The Nonlinear Library

The Nonlinear Library

By The Nonlinear FundEducation
Download on the App Store

All The Nonlinear Library episodes

  • LW - DeepMind: Model evaluation for extreme risks by Zach Stein-Perlman
    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: DeepMind: Model evaluation for extreme risks, published by Zach Stein-Perlman on May 25, 2023 on LessWrong.
    Current approaches to building general-purpose AI systems tend to produce systems with both beneficial and harmful capabilities. Further progress in AI development could lead to capabilities that pose extreme risks, such as offensive cyber capabilities or strong manipulation skills. We explain why model evaluation is critical for addressing extreme risks. Developers must be able to identify dangerous capabilities (through "dangerous capability evaluations") and the propensity of models to apply their capabilities for harm (through "alignment evaluations"). These evaluations will become critical for keeping policymakers and other stakeholders informed, and for making responsible decisions about model training, deployment, and security.
    This is the first great public writeup on model evals for averting existential catastrophe. I think it's likely that if AI doesn't kill everyone, developing great model evals and causing everyone to use them will be a big part of that. So I'm excited about this paper both for helping AI safety people learn more and think more clearly about model evals and for getting us closer to it being common knowledge that responsible labs should use model evals and responsible authorities should require them (by helping communicate model evals more widely, in a serious/legible manner).
    Non-DeepMind authors include Jade Leung (OpenAI governance lead), Daniel Kokotajlo (OpenAI governance), Jack Clark (Anthropic cofounder), Paul Christiano, and Yoshua Bengio.
    For more on model evals for AI governance, see ARC Evals, including Beth's EAG talk Safety evaluations and standards for AI and the blogpost Update on ARC's recent eval efforts (LW).
    Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org.
    2 min
  • EA - Introducing Allied Scholars for Animal Protection by Dr Faraz Harsini
    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 Allied Scholars for Animal Protection, published by Dr Faraz Harsini on May 24, 2023 on The Effective Altruism Forum.
    We’re excited to introduce Allied Scholars for Animal Protection, a nonprofit creating a unified infrastructure for effective and sustainable animal advocacy at universities. Our mission is to organize, train, and mentor students who are interested in advocating for animal welfare and pursuing impactful careers.
    The Problem
    Universities play a critical role in shaping the future of society and effecting systemic change. As future leaders, college students hold immense potential for driving progress and cultural transformation.
    Unfortunately, animal advocacy in universities tends to be limited, sporadic, and unsustainable. The existing clubs on campuses operate independently with no coordination, and students are often hindered by a lack of time, training, experience, and support. Often, when active students graduate, their animal advocacy clubs become inactive. Much time and effort are wasted due to a lack of continuity and longevity of animal advocacy on campuses because students have to reinvent the proverbial wheel each time they restart a group.
    One of the worst consequences of this is that much talent goes untapped due to insufficient education and inspiration for vegans to choose effective and impactful careers. The EA Community is working hard to reach this talent through career advising and community building. We believe that on-the-ground support for university animal rights clubs can complement EA recruitment efforts and can encourage vegan college students to engage with critical intellectual work being done by the EA community. We also think that community building work focused specifically on animal advocacy can help reach vegans who might not be as interested in other cause areas or the broader EA project.
    Some animal advocacy organizations provide opportunities for students to volunteer, but enabling a strong campus movement is not the sole focus of these organizations. Having a single organization dedicated to providing infrastructure for campus activism would therefore fill a highly neglected niche in the current animal advocacy ecosystem.
    Building a strong campus animal advocacy movement is also highly tractable. There are many vegan students out there who care deeply about these issues but do not feel they have the knowledge or resources to organize a group of their own. By providing the needed support, we can dramatically lower the barrier of entry to vegan advocacy and broaden the pool of talent going towards highly impactful careers.
    Our Approach
    Animal organizations often focus on training individual students rather than on building a sustainable vegan community. ASAP takes a more holistic approach. Our strategy for constructing a strong campus animal movement involves the following:
    Building a growing vegan community while investing in and strengthening individuals. This means conducting outreach to vegans who might like to become more active as advocates, and training vegans to conduct effective outreach to nonvegans.
