The Nonlinear Library

The Nonlinear Library

By The Nonlinear FundEducation
Download on the App Store

All The Nonlinear Library episodes

  • AF - Trying to find the underlying structure of computational systems by Matthias G. Mayer
    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: Trying to find the underlying structure of computational systems, published by Matthias G. Mayer on September 13, 2022 on The AI Alignment Forum.
    This post was written in the SERI MATS program under John Wentworth.
    In our quest to find steering mechanisms in computational systems, we first have to find the right framework to look at the internals of such a system. This framework should naturally express the systems we normally associate with computation, like the brain and computers themselves, as well as more abstract interfaces like Turing Machines. Turing Machines, arithmetic operations, and Neural Nets are in some sense equivalent, so if we use our framework on equivalent computational objects (E.g. a Neural Net and a Turing Machine that simulates it), the representations we get should also be equivalent.
    We also want to find meaningful concepts in our chosen framework, like modularity, features, search, and steering mechanisms. But it is hard to evaluate beforehand if we will be able to find these things in a given framework.
    My first intuition was that Computational Graphs is what we want. They don’t impose a rigid structure like e.g. Turing Machines would, and we can naturally express many things with them. (I don’t know if they can express the brain’s computation.)
    This is essentially the same as causal DAG.
    There is some ambiguity in which nodes we choose. We can express certain bit flips in a processor aggregated as addition.
    If we choose the algebraic level of detail (left in picture), we implicitly assume that there is no relevant thing going on if we look at the bits (right in picture). Intuitively this seems likely. If we would, for example, look for modularity in a NN, it feels like a module wouldn’t ‘begin’ inside an addition process. Features should not be encoded inside it either.
    Computational Graphs work very well for Neural Networks because the graph does not change when we change the inputs. In general, this is not the case. I have not found a crisp way of expressing if-branching in arithmetic graphs. Let's say we want to express the following program:
    else: return b - c
    It is clear that we can use a Turing Machine for this. When we change the detail level to arithmetic, however, it is not clear to me how to express that the graph depends on a.
    Here the result of a computation does not change which values are propagated through the graph. Rather, it changes the graph itself. To be clear, I am not looking for a way to express an if-statement in arithmetic (possibly with activation functions). I want to extend the notion of graphs to allow for naturally expressing that a computation result changes the nodes of the graph itself.
    Nonetheless, it might still be worth looking at computational graphs since we don't face those issues when looking at Neural Nets.
    Computational States
    If we take an actual implementation of a program that produced a computational graph and stop it at a time t, then at this time, the computational state of the program corresponds to the value of all nodes for which a successor is still not computed. In this case, the result of their computation is still relevant to the process and can’t be thrown away. I call a subset of nodes that correspond to a computational state an incomplete Cut, and the full set a complete Cut.
    A Cut is complete if and only if every path from input to output goes through it. In the causal DAG setting it corresponds to a Markov Blanket.
    Neuron vs Layer vs Cut Activation Space
    Layer Activation Space is a generalization of looking at neurons: If we optimize activations for the length of the projection onto ei then this is the same as disregarding all components except the ith neuron and maximizing its activation. It is not intuitively clear to me that 'projection on a vector' is the right notion of 'fea...
    7 min
  • EA - Podcast: Questions for John Halstead by Ben Yeoh
    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: Questions for John Halstead, published by Ben Yeoh on September 13, 2022 on The Effective Altruism Forum.
    Podcast: Questions for John Halstead
    I'm podcasting with John soon hoping to cover his climate work, long-termism and general worldview. If you haven't read it, his work on climate change & longtermism is, IMO, very impressive:
    What questions should I ask?
    The podcast is generalist as opposed to EA-specific but has covered some EA topics in depth (both arguments for and against, eg Leopold Aschenbrenner, Larry Temkin).
    Let me know any good questions. Thanks!
    Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org.
    1 min
  • LW - AI Safety field-building projects I'd like to see by Akash
    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: AI Safety field-building projects I'd like to see, published by Akash on September 11, 2022 on LessWrong.
    People sometimes ask me what types of AIS field-building projects I would like to see.
    Here’s a list of 11 projects.
    Background points/caveats
    But first, a few background points.
    These projects require people with specific skills/abilities/context in order for them to go well. Some of them also have downside risks. This is not a “list of projects Akash thinks anyone can do” but rather a “list of projects that Akash thinks could Actually Reduce P(Doom) if they were executed extremely well by an unusually well-qualified person/team.”
