The Nonlinear Library

The Nonlinear Library

By The Nonlinear FundEducation
Download on the App Store

All The Nonlinear Library episodes

  • AF - Thoughts on sharing information about language model capabilities by Paul Christiano
    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: Thoughts on sharing information about language model capabilities, published by Paul Christiano on July 31, 2023 on The AI Alignment Forum.
    Core claim
    I believe that sharing information about the capabilities and limits of existing ML systems, and especially language model agents, significantly reduces risks from powerful AI - despite the fact that such information may increase the amount or quality of investment in ML generally (or in LM agents in particular).
    Concretely, I mean to include information like: tasks and evaluation frameworks for LM agents, the results of evaluations of particular agents, discussions of the qualitative strengths and weaknesses of agents, and information about agent design that may represent small improvements over the state of the art (insofar as that information is hard to decouple from evaluation results).
    Context
    ARC Evals currently focuses on evaluating the capabilities and limitations of existing ML systems, with an aim towards understanding whether or when they may be capable enough to pose catastrophic risks. Current evaluations are particularly focused on monitoring progress in language model agents.
    I believe that sharing this kind of information significantly improves society's ability to handle risks from AI, and so I am encouraging the team to share more information. However this issue is certainly not straightforward, and in some places (particularly in the EA community where this post is being shared) I believe my position is controversial.
    I'm writing this post at the request of the Evals team to lay out my views publicly. I am speaking only for myself. I believe the team is broadly sympathetic to my position, but would prefer to see a broader and more thorough discussion about this question.
    I do not think this post presents a complete or convincing argument for my beliefs. The purpose is mostly to outline and explain the basic view, at a similar level of clarity and thoroughness to the arguments against sharing information (which have mostly not been laid out explicitly).
    Accelerating LM agents seems neutral (or maybe positive)
    I believe that accelerating LM agents increases safety through two channels:
    LM agents are an unusually safe way to build powerful AI systems. Existing concerns about AI takeover are driven primarily by scaling up black-box optimization, and increasing our reliance on human-comprehensible decompositions with legible interfaces seems like it significantly improves safety. I think this is a large effect size and is suggested by several independent lines of reasoning.
    To the extent that existing LM agents are weak it is due to exceptionally low investment. This is creating "dry tinder:" as incentives rise that investment will quickly rise and so low-hanging fruit will be picked. While there is some dependence on serial time, I think that increased LM investment now will significantly slow down progress later.
    I will discuss these mechanisms in more detail in the rest of this section.
    I also think that accelerating LM agents will drive investment in improving and deploying ML systems, and so can reduce time available to react to risk. As a result I'm ambivalent about the net effect of improving the design of LM agents - my personal tentative guess is that it's positive, but I would be hesitant about deliberately accelerating LM agents to improve safety (moreover I think this would be a very unleveraged approach to improving safety and would strongly discourage anyone from pursuing it).
    But this means that I am significantly less concerned about information about LM agent capabilities accelerating progress on LM agents. Given that I am already positively disposed towards sharing information about ML capabilities and limitations despite the risk of acceleration, I am particularly positive about sharing...
    21 min
  • EA - Thoughts on far-UVC after working in the field for 8 months by Max Görlitz
    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: Thoughts on far-UVC after working in the field for 8 months, published by Max Görlitz on July 31, 2023 on The Effective Altruism Forum.
    Views expressed in this article are my own and do not necessarily reflect those of my employer SecureBio.
    Summary
    Far-UVC has great promise, but a lot of work still needs to be done
    There still are many important open research questions that need to be answered before the technology can become widely adopted
    Right now, a key priority is to grow the research field and improve coordination
    The main reason far-UVC is so promising is that widespread installation could passively suppress future pandemics before we even learn that an outbreak has occurred
    Higher doses mean more rapid inactivation of airborne pathogens but also more risk for harm to skin, eyes, and through indoor air chemistry. Therefore, the important question in safety is, "How high can far-UVC doses go while maintaining a reasonable risk profile?"