    Providing on-the-ground support to student groups.
    Collecting thousands of signatures through petitions.
    Streamlining the process of starting and running student animal advocacy groups.
    Facilitating systemic and long-term educational frameworks, rather than just one-time events. We will provide educational seminars to empower vegans and educate the general student population, with a special emphasis on plant-based nutrition for future healthcare professionals.
    Fighting speciesism and humane-washing while promoting plant-based options.
    By facilitating more effective student advocacy, we believe ASAP can help produce more influential vegans who push for change. We want to inspire the next Eric Adams, Co...
    8 min
  • LW - Adumbrations on AGI from an outsider by nicholashalden
    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: Adumbrations on AGI from an outsider, published by nicholashalden on May 24, 2023 on LessWrong.
    Preamble
    A lot of people have written against AI Doom, but I thought it might be interesting to give my account as an outsider encountering these arguments. Even if I don’t end up convincing people who have made AI alignment central to their careers and lives, maybe I’ll at least help some of them understand why the general public, and specifically the group of intelligent people which encounters their arguments, is generally not persuaded by their material. There may be inaccuracies in my account of the AI Doom argument, but this is how I think it’s generally understood by the average intelligent non-expert reader.
    I started taking AI alignment arguments seriously when GPT-3 and GPT-4 came out, and started producing amazing results on standardized testing and writing tasks. I am not an ML engineer, do not know much about programming, and am not part of the rationalist community that has been structured around caring deeply about AI risk for the last fifteen years. It may be of interest that I am a professional forecaster, but of financial asset prices, not of geopolitical events or the success of nascent technologies. My knowledge of the arguments comes mostly from reading LessWrong, ACX and other online articles, and specifically I’m responding to Eliezer’s argument detailed in the pages on Orthogonality, Instrumental Convergence, and List of Lethalities (plus the recent Time article).
    I. AI doom is unlikely, and it’s weird to me that clearly brilliant people think it’s >90% likely
    I agree with the following points:
    An AI can probably get much smarter than a human, and it’s only a matter of time before it does
    Something being very smart doesn’t make it nice (orthogonality, I think)
    A superintelligence doesn’t need to hate you to kill you; any kind of thing-maximizer might end up turning the atoms you’re made of into that thing without specifically wanting to destroy you (instrumental convergence, I think)
    Computers hooked up to the internet have plenty of real-world capability via sending emails/crypto/bank account hacking/every other modern cyber convenience.
    The argument then goes on to say that, if you take a superintelligence and tell it to build paperclips, it’s going to tile the universe with paperclips, killing everyone in the process (oversimplified). Since the people who use AI are obviously going to tell it to do stuff–we already do that with GPT-4–as soon as it gains superintelligence capabilities, our goose is collectively cooked. There is a separate but related argument, that a superintelligence would learn to self-modify, and instead of building the paperclips we asked it to, turn everything into GPUs so it can maximize some kind of reward counter. Both of these seem wrong to me.
    The first argument–paperclip maximizing–is coherent in that it treats the AGI’s goal as fixed and given by a human (Paperclip Corp, in this case). But if that’s true, alignment is trivial, because the human can just give it a more sensible goal, with some kind of “make as many paperclips as you can without decreasing any human’s existence or quality of life by their own lights”, or better yet something more complicated that gets us to a utopia before any paperclips are made. We can argue over the hidden complexity of wishes, but it’s very obvious that there’s at least a good chance the populace would survive, so long as humans are the ones giving the AGI its goal. And, there’s a very good chance the first AGI-wishers will be people who care about AI safety, and not some random guy who wants to make a few million by selling paperclips.
    At this point, the AGI-risk argument responds by saying, well, paperclip-maximizing is just a toy thought experiment for people to understand. In fac...
    13 min
  • EA - How I solved my problems with low energy (or: burnout) by Luise
    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 I solved my problems with low energy (or: burnout), published by Luise on May 24, 2023 on The Effective Altruism Forum.
    I had really bad problems with low energy and tiredness for about 2 years. This post is about what I’ve learned. This is not a general guide to solving any and all energy problems since I mostly only have one data point. I still hope reading about my personal problems and solutions will help people solve theirs.