    I strongly encourage people to reach out to experienced researchers/community-builders before doing big versions of any of these. (You may disagree with their judgment, but I think it’s important to at least have models of what they believe before you do something big.)
    This list represents my opinions. As always, you should evaluate these ideas for yourself.
    If you are interested in any of these, feel free to reach out to me. If I can’t help you, I might know someone else who can.
    Reminder that you can apply for funding from the long-term future fund. You don’t have to apply to execute a specific project. You can apply for career exploration grants, grants that let you think about what you want to do next, and grants that allow you to test out different hypotheses/uncertainties.
    I sometimes use the word “organization”, which might make it seem like I’m talking about 10+ people doing something over the course of several years. But I actually mean “I think a team of 1-3 people could probably test this out in a few weeks and get something ambitious started here within a few months if they had relevant skills/experiences/mentorship.
    These projects are based on several assumptions about AI safety, and I won’t be able to articulate all of them in one post. Some assumptions include “AIS is an extremely important cause area” and “one of the best ways to make progress on AI safety is to get talented people working on technical research.” If I’m wrong, I think I’m wrong because I’m undervaluing non-technical interventions that could buy us more time (e.g., strategies in AI governance/strategy or strategies that involve outreach to leaders of AI companies). I plan to think more about those in the upcoming weeks.
    Some projects I am excited about
    Global Talent Search for AI Alignment Researchers
    Purpose: Raise awareness about AI safety around the world to find highly talented AI safety researchers.
    How this reduces P(doom): Maybe there are extremely promising researchers (e.g., people like Paul Christiano and Eliezer Yudkowsky) out in the world who don’t know about AI alignment or don’t know how to get involved. One global talent search program could find them. Alternatively, maybe we need 1000 full-time AI safety researchers who are 1-3 tiers below “alignment geniuses”. A separate global talent search program could find them.
    Imaginary example: Crossover between the Atlas Fellowship, old CFAR, and MIRI. I imagine an organization that offers contests, workshops, and research fellowships in order to attract talented people around the world.
    Skills needed: Strong models of community-building, strong understanding of AI safety concepts, really good ways of evaluating who is promising, good models of downside risks when conducting broad outreach
    Olivia Jimenez and I are currently considering working on this. Please feel free to reach out if you have interest or advice.
    Training Program for AI Alignment researchers
    Purpose: Provide excellent training, support, internships, and mentorship for junior AI alignment researchers.
    How this reduces P(doom): Maybe there are people who would become extremely promising researchers if they were provided sufficient...
    13 min
  • EA - Requesting feedback: proposal for a new nonprofit with potential to become highly effective by Marshall
    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: Requesting feedback: proposal for a new nonprofit with potential to become highly effective, published by Marshall on September 13, 2022 on The Effective Altruism Forum.
    TL;DR
    I make the case for funding a new initiative that addresses an important, tractable, and neglected problem with potential to reach GiveWell-level cost-effectiveness. Please review this and share your thoughts!
    Millions of people die every year from preventable conditions, especially in low- and middle-income countries (LMICs).
    Some of these deaths can be avoided by improving health workers’ skills and knowledge.
    Working with a nonprofit, I’ve been piloting a way to train health workers online with a scrappy WordPress-based MVP, and still getting good results.
    I’m looking for funding to create and scale a solution that’s 10X better than anything that exists.
    Why I’m sharing this
    I’m pretty new to EA and I’d like to submit a version of this to EA organizations for funding. I’m seeking feedback to help make this as clear, logical, and succinct as possible.
    Importance: what is the scale of problems that can be addressed with health worker training?
    Every year, there are tens of millions of deaths that can be averted with simple interventions, including:
    2.9 million deaths of newborn babies readily preventable with standard care,
    1.5 million deaths caused by diarrheal diseases; many of which can be averted with improved sanitation practices and low-cost treatment,
    2.5 million deaths caused by lower respiratory infections; some of which can be prevented with appropriate diagnosis and treatment, and
    10.8 million deaths caused by high blood pressure, which is inexpensively controlled with generic medications.
    Pandemics are an equally important global problem. COVID-19 has caused an estimated 17.5 million deaths. Unfortunately, experts agree that COVID-19 isn’t a worst-case scenario: future pandemics could pose an existential threat to humanity.