    Existing evidence for skin safety within current exposure guidelines seems pretty robust, and I expect that skin safety won't be the bottleneck for far-UVC deployment at higher doses.
    Current evidence around eye safety is much more sparse than for skin safety. Eye safety seems like it could be the bottleneck to what doses of far-UVC can be reasonably used.
    Undoubtedly, far-UVC has a substantial impact on indoor air chemistry by producing ozone, which oxidizes volatile organic compounds in the air that can result in harmful products such as particulate matter.
    Little research has been done on methods to mitigate this issue.
    This might turn out to be a bottleneck to what doses of far-UVC can be reasonably used, but I am really uncertain here.
    There is no doubt that far-UVC can dramatically reduce the amount of airborne pathogens within a room (inactivation of ~98% of aerosolized bacteria within 5 minutes). Crucially, we don't know how well this translates into an actual reduction in the total number of infections.
    Very few people have thought about how the adoption of far-UVC could be driven and what a widespread deployment of the technology could look like
    So far, there is little to no regulation of far-UVC.
    In the US, (potential) regulation of far-UVC seems quite messy, as no authority has clear jurisdiction over it.
    Introduction
    Far-UVC (200-235 nm) has received quite a bit of attention in EA-adjacent biosecurity circles as a technology to reduce indoor airborne disease spread and is often discussed in the context of indoor air quality (IAQ). Notably, Will MacAskill mentioned it often throughout various media appearances in 2022.
    I have been working on research around far-UVC for the past 8 months. More specifically, we wrote an extensive literature review on skin and eye safety (submitted & soon™ to be published as an academic paper). We also coordinated with many researchers in the field to lay out a plan for the studies that still need to be done to get a more comprehensive understanding of the technology's safety & efficacy.
    Although far-UVC has been discussed on the forum, the existing information is relatively shallow, and most in-depth knowledge is either buried in technical research papers or not publicly available since a lot of intricacies are mostly discussed informally within the research community.
    In this post, I will first offer high-level thoughts and then go over different categories of information around far-UVC (safety, efficacy, indoor air chemistry, adoption, and regulation) to provide my current perspectives & takes. Please note that I am much more familiar with safety aspects than with the other categories. Also, this is not a general overview of far-UVC, what it is, and how it works. For a relatively recent and comprehensive introduction, I recommend "Far UV-C radiation: An emerging tool for pandemic co...
    21 min
  • AF - Open Problems and Fundamental Limitations of RLHF by Stephen Casper
    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: Open Problems and Fundamental Limitations of RLHF, published by Stephen Casper on July 31, 2023 on The AI Alignment Forum.
    Reinforcement Learning from Human Feedback (RLHF) has emerged as the central alignment technique used to finetune state-of-the-art AI systems such as GPT-4, Claude, Bard, and Llama-2. Given RLHF's status as the default industry alignment technique, we should carefully evaluate its shortcomings. However, there is little public work formally systematizing problems with it.
    In a new survey of over 250 papers, we review open challenges and fundamental limitations with RLHF with a focus on applications in large language models.
    Abstract
    Reinforcement learning from human feedback (RLHF) is a technique for training AI systems to align with human goals. RLHF has emerged as the central method used to finetune state-of-the-art large language models (LLMs). Despite this popularity, there has been relatively little public work systematizing the challenges it poses. In this paper, we (1) survey concrete challenges and open questions with RLHF; (2) overview techniques to better understand, improve, and complement it in practice; and (3) discuss how standards for transparency and auditing could help to reduce risks. We emphasize the importance of recognizing the limitations of RLHF and integrating it into a more complete framework for developing safer AI.
    Contributions
    Concrete challenges with RLHF: We taxonomize and survey problems with RLHF, dividing them into three primary categories: challenges with feedback, challenges with the reward model, and challenges with the policy. We also distinguish between challenges that are relatively tractable versus ones that are more fundamental limitations of alignment with RLHF.
    Incorporating RLHF into a broader technical safety framework: We discuss how RLHF is not a complete framework for developing safe AI and highlight additional approaches that can help to better understand, improve, and complement it.