    Summary
    I had basically two periods of very low energy: college and last summer.
    In college, I felt tired literally all day, especially as soon as I tried to study. I was also depressed.
    In the summer, I was very happy but I had days at a time where I wouldn’t do anything productive. All tasks seemed unbearably hard to me, sometimes even writing a simple email. I also became introverted.
    I thought I was being lazy and just needed to “get over it”. Starting to notice I had a ‘real’ problem was a big step forward.
    I learned that I actually had multiple hard-to-disentangle problems:
    I’m sensitive to disruptions in my sleep leading to feeling tired.
    Certain types of work that are both hard and demotivating also make me feel physically tired.
    My biggest realization was that I was burned out much of last summer. This was because I didn’t give myself rest, even though I didn’t see it that way at the time. This led to the unproductive days (not laziness).
    In college, I lived a weird lifestyle regarding sleep, social life, and other things. Some part of this was probably bad. Having common sense would’ve helped.
    I can now notice symptoms of overloading myself before it leads to burnout. Learning to distinguish this from “being lazy” phenomenologically was crucial.
    My problems had nothing to do with physical health or stress.
    When experimenting to solve my problems, it was useful for me to track when I had unproductive days. This way I could be sure how much the experiments impacted me.
    What my problems were like (so you know whether they’re similar to yours)
    A typical low-energy day while I was in college in first year:
    I wake up at 12 pm. I slept 9 hours but I’m tired. It doesn’t go away even after an hour. I open my math book. But literally as soon as I read the first sentences, I feel so tired that I physically want to lie down and close my eyes. It feels very hard to keep reading. Often I just stare at the wood of the table right next to my book. Not doing anything, just avoiding thinking. Even staring at the wall for 10 minutes sounds great right now. I never really stop feeling tired until it’s night again.
    A typical low-energy day while I was working on EA community building projects in the summer:
    I have to do a task I usually love doing, maybe reading applications for an event I’m running. But as soon as I look at the wall of Airtable fields and text, the task feels way too large. I will have to think deeply about these answers people wrote in the application form and make difficult decisions, drawing on information from over 20 fields. That depth of thinking and amount of working memory sounds way too hard right now. I try, but 3 minutes later I give up. I decide to read something instead. I feel the strong desire to sit in a comfy bean bag and get a blanket. Even sitting upright in an office chair feels hard. I start reading. The text requires slight cognitive effort on the part of the reader to understand. It sounds too hard. I stare at a sentence, willing myself to think. I give up after 3 sentences.
    It’s lunchtime. I used to love lunchtime at the office because I get to chat with all these super cool people and because I’m quite extraverted. But now the idea of a group conversation sounds way too much. I don’t even want to chat to a single person. I would have to be ‘switched on’, think of things to say, smile, and I just don’t have...
    22 min
  • EA - AGI Catastrophe and Takeover: Some Reference Class-Based Priors by zdgroff
    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: AGI Catastrophe and Takeover: Some Reference Class-Based Priors, published by zdgroff on May 24, 2023 on The Effective Altruism Forum.
    This is a linkpost for
    I am grateful to Holly Elmore, Michael Aird, Bruce Tsai, Tamay Besiroglu, Zach Stein-Perlman, Tyler John, and Kit Harris for pointers or feedback on this document.
    Executive Summary
    Overview
    In this document, I collect and describe reference classes for the risk of catastrophe from superhuman artificial general intelligence (AGI). On some accounts, reference classes are the best starting point for forecasts, even though they often feel unintuitive. To my knowledge, nobody has previously attempted this for risks from superhuman AGI. This is to a large degree because superhuman AGI is in a real sense unprecedented. Yet there are some reference classes or at least analogies people have cited to think about the impacts of superhuman AI, such as the impacts of human intelligence, corporations, or, increasingly, the most advanced current AI systems.