    These problems and their causes have an uneven global distribution. Newborn deaths, lower respiratory infections, and diarrheal disease, in particular, are concentrated in LMICs. Moreover, for a variety of reasons, LMICs are particularly susceptible to epidemics. Solutions to these problems must be implemented in LMICs.
    Tractability: how can online health worker training solve these problems?
    There is ample evidence that health worker (HW) training can have benefits on its own or when combined with other interventions.
    My analysis of existing research studies shows that training HWs to properly care for newborn babies is likely to be highly cost-effective, with an average cost of $59 per DALY averted ($100 per DALY averted is sometimes cited as a benchmark for highly effective interventions). Please see the full forum post for the details of this analysis.
    Management of childhood illness training, which covers simple interventions (such as appropriately diagnosing and treating lower respiratory infections and diarrhea), measurably improves HW clinical skills.
    A recent study estimated that properly training HWs and giving them personal protective equipment would have saved $7.2 billion in just the first six months of the COVID-19 pandemic.
    The decisions that HWs make early in a disease outbreak can save countless lives. Notably, a doctor who acted quickly to quarantine a sick traveler is credited with stopping the spread of Ebola into Nigeria in 2014. HWs need to be well trained to be ready to make these critical decisions at any time.
    There is also emerging evidence that online learning is an effective way to train HWs.
    Online neonatal care training leads to significant knowledge and skills gains among HWs.
    The infection control training I’ve been working on has high completion rates, learning gains comparable to those seen in more resource-intensive in-person t...
    14 min
  • EA - EA & LW Forums Weekly Summary (5 - 11 Sep 22’) by Zoe Williams
    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: EA & LW Forums Weekly Summary (5 - 11 Sep 22’), published by Zoe Williams on September 12, 2022 on The Effective Altruism Forum.
    Supported by Rethink Priorities
    This is part of a weekly series - you can see the full collection here. The first post includes some details on purpose and methodology.
    If you'd like to receive these summaries via email, you can subscribe here.
    Top Readings / Curated
    A new section, designed for those without the time to read all the summaries. Everything here is also within the relevant sections later on so feel free to skip if you’re planning to read it all.Methodology - I’ve pulled out a few posts I think offer some particularly new and interesting ideas, or would be useful to have more people giving feedback on. This is purely based on my personal opinion. Keen on any thoughts in the comments if this section is useful or not!
    A California Effect for Artificial Intelligence
    By henryj
    The California Effect occurs when companies adhere to California regulations even outside California’s borders. The author evaluates three types of AI regulation California could adopt, and finds two 80% likely to produce a california effect - regulating training data through data privacy, and risk-based regulation like that proposed by the EU. A full research paper is linked, the results of a summer research fellowship.
    Developing Next Gen PPE to Reduce Biorisks
    by AndyGraham, tsmilton
    The authors are engineers planning to apply to LTTF for funding to run a feasibility study on improved PPE, particularly high protection versions (ie. suits). The current versions are bulky, expensive, hard to use, and have changed little since 1979. Feasibility will be primarily research to validate the need and potential impact on GCBRs. If that resolves favorably, they hope to finalize a product within 2-5 years.
    They’re looking for advice and thoughts - particularly from those with bioexpertise or PPE expertise, or who can help with the grant application.
    Cause Exploration Prizes: Announcing our prizes
    Over 150 submissions. Top prize to Organophosphate pesticides and other neurotoxicants by Ben Stewart. Second prize to each of Violence against women and girls by Akhil, Sickle Cell Disease (anonymous) and Shareholder activism by sbehmer. 20 honourable mentions. Post on learnings may come in the future.
    See section ‘Bonus: winners of Open Philanthropy’s cause exploration prizes’ section below for more detailed summaries.
    Announcing the Change Our Mind Contest for critiques of our cost-effectiveness analyses
    By Givewell
    Givewell is looking for critiques on their cost effectiveness analyses. Entries due 31st October, prizes up to 20K + a chance to improve allocation of millions of dollars.
    Monitoring for deceptive alignment
    by evhub
    Requests DeepMind, OpenAI, and Anthropic to actively monitor and run experiments on narrow deceptive alignment (ie. where a model looks aligned only because it’s trying to, for some ulterior motive). Early examples may be relatively easy to detect (eg. because early AIs are bad at it, or defect quickly) and therefore study. This could include monitoring for pre-cursors like when an AI first develops an instrumental goal not to have itself shut down.