    Governance and transparency: We consider the challenge of improving industry norms and regulations affecting models trained with RLHF. Specifically, we discuss how the disclosure of certain details by companies using RLHF to train AI systems can improve accountability and auditing.
    Transparency
    We argue that a sustained commitment to transparency (e.g. to auditors) would make the RLHF research environment more robust from a safety standpoint. First, the disclosure of some details behind large RLHF training runs would clarify a given organization's norms for model scrutiny and safety checks. Second, increased transparency about efforts to mitigate risks would improve safety incentives and suggest methods for external stakeholders to hold companies accountable. Third, transparency would improve the AI safety community's understanding of RLHF and support the ability to track technical progress on its challenges.
    Specific policy prescriptions are beyond the paper's scope, but we discuss specific types of details that, if disclosed, could be indicative of risks and should be accounted for when auditing AI systems developed using RLHF.
    RLHF = Rehashing Lessons from Historical Failures?
    RLHF offers new capabilities but faces many old problems. Researchers in the safety, ethics, and human-computer interaction fields have been demonstrating technical and fundamental challenges with the system and its components for years. In 2023, Paul Christiano (the first author of the 2017 paper, Christiano et al. (2017), prototyping RLHF) described it as a "basic solution" intended to make it easier to "productively work on more challenging alignment problems" such as debate, recursive reward modeling, etc.
    Instead of being used as a stepping stone toward more robust alignment techniques, RLHF seems to have also undergone a sort of "capabil...
    5 min
  • LW - Is Light Drinking Protective? 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: Is Light Drinking Protective?, published by jefftk on July 31, 2023 on LessWrong.
    There are a lot of claims about how alcohol affects the body, and some sort of "heavy drinking is bad for you but light or moderate drinking is better than no drinking" is a common one. I've not paid a lot of attention to these, however, since non-drinkers as a group include a bunch of people who've given up alcohol due to health-related issues. I was interested, however, to see a study ( Tian et al. 2023) that compares light and moderate drinkers to people who haven't
    ever been drinkers. Unfortunately, after getting into the study I don't think it tells us much and I haven't updated my views here.
    The study finds:
    Compared with lifetime abstainers, current infrequent, light, or moderate drinkers were at a lower risk of mortality from all causes, CVD, chronic lower respiratory tract diseases, Alzheimer's disease, and influenza and pneumonia. Also, light or moderate drinkers were associated with lower risk of mortality from diabetes mellitus and nephritis, nephrotic syndrome, or nephrosis.
    To get this association they analyzed data from the NHIS survey, which includes questions on drinking habits, and the linked death records.
    As you might expect, lifetime abstainers look quite different from light and moderate drinkers. For example, 26% of lifetime abstainers had less than a high school education, while only 10% of light or moderate drinkers did. The lifetime abstainer group is also a lot older (22% 65+ vs 10% or 13%), more female (66% vs 50% or 27%), more
    Black (17% vs 9% or 8%), more Hispanic (19% vs 12% or 10%), more nonsmoker (82% vs 56% or 44%), and more physically inactive (69% vs
    48% or 43%). See the full table. Visually:
    These are really big!
    Now, they did adjust for these, along with many other differences between these groups:
    The above-mentioned associations were investigated by adjusting for the following covariates: age, sex, race, or ethnicity (model 1); model 1 plus education level, physical activity, body mass index, smoking status, hypertension, heart disease, stroke, cancer, diabetes, asthma, emphysema, or chronic bronchitis in a separate model (model
    2).
    The problem is, there could easily be other important confounders. If something they are not adjusting for leads to both abstinence from alcohol and higher mortality, they aren't able to distinguish that from the causal effect of alcohol on mortality.
    They do explicitly list this concern, the first one in their list of limitations, and it's always something I'm worried about with correlational studies. But this is worse than usual: the known differences between these groups going into the study are so large that I am very skeptical that adjusting for these is enough to cover the unknown differences. So I haven't updated my views appreciably from this study.
    (I don't drink, but not for health reasons.)