    My high-level takeaway is that different ways of integrating and interpreting reference classes generate priors on AGI-caused human extinction by 2070 anywhere between 1/10000 and 1/6 (mean of ~0.03%-4%). Reference classes offer a non-speculative case for concern with AGI-related risks. On this account, AGI risk is not a case of Pascal’s mugging, but most reference classes do not support greater-than-even odds of doom. The reference classes I look at generate a prior for AGI control over current human resources anywhere between 5% and 60% (mean of ~16-26%). The latter is a distinctive result of the reference class exercise: the expected degree of AGI control over the world looks to far exceed the odds of human extinction by a sizable margin on these priors. The extent of existential risk, including permanent disempowerment, should fall somewhere between these two ranges.
    This effort is a rough, non-academic exercise and requires a number of subjective judgment calls. At times I play a bit fast and loose with the exact model I am using; the work lacks the ideal level of theoretical grounding. Nonetheless, I think the appropriate prior is likely to look something like what I offer here. I encourage intuitive updates and do not recommend these priors as the final word.
    Approach
    I collect sets of events that superhuman AGI-caused extinction or takeover would be plausibly representative of, ex ante. Interpreting and aggregating them requires a number of data collection decisions, the most important of which I detail here:
    For each reference class, I collect benchmarks for the likelihood of one or two things:
    Human extinction
    AI capture of humanity’s available resources.
    Many risks and reference classes are properly thought of as annualised risks (e.g., the yearly chance of a major AI-related disaster or extinction from asteroid), but some make more sense as risks from a one-time event (e.g., the chance that the creation of a major AI-related disaster or a given asteroid hit causes human extinction). For this reason, I aggregate three types of estimates (see the full document for the latter two types of estimates):
    50-Year Risk (e.g. risk of a major AI disaster in 50 years)
    10-Year Risk (e.g. risk of a major AI disaster in 10 years)
    Risk Per Event (e.g. risk of a major AI disaster per invention)
    Given that there are dozens or hundreds of reference classes, I summarise them in a few ways:
    Minimum and maximum
    Weighted arithmetic mean (i.e., weighted average)
    I “winsorise”, i.e. replace 0 or 1 with the next-most extreme value.
    I intuitively downweight some reference classes. For details on weights, see the methodology.
    Weighted geometric mean
    Findings for Fifty-Year Impacts of Superhuman AI
    See the full document and spreadsheet for further details on how I arrive at these figures.
    ...
    13 min
  • EA - New s-risks audiobook available now by Alistair Webster
    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 s-risks audiobook available now, published by Alistair Webster on May 24, 2023 on The Effective Altruism Forum.
    Tobias Baumann's first-of-its-kind introduction to s-risks, Avoiding the Worst: How to Prevent a Moral Catastrophe is now available to listen to for free.
    Professionally narrated by Adrian Nelson, the full audiobook is out now on Audible and other audiobook stores. Additionally, a captioned video can be listened to for free on the CRS YouTube channel.
    Running at just 2 hours and 40 minutes, the audiobook packs in a comprehensive introduction to the topic, explaining what s-risks are, whether we should focus on them, and what we can do now to reduce the likelihood of s-risks occurring.
    The eBook is also available in various formats, or you can read a PDF version.
    Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org.
    1 min
  • EA - KFC Supplier Sued for Cruelty by alene
    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: KFC Supplier Sued for Cruelty, published by alene on May 24, 2023 on The Effective Altruism Forum.
    Dear EA Forum readers,
    The EA charity, Legal Impact for Chickens (LIC), just filed our second lawsuit!
    As many of you know, LIC is a litigation nonprofit dedicated to making factory-farm cruelty a liability. We focus on chickens because of the huge numbers in which they suffer and the extreme severity of that suffering.
    Today, we sued one of the country’s largest poultry producers and a KFC supplier, Case Farms, for animal cruelty.
    The complaint comes on the heels of a 2021 undercover investigation by Animal Outlook, revealing abuse at a Morganton, N.C. Case Farms hatchery that processes more than 200,000 chicks daily.
    Our lawsuit attacks the notion that Big Ag is above the law. We are suing under North Carolina's 19A statute, which lets private parties enjoin animal cruelty.
    Case Farms was documented knowingly operating faulty equipment, including a machine piston which repeatedly smashes chicks to death and a dangerous metal conveyor belt which traps and kills young birds. Case Farms was also documented crushing chicks’ necks between heavy plastic trays.