    Concrete things to test models on include if behavior changes with / without oversight, or catching deception at source via interpretability / transparency tools.
    From twitter: Trials show malaria vaccine gives 80% protection (3 doses + yearly booster). It’s cheap, they have a deal to manufacture 100 million doses a year, and hope to roll it out next year. (link)
    EA Forum
    Philosophy and Methodologies
    An entire category of risks is undervalued by EA [Summary of previous forum post]
    By Richard Ren
    The cascading sociopolitical & economic effects of climate change, pandemics, and conflicts are undervalued i...
    28 min
  • AF - New tool for exploring EA Forum, LessWrong and Alignment Forum - 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, LessWrong and Alignment Forum - Tree of Tags, published by Filip Sondej on September 13, 2022 on The AI Alignment 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.
    Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org.
    3 min
  • EA - EA architect: Building an EA hub in the Great Lakes region by Tereza Flidrova
    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: EA architect: Building an EA hub in the Great Lakes region, published by Tereza Flidrova on September 13, 2022 on The Effective Altruism Forum.
    EA is growing, and there is a great need for a variety of spaces for the community to live, work and socialize. Traditionally, EA communities formed around hubs like Boston, the Bay Area or London. But, I feel like recently, there has been a growing demand for diversification of EA hubs around the world. This article aims to introduce the Great Lakes region as a potential place for a new EA hub. I will argue that a new hub could be set up in the green and walkable town of Sandusky, located right on Lake Erie, allowing EAs to benefit from the existing thriving community, facilities and infrastructure.
    My involvement in the project started a few weeks ago, when I spent a week in Sandusky, Ohio, visiting Andy Weber (a great 80k episode with Andy here), a former Assistant Secretary of Defense for Nuclear, Chemical & Biological Defense Programs, and his wife Christine Parthemore, Chief Executive Officer of Council on Strategic Risks. I was on a mission: to explore Sandusky and help figure out whether it would be a suitable location for an EA hub. Those of you living a nomadic lifestyle, working remotely or seeking to improve your mental health, productivity and wellness, stay tuned: Sandusky might be a place for you!
    So why the Great Lakes region, and why Sandusky in particular?
    In Andy's words: 'Sandusky is the remote worker's paradise', and after visiting, I have to say I wouldn't disagree! The following paragraphs describe the main reasons why Sandusky is a great place for remote work or perhaps establishing an EA hub.
    In Sandusky, you get more for less.
    Sandusky is much cheaper than the usual places where EAs tend to gather. According to Redfin (Redfin n.d.), the median price of Sandusky home prices for July 2022 was $200K, as compared to $825K for NYC and $1,460K for SF. In times of remote working, places like Sandusky might offer EAs a chance to work from one place, offering an alternative to living in the most expensive places in the world. Living costs as well as costs associated with running an EA hub in Sandusky are much more favorable and would buy EAs better housing and healthier work environments for far less money.
    Life at the water.
    Living near the water, similarly to living surrounded by greenery, seems to be associated with many positive measures of physical and mental wellbeing (Hunt 2019).
    Sandusky is located at Lake Erie, a freshwater reservoir offering life right at the waterfront.
    It's one of the cities of the Green Cities Initiative, an initiative of cities by the Great Lakes that strive to offer a place to do remote work in times of global warming and covid. The initiative wants to offer life at a freshwater resource (the Great Lakes provide 84% of North America’s surface freshwater (“Green Cities Initiative” n.d.)), with access to nature. I like the idea from an urban design perspective: it brings cities from the region together to work collaboratively on developing the region as a whole. For a smaller town like Sandusky, this means better connectivity to the communities and facilities surrounding it.
    Access to nature.
    People living near green spaces seem to be less likely to have anxiety and depression and are more likely to be physically active (Jennings, Johnson Gaither, and Gragg 2012). Many of us choose to live in cities despite the distance from nature because of all their benefits, like access to jobs, opportunities, and people. However, living in loud, densely populated and congested urban areas can get overwhelming and stressful and can negatively impact our mental health or sleep quality. Sandusky could offer an escape with more immediate access to nature. It is surrounded by parks, and the Magee...
    11 min
  • AF - The interpretable composition hypothesis: mundane lessons from golang by Luke H Miles
    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 interpretable composition hypothesis: mundane lessons from golang, published by Luke H Miles on September 13, 2022 on The AI Alignment Forum.