    Comment via: facebook, mastodon
    Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org
    4 min
  • LW - Apollo Neuro Results 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: Apollo Neuro Results, published by Elizabeth on July 30, 2023 on LessWrong.
    Introduction
    Two months ago I recommended the Apollo Neuro for sleep/anxiety/emotional regulation. A number of people purchased it based on my recommendation- at least 25, according to my referral bonuses. Last week I asked people to fill out a form on their experience.
    Take-home messages:
    If you are similar to people who responded to my first post on the Apollo, there's a ~4% chance you end up getting a solid benefit from the Apollo.
    The chance of success goes up if you use it multiple hours per day for 4 weeks without seeing evidence of it working, but unless you're very motivated you're not going to do that.
    The long tail of upside is very, very high; I value the Apollo Neuro more than my antidepressant. But you probably won't.
    There's a ~10% chance the Apollo is actively unpleasant for you; however no one reported cumulative bad effects, only one-time unpleasantness that stopped as soon as they stopped using it.
    With Numbers
    The following graphs include only people who found the Apollo and the form via my recommendation post. It does not include myself or the superresponders who recommended it to me.
    (that's one person reporting it definitely helped)
    An additional six people filled out an earlier version of the form, none of whom found it helpful, bringing the total to 24 people.
    Obviously I was hoping for a higher success rate. OTOH, the effects are supposed to be cumulative and most people gave up quickly (I base this on conversations with a few people, there wasn't a question for it on the form). Some of that is because using the Apollo wasn't rewarding, and I'll bet a lot of the problem stems from the already pretty mediocre app getting an update to be actively antagonistic. It probably is just too much work to use it long enough to see results, unless you are desperate or a super responder.
    Of people who weren't using it regularly: 55% returned it, 20% failed to return it, and the remaining 35% chose to keep it. I think that last group is probably making a mistake; the costs of luck-based medicine add up, so if you're going to be a serious practitioner you need to get good at cutting your losses. It's not just about the money, but the space and mental attention.
    Of 6 people in the earlier version of the form, 1-2 found it actively unpleasant.
    The downside turned out to be worse than I pictured. I'm fond of saying "anything with a real effect can hurt you", but I really couldn't imagine how that would happen in this case. The answer is: nightmares and disrupted sleep. In both cases I know of they only experienced this once and declined to test it again, so it could be bad luck, but I can't blame them for not collecting more data. No one reported any ill effects after they stopped using it.
    I would also like to retract my previous description of the Apollo return policy as "good". You do get most of your money back, but a 30-day window for a device you're supposed to test for 28 days before passing judgment is brutal.
    It's surprisingly hard for me to find referral numbers, but I know I spurred at least 25 purchases, and almost certainly less than 30. That implies an 80% response rate to my survey, which is phenomenal. It would still be phenomenal even if I'd missed half the purchasers and it was only a 40% response rate. Thanks guys.
    Life as a superresponder
    Meanwhile, the Apollo has only gotten better for me. I've basically stopped needing naps unless something obvious goes wrong, my happiness has gone up 2 points on a 10 point scale (probably because of the higher quality sleep)1, sometimes my body just feels good in a way it never has before. I stress-tested the Apollo recently with a very grueling temp gig (the first time in 9 years I've regularly used a morning alarm. And longer h...
    5 min
  • LW - How to make real-money prediction markets on arbitrary topics by yutaka
    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 to make real-money prediction markets on arbitrary topics, published by yutaka on July 30, 2023 on LessWrong.
    Intro
    On current real-money event market platforms, traders are limited to viewing and betting on markets made by platform operators. This means the questions tend to be confined to popular subjects like politics and sporting outcomes, and the sites can't be used for financing estimates on practical or limited-interest questions, like alignment expectations or CEO performance.
    OpenPredict is a toolkit for managing prediction markets which regular people can use for this purpose. We're calling it a "toolkit" instead of a "website" because OpenPredict's developers do not act as either arbitrators or market makers; instead, traders set criteria and manage resolution themselves. Likewise, all of OpenPredict's components, including its web client, are entirely open-source and store important metadata about markets on Solana and IPFS. This means that while you are free to use the public instance at openpredict.org it is an unprivileged, replaceable frontend that you may (and are in fact encouraged to) forgo in favor of a self-hosted client.