    Case Farms supplies its chicken to KFC, Taco Bell, and Boar’s Head, among other customers.
    Thank you so much to all the EA Forum readers who helped make this happen, by donating to, and volunteering for, Legal Impact for Chickens!
    Sincerely,
    Alene
    Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org.
    2 min
  • LW - The Office of Science and Technology Policy put out a request for information on A.I. by HiroSakuraba
    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 Office of Science and Technology Policy put out a request for information on A.I., published by HiroSakuraba on May 24, 2023 on LessWrong.
    This request for information does cover some of the possible existential risks, in the first section. I am going to submit a few responses of my own, and I am hopeful others will do the same.
    1. What specific measures – such as standards, regulations, investments, and improved trust and safety practices – are needed to ensure that AI systems are designed, developed, and deployed in a manner that protects people’s rights and safety? Which specific entities should develop and implement these measures?
    2. How can the principles and practices for identifying and mitigating risks from AI, as outlined in the Blueprint for an AI Bill of Rights and the AI Risk Management Framework, be leveraged most effectively to tackle harms posed by the development and use of specific types of AI systems, such as large language models?
    3. Are there forms of voluntary or mandatory oversight of AI systems that would help mitigate risk? Can inspiration be drawn from analogous or instructive models of risk management in other sectors, such as laws and policies that promote oversight through registration, incentives, certification, or licensing?
    4. What are the national security benefits associated with AI? What can be done to maximize those benefits?
    5. How can AI, including large language models, be used to generate and maintain more secure software and hardware, including software code incorporating best practices in design, coding and post deployment vulnerabilities?
    6. How can AI rapidly identify cyber vulnerabilities in existing critical infrastructure systems and accelerate addressing them?
    7. What are the national security risks associated with AI? What can be done to mitigate these risks?
    8. How does AI affect the United States’ commitment to cut greenhouse gases by 50-52% by 2030, and the Administration’s objective of net-zero greenhouse gas emissions no later than 2050? How does it affect other aspects of environmental quality?
    Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org.
    3 min
  • EA - Top Idea Reports from the EA Philippines Mental Health Charity Ideas Research Project by Shen Javier
    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: Top Idea Reports from the EA Philippines Mental Health Charity Ideas Research Project, published by Shen Javier on May 24, 2023 on The Effective Altruism Forum.
    In October 2021 to May 2022, EA Philippines organized the Mental Health Charity Ideas Research project. The project's goal was to find ideas that can become highly impactful and cost-effective charities in improving the well-being of people living in the Philippines and other low- to middle-income countries. It focused on children and adolescent mental health.
    This was a follow-up to the participation of Brian Tan and myself in Charity Entrepreneurship’s 2021 Incubation Program, in their region-specific track for training people to research the top charity ideas in a region. The project was awarded $11,000 in funding from the EA Infrastructure Fund in 2021 for 1.2 FTE in salary for the project for 8 months. Brian transitioned to being an advisor of the project early on, and AJ Sunglao was brought on as a part-time project co-lead, while two part-time researchers (Mae Muñoz, and Zam Superadble) were also hired.
    Links to our reports
    We already held a brown bag session last June 11, 2022 discussing the research process and introducing the top four charity ideas we found last year. Now, we share deep reports on those ideas that detail the evidence supporting their effectiveness and how one might implement the charities in the Philippines. We also share the shallow reports made for the other top mental health interventions.
    Access the reports here:
    Deep Reports
    Self-Help Workbooks for Children and Adolescents in the Philippines and Low-to-Middle-Income Countries
    School-based Psychoeducation in the Philippines and Low-to-Middle-Income Countries
    Guided Self-Help Game-based App for Adolescents in the Philippines and Low-to-Middle-Income Countries
    Shallow Reports
    Here’s a quick guide to our top ideas:
    Idea NameDescriptionCost-Effectiveness ($ per unit, total costs)Self-Help Workbooks for Children and AdolescentsThis intervention will develop and distribute self-help workbooks to improve depression and anxiety symptoms in children and young adolescents, particularly 6 to 18-year-olds. Depending on the severity of mental health disorders, the workbook can be accompanied by weekly guidance by lay counselors through telephone, email, social media, or other available platforms.School-based PsychoeducationThis preventive approach entails training and supervising teachers to deliver psychoeducation on mental health topics in their respective schools. Through weekly participatory learning sessions, students would learn to apply positive coping strategies, build interpersonal skills, and/or develop personal characteristics that would empower them to care for their mental health and navigate important life transitions.