    An AGI will probably rely on correctly-implemented multithreaded software and networking code.
    Go has some lessons in this exact domain of course but I claim there are broader implications.
    How much better is go?
    Most smart contracts are written in macro-crusted rust or in solidity. Most major smart contracts have been hacked. Most blockchain validators were written in go and have not been hacked.
    The validators are implementing objectively more complex and difficult-to-secure algorithms than the smart contracts in most cases. Simple deposit + releases contracts have had losses in the hundreds of millions of dollars.
    Most JavaScript and Python server libraries have had major vulnerabilities. Most go server libs have not.
    Go has the advantage of being created twenty years later than Python and Js but much of its innovation is just the removal of confusing stuff. Well-understood primitives and a small feature set go a long way. Often far enough to be the difference between hacked and not hacked.
    Go does not have compile time array bounds checking or verification of input & output constraints or anything else sophisticated like this. People simply write code line by line and check it line by line.
    If you have a small primitive set and careful review then can you build towers that do not fall? Is that alone enough? It seemed to work in math.
    Interpretable composition hypothesis: weak form
    Extremely complex systems can consistently be controlled, reliable, and interpretable if they are hand-made of 100% understood parts.
    Breaking this down:
    (This is not completely crisp; it is a claim about what humans can achieve under good circumstances; it cannot be outright proven or disproven. I expect strong evidence towards one side or the other to be relatively easy to find, to the point of making it practically falsifiable/provable.)
    A part being 99% understood is inadequate because adversaries (or gradient descent) can exploit footguns.
    eg most js developers mostly understand js prototyping, but some people missed the __proto__ secret feature and it caused countless vulnerabilities.
    The parts must be understood by the composers themselves. It is not adequate for someone somewhere to understand the parts.
    (Related thought: Interpretability work will likely achieve little if it is not strongly used by the AI builder or not fully understood in its use.)
    By "can consistently be" I mean that there exists some structure of people and method of work or review such that they produce a reliable and understood product for almost any task assigned.
    The parts must be put together by hand! If you do an evolutionary algorithm on a million transistors then who knows what you'll get. But transistors were pretty much perfectly understood and we got pretty much perfect adders etc with them.
    (I cannot find John Wentworth's post where he gives the trillion transistors example.)
    Observe that we do have a few samples of complex code in the wild that has basically never fallen down despite harsh or adversarial conditions, such as Voyager or the Bitcoin validators.
    Interpretable composition hypothesis: strong form
    General optimizers can consistently be controlled, reliable, and interpretable across environments if they are hand-made of 100% understood parts, even as they gain superhuman capabilities, and even if they are very complex.
    (Again, "can consistently be" means roughly "there exists a process that humans can follow without fail". Think of the confidence we have in a mathematical proof that has been checked by a handful of qualified people.)
    A standard game tree search algorithm will solve tic tac toe and will do nothing surprising if you give it a bi...
    8 min
  • EA - CEA Ops is now EV Ops by EV Ops
    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: CEA Ops is now EV Ops, published by EV Ops on September 13, 2022 on The Effective Altruism Forum.
    Effective Ventures (EV) is a federation of organisations and projects working to have a large positive impact in the world. EV was previously known as the Centre for Effective Altruism but the board decided to change the name to avoid confusion with the organisation within EV that goes by the same name.
    EV Operations (EV Ops) provides operational support and infrastructure that allows effective organisations to thrive.
    Summary
    EV Ops is a passionate and driven group of operations specialists who want to use our skills to do the most good in the world.
    You can read more about us at.
    What does EV Ops look like?
    EV Ops began as a two-person operations team at CEA. We soon began providing operational support for 80,000 Hours, EA Funds, the Forethought Foundation, and Giving What We Can. And eventually, we started supporting newer, smaller projects alongside these, too.
    As the team expanded and the scope of these efforts increased, it made less sense to remain a part of CEA. So at the end of last year, we spun out as a relatively independent organisation, known variously as “Ops”, “the Operations Team”, and “the CEA Operations team”.
    For the last nine months or so, we’ve been focused on expanding our capacity so that we can support even more high-impact organisations, including the GovAI, Longview Philanthropy, Asterisk, and Non-trivial. We now think that we have a comparative advantage in supporting and growing high-impact projects — and are happy that this new name, “Effective Ventures Operations”' or “EV Ops”, accords with this.