    The OpenPredict network uses cryptocurrency to settle accounts, but even if you don't have a crypto wallet, creating a new prediction market on OpenPredict takes less than five minutes. As a tutorial, I'll be making a market about the outcome of this $150,000 bet between EliezerYudkowsky and the UFO guy. You can view and trade outcomes on the resulting market here.
    Market Creation
    Step One: Sign up
    On OpenPredict, your identity is synonymous with your cryptocurrency wallet. If you have one, you can login with that. If you don't have one, all you need is an email address and the site will provision one for you. Simply enter it into the signup box in the top right and follow the shown instructions:
    Once you've done this, you'll see two wallet addresses generated automatically on the right. The one marked "SOL" is your user ID and what will initially be displayed when you comment, make markets, or trade. When you fill your wallet with some money, you'll be able to register a username that will be shown instead of the wallet address.
    Step Two: Fill your Solana Wallet
    You need one of two things to trade on OpenPredict: SOL or USDC. SOL is the network currency of Solana and is what's used to pay transaction fees - generally under a tenth of a cent per transaction. USDC is a "stablecoin", which means it's a cryptocurrency that stays 1:1 with the U.S. dollar, and is what you actually use to trade. When your wallet has enough of one but not the other, OpenPredict will automatically convert between the two at the best available price.
    The site will prompt you to buy SOL with normal credit or debit card on changenow.io here. Simply copy your wallet address and put it in the box; Solana transactions confirm in a couple seconds:
    After you have your SOL, you may prefer to convert the majority of it to USDC, to prevent yourself from being subject to its price movements. There's a button in that same sidebar you got your address from which you can use to do this; just make sure to keep a small amount of SOL in your account so that you can make trades. 50 cents (0.1 SOL at the time of writing) should suffice indefinitely.
    Step Three: Make and describe your market
    Click the "Create Market" button and follow the instructions. Be as unambiguous about resolution criteria as you can; you will not be able to edit either the title or description after you post the market! In my market description and title I defer to the two parties involved in the bet:
    Besides resolution criteria, before you create a market, you'll have to decide on a subsidy. The subsidy is basically the amount of money you're artificially providing as an incentive for good predi...
    6 min
  • EA - About 'subjective' wellbeing and cost-effectiveness analysis in mental health by LondonGal
    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: About 'subjective' wellbeing and cost-effectiveness analysis in mental health, published by LondonGal on July 30, 2023 on The Effective Altruism Forum.
    Hello everyone,
    I was first 'sucked in' to this forum when I was directed to a post I might find interesting - it was about a research organisation with EA endorsement that was straying into my area of work, mental health. I'm a UK doctor specialising in psychiatry, with some research experience. To be honest, I was baffled and a little frustrated by how far this organisation strayed from what I would expect from mental health research - hence the (perhaps overly) technical diatribe I launched into on a website I hadn't visited before, about an organisation I hadn't heard of prior.
    However, that's not usually my style, and once I took a step back from my knee-jerk reaction, I wanted to understand how people with the same goals could arrive at completely different conclusions. It's led me to do a lot of reading, and I wanted to see if I could try on a makeshift 'EA' hat, with most of my philosophy knowledge gained from The Good Place, no economics experience, and see where it went.
    What I wanted to understand:
    Where has the interest in 'wellbeing' arisen from, and what does it mean?
    What are 'subjective wellbeing' (SWB) measures, and are they useful?
    Are we at a point of putting monetary value on SWB (e.g. like QALYs) for the sake of cost-effectiveness analysis (CEA)?
    When people are in this space talking about mental health, are we talking the same language?
    Why are RCTs the 'best' evidence for subjective wellbeing?
    What would I come up with from my perspective of working within mental health for a way of comparing different interventions based on their intended effects on wellbeing? a. Spillover effects b. Catastrophic multipliers
    How does my guess stack up against existing research into wellbeing?
