiplication algorithms. They do this by turning matrix multiplication into a one-player game, where the state represents how far from correct the current output is, moves are algorithmic instructions, and the reward is -1 per step (plus a terminal reward of -rank(final state), if the final state is not a zero tensor). On small matrices, they find that AlphaTensor can discover algorithms that use fewer scalar multiplications than the best known human-designed matrix multiplication algorithms. They apply this to find hardware-specific matmuls (by adding an additional reward equal to -time to the terminal state) that have a 10-20% larger speedup than Strassen's algorithm on NVIDIA V100s and TPU V2s (saving 4%/7.5% wall clock time).
    Paper abstract:
    Improving the efficiency of algorithms for fundamental computations can have a widespread impact, as it can affect the overall speed of a large amount of computations. Matrix multiplication is one such primitive task, occurring in many systems—from neural networks to scientific computing routines. The automatic discovery of algorithms using machine learning offers the prospect of reaching beyond human intuition and outperforming the current best human-designed algorithms. However, automating the algorithm discovery procedure is intricate, as the space of possible algorithms is enormous. Here we report a deep reinforcement learning approach based on AlphaZero for discovering efficient and provably correct algorithms for the multiplication of arbitrary matrices. Our agent, AlphaTensor, is trained to play a single-player game where the objective is finding tensor decompositions within a finite factor space.
    AlphaTensor discovered algorithms that outperform the state-of-the-art complexity for many matrix sizes. Particularly relevant is the case of 4 × 4 matrices in a finite field, where AlphaTensor’s algorithm improves on Strassen’s two-level algorithm for the first time, to our knowledge, since its discovery 50 years ago. We further showcase the flexibility of AlphaTensor through different use-cases: algorithms with state-of-the-art complexity for structured matrix multiplication and improved practical efficiency by optimizing matrix multiplication for runtime on specific hardware. Our results highlight AlphaTensor’s ability to accelerate the process of algorithmic discovery on a range of problems, and to optimize for different criteria.
    Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org.
    3 min
  • LW - Adversarial vs Collaborative Contexts 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: Adversarial vs Collaborative Contexts, published by jefftk on October 4, 2022 on LessWrong.
    You can often make pretty clear predictions about what sort of effect something would have in the world as it exists today, but when you start to take into account how the world might change in response it gets very difficult. One tool here is thinking about whether you mostly expect collaborative or adversarial effects.
    In a collaborative context other people respond to your change by building on it, and it typically has a larger effect than you'd naively expect. For example, say you create an icon to communicate an idea in your program, other people start using it, and it becomes the standard representation for the concept. While an A/B test might have been only mildly positive, or even negative, after the world adapted to the new icon it would have become very clearly the right symbol to use.
    In an adversarial context other people respond to your change by countering it, and it typically has a smaller effect. You figure out a way to detect bot traffic, but the bot operators are very motivated to shift their behavior to undo your work.
    Some contexts are also neutral, where both collaborative and adversarial effects are minimal or nonexistent. A baby toy design doesn't get better or worse with the generations.
    Some examples:
    Diseases are adversarial: antibiotics lose resistance, new variants evade immunity. While something looks good in initial testing, that advantage will erode over time. An interesting exception here is cancer, which (almost always) has to evolve anew in each patient, and so cancer treatments don't become less effective over time (in new patients).
    Adopting standards is collaborative: each computer with USB increased demand for devices that spoke USB and vice-versa. Someone thinking "let's not bother with USB, it's not that widely supported yet" would have been ignoring collaborative effects.
    Same with building on a platform: the more iPhones sold the more people wanted apps, and the more apps there were the more people wanted iPhones.
    Interaction with nature is generally adversarial: cockatoos figure out how to get into trash, weeds evolve herbicide resistance.
    A new kind of ad has both kinds of effects: adversarial in that people will notice novel ads more, but cooperative in that advertisers haven't yet learned how to make the most of the format.
    This sort of effect is especially important to think through when considering rolling something out based on the results of an experiment. You've almost certainly not run the experiment long enough to see how the world will adjust, but you can use these patterns to predict whether the long-term results will be stronger or weaker than you've measured.
    Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org.
    3 min
  • AF - Reflection Mechanisms as an Alignment target: A follow-up survey by Marius Hobbhahn
    Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Reflection Mechanisms as an Alignment target: A follow-up survey, published by Marius Hobbhahn on October 5, 2022 on The AI Alignment Forum.
    This is the second of three posts (part I) about surveying moral sentiments related to AI alignment. This work was done by Marius Hobbhahn and Eric Landgrebe under the supervision of Beth Barnes as part of the AI safety camp 2022.
    TL;DR: We find that the results of our first study, i.e. that humans tend to agree with conflict resolution mechanisms hold under different wordings but are weakened in adversarial scenarios (where we actively try to elicit less agreement). Furthermore, we find that people tend to agree much less with a mechanism when the decision-maker is a smart benevolent AI rather than a smart benevolent human. A positive interpretation of these findings is that humans are fine with giving power to a conflict resolution mechanism as long as humans are ultimately in control.
    Introduction
    In the first post, we surveyed 1000 US respondents about their moral beliefs, the conditions under which they would change their moral beliefs and how they felt about mechanisms to resolve moral disagreements such as democracy or debate.
    Our main findings of the first survey were
    Unsurprisingly, people have very different moral beliefs, e.g. on the morality of abortions, immigration or eating meat.
    They very rarely report to changing or wanting to change these beliefs, e.g. most participants (>80%) report having not changed them in the last 10 years and do not expect to change them in the next 10 years.
    However, they mostly think that mechanisms to resolve moral disagreements, such as democracy or debate, are good even when they disagree with the outcome. In other words, people are willing to accept outcomes that are different from their beliefs if they trust the process by which it was derived.
    We think this finding has some implications for AI alignment. Most importantly, we think that the alignment target of an AI should be a mechanism to resolve moral conflicts such as democracy or debate rather than a set of individual beliefs, e.g. about abortion or veganism. This is more likely to lead to decisions that can be accepted by most people and prevent a race from “aligning an AI to your personal beliefs at the cost of other people’s beliefs”.
    We think the main weakness of our first result was that it could have been a result of confounders or randomness. We asked people for their opinions on multiple scenarios, but our questions still had some shortcomings. These include:
    Specificity: Maybe people would change their minds if they had a very specific scenario in mind rather than just an abstract question.
    Robustness: Maybe the way we ask our questions makes one answer more likely than another.
    Social desirability: Maybe people answer what they think is socially desirable.
    Robustness to different scenarios: Maybe people’s opinions change when we change key variables of the process, e.g. when we specify a bad outcome or when we change who makes the decisions.
    Relation to AI: We never talk about AIs specifically, we always talk about smart, benevolent and nice future humans (see below for details).
    Therefore, we ran a follow-up study that attempts to address these problems. Concretely, we first run robustness tests by asking the same question in different ways. Then we run adversarial tests where we actively try to get people to give more disagreeing answers. Finally, we ask the same question that was asked with smart, benevolent and nice future humans but replace “humans” with “AIs”.
    Main findings
    Robustness
    To test the robustness of our questions, we randomly assigned one of the following three questions to a participant.
    Basic: This was the same scenario as in the previous study and was mostly used to provide a reference. People wer...
    12 min