    EV Ops is arranged into the following six teams:
    The organisations EV Ops supports
    We now support and fiscally sponsor several organisations (learn more on our website). Alongside these we support a handful of Special Projects: smaller, 1-2 person, early-stage projects which may grow into independent organisations of their own.
    We’re keen to support a wide range of projects looking to do good in the world, although we’re close to current capacity. To see if we could help your project grow and develop, visit or complete the expression of interest form.
    Get involved
    We’re currently hiring for the following positions:
    Project Manager for Oxford EA hub
    Senior Bookkeeper / Accountant
    Operations Associate
    Executive Assistant for the Property team
    Operations Associate - Salesforce Admin
    Finance Associate
    If you’re interested in joining our team, visit.
    If you have any questions about EV or EV Ops, just drop a comment below. Thanks for reading!
    Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org.
    3 min
  • LW - Argument against 20% GDP growth from AI within 10 years [Linkpost] by aogara
    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: Argument against 20% GDP growth from AI within 10 years [Linkpost], published by aogara on September 12, 2022 on LessWrong.
    Mohammed Bavarian, a research scientist at OpenAI, tweeted this thread arguing that he could see "the overall US GDP growth rising from recent avg 2-3% to 20+% in 10 years." Feel free to check out those arguments, though they'll probably be familiar to you: GPT-3, GitHub Copilot, and image synthesis will drive unprecedented improvements.
    Cameron Fen, an economics PhD student at the University of Michigan, responded with this thread disagreeing with Bavarian's argument. I wanted to share some of the arguments that I found novel and interesting.
    Argument #1: The impacts of previous transformative technologies
    There have been 3 industrial revolutions in history, mechanization, electricity and mass production, and IT and the internet. China went through all 3 at the same time and was barely able to go above 7% growth annually.
    The newest industrialization will be big, but will it be bigger than moving 95% of the population working in agriculture to working factories, not to mention all 3 combined?
    In particular, the growth of the first industrial revolution accelerated growth from 1.5% to 3% a year in the UK (Source). Growth from electrification was 1.5% a year on average (source) and growth during the information industrialization was 3.5% a year (source).
    Given these growth rates, it seems unlikely that a single industrial revolution can move the needle to such an extent that US GDP growth accelerates from 2% to 7%.
    Argument #2: The size of the tech industry
    Perhaps you can argue that the industrial revolution on the horizon is going to be 3x bigger than any industrial revolution in the past. Let’s see what that would imply:
    7% growth implies 5% growth over our current trend rate of about 2%. This comes out to be about 1.15 trillion GDP additional growth this year and increasing as the base gets larger.
    According to this article, Facebook contributes 100 billion to US economic activity. While much of this is uncounted because GDP methodology is imperfect, I’ll go with that number.
    To get to the 1.15 trillion additional growth you need 11.5 Facebook created every year. This comes out to a market cap of 5.5 trillion dollars created every year in tech to get something like a 5% growth. The market cap of the entire tech sector is 13.5 trillion dollars.
    Call adjacent markets another 6.5 trillion dollars. Thus a 20 trillion dollar market cap sector needs to add 6 trillion (5.5 + .0220) dollars a year. Has the tech sector ever grown so fast that it is creating the equivalent of 12.5 Facebooks from nothing every year? No.
    Let’s think about this another way, assume the US is growing at 2% a year. The tech sector is 10% of the US economy. If the US were to grow 7% a year on all tech sector growth, the tech sector would have to grow at 52% a year.
    Even if you assume adjacent sectors (20% share) growing at 12% a year, you need 32% growth from the tech sector to just get 7% total GDP growth. This seems infeasible to me.
    Argument #3: Estimating the market impact of LLMs, image synthesis, and AlphaFold
    Just to get a sense of how massive innovation has to be to move the needle, I'm going to discuss three central innovations. Let me know if I missed something--more will be invented--but I plan on showing that these game changing techs will not drastically improve growth.
    The triumvirate of GDP impactful techs are 1) GPT-3 and other LLMs, 2) Self-driving cars and other robotic control, 3) and Alpha Fold. I don’t include text-to-image models like DALLE and other diffusion models because idk any commercial applications that move the needle.
    Market impact of natural language generation: 0.5% of GDP per year because there are free alternatives. (My response: Wouldn't t...
    9 min