    Guided Self-Help Game-based App for Adolescents The intervention is a self-help game-based mobile application for help-seeking adolescents aged 12 - 19 years old. As a self-help format, the app aims to teach service users concepts and skills that will aid them in addressing MH concerns. The content of the app will be based on evidence-based therapeutic modalities. The game-based format is used to enhance service user engagement and prevent dropout. Youth-led Mental Health SupportThis intervention is a community-based intervention for adolescents aged 13-18. It uses task-sharing principles in delivering basic para-mental health support by training community members like SK officials and student leaders in basic mental health skills such as psychoeducation, peer counseling, and psychological first aid. The content of the training would be based on other community-based interventions like Thinking Healthy Programme, PM+, and Self Help+.
    $2.67 per WHO-5 improvement
    $85.93 per GSES improvement
    $69.47 per SWEMWBS impro...
    11 min
  • AF - [Linkpost] Interpretability Dreams by DanielFilan
    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: [Linkpost] Interpretability Dreams, published by DanielFilan on May 24, 2023 on The AI Alignment Forum.
    A brief research note by Chris Olah about the point of mechanistic interpretability research. Introduction and table of contents are below.
    Interpretability Dreams
    An informal note on the relationship between superposition and distributed representations by Chris Olah. Published May 24th, 2023.
    Our present research aims to create a foundation for mechanistic interpretability research. In particular, we're focused on trying to resolve the challenge of superposition. In doing so, it's important to keep sight of what we're trying to lay the foundations for. This essay summarizes those motivating aspirations – the exciting directions we hope will be possible if we can overcome the present challenges.
    We aim to offer insight into our vision for addressing mechanistic interpretability's other challenges, especially scalability. Because we have focused on foundational issues, our longer-term path to scaling interpretability and tackling other challenges has often been obscure. By articulating this vision, we hope to clarify how we might resolve limitations, like analyzing massive neural networks, that might naively seem intractable in a mechanistic approach.
    Before diving in, it's worth making a few small remarks. Firstly, essentially all the ideas in this essay were previously articulated, but buried in previous papers. Our goal is just to surface those implicit visions, largely by quoting relevant parts. Secondly, it's important to note that everything in this essay is almost definitionally extremely speculative and uncertain. It's far from clear that any of it will ultimately be possible. Finally, since the goal of this essay is to lay out our personal vision of what's inspiring to us, it may come across as a bit grandiose – we hope that it can be understood as simply trying to communicate subjective excitement in an open way.
    Overview
    An Epistemic Foundation - Mechanistic interpretability is a "microscopic" theory because it's trying to build a solid foundation for understanding higher-level structure, in an area where it's very easy for us as researchers to misunderstand.
    What Might We Build on Such a Foundation? - Many tantalizing possibilities for research exist (and have been preliminarily demonstrated in InceptionV1), if only we can resolve superposition and identify the right features and circuits in a model.
    Larger Scale Structure - It seems likely that there is a bigger picture, more abstract story that can be built on top of our understanding of features and circuits. Something like organs in anatomy or brain regions in neuroscience.
    Universality - It seems likely that many features and circuits are universal, forming across different neural networks trained on similar domains. This means that lessons learned studying one model give us footholds in future models.
    Bridging the Microscopic to the Macroscopic - We're already seeing that some microscopic, mechanistic discoveries (such as induction heads) have significant macroscopic implications. This bridge can likely be expanded as we pin down the foundations, turning our mechanistic understanding into something relevant to machine learning more broadly.
    Automated Interpretability - It seems very possible that AI automation of interpretability may help it scale to large models if all else fails (although aesthetically, we might prefer other paths).
    The End Goals - Ultimately, we hope this work can eventually contribute to safety and also reveal beautiful structure inside neural networks.
    Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org.
    4 min