    How could my framework be helpful in practice?
    What would I be suggesting as research areas for maximal gains in wellbeing from my biased perspective?
    I'm aware this might be well-trodden ground in EA, which would make me embarrassingly late to the party, and consequently a complete bore. To lay my cards firmly on the table, I did approach these questions from the perspective that mental health is desperately underfunded, I spend a lot of time with patients who are severely affected by mental illness and therefore I'm biased towards seeing 'wellbeing' as an opportunity to rebalance this scale and acknowledge the impact mental illnesses have on people. I also feel the term 'mental health' is used in a way which is often confusing and occasionally unhelpful or stigmatising.
    This is not meant as an attempt to further an argument against any person or organisation; it will also not be high in tech-speak as this was the first lesson I learnt very quickly on my journey - while jargon is a useful shorthand for talking with people in the same field, as an outsider it is exhausting. This post does not reflect the attitudes or opinions of anyone but me - this is my personal quest for common ground and understanding, not a representation of 'UK psychiatry' - I'm speaking in an entirely personal capacity and, accordingly, I'm assuming I've gotten a lot of it completely wrong.
    To make this less self-indulgent, I've arranged this post to follow that question-and-answer format. For the sake of transparency, this was how this work came to be: I started with a long piece of writing about my concerns with assumptions made about mental health interventions in low- or middle-income country (LMIC) settings. I then did a quick Google on the WELLBY and wrote a lot about the idea of asking people to rate their 'satisfaction with life' on a scale from 0-10 which was essentially just entirely critical. I subsequently wrote out my concept of wellbei...
    1 hr 6 min
  • EA - Shutting down AI Safety Support by JJ Hepburn
    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: Shutting down AI Safety Support, published by JJ Hepburn on July 30, 2023 on The Effective Altruism Forum.
    This is a quick post to officially announce that AI Safety Support is on an indefinite pause. We have been on pause for a while now, but I wanted to make it official. I might have more to say on this later but expect that if I wait till I write a longer explanation, it will not happen.
    I'm calling it a pause because maybe AI Safety Support does reopen in some capacity in the future, possibly with a different person running it.
    A lack of funding is a part of the decision but is not the only factor. That is to say that I'm not looking for offers of funding.
    AI Safety Support Ltd, the Australian-based charity, supports several other projects and will continue to do so. None of these projects should be affected by this.
    If you have any questions, you can ask them in the comments here, but I may not respond to them.
    It has been a wild ride these last few years. I have so much love and appreciation for everyone I have worked with.
    Thank you all so much.
    Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org
    2 min
  • EA - The harm cascade: why helping others is so hard by Stijn
    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 harm cascade: why helping others is so hard, published by Stijn on July 29, 2023 on The Effective Altruism Forum.
    Altruism, helping others, is so difficult. Suppose you save a child's live, and many years later that child becomes a new Hitler, killing thousands of people. Of course this is very unlikely, because there are not many Hitlers in the world. And the child could also become a scientist who invents a new vaccine that saves thousands of people. So in expectation, saving a child is good. But, it's complicated, because of.
    The harm cascade
    What does it mean to help others? Simply put, it means choosing something that is desired, or the consequences of which are desired, by the individuals that are helped. The individuals have positive evaluations of the choice or the consequences of the choice. This means the individuals must have personal experiences and personal desires or preferences. In this sense, the others are persons or sentient beings (including almost all humans and many non-human animals). We cannot help non-sentient objects or things that have no personal experiences and desires. The opposite of helping is harming: doing something such that individuals have negative evaluations of the consequences.
    Now here is the real problem: almost everyone that we will help, is very likely to harm others as a consequence of the provided help. The more we help sentient beings, the more harm they will cause to others. If you help one individual, that individual is likely to cause harm to more than one other individual. Not helping that one individual will result in less harm caused to other individuals. So that means those other individuals are helped. But to make it more complex, when they are helped, those other individuals will cause harm to even more other individuals.
    Suppose each saved individual will - as a consequence of the help - kill two other individuals. Or vice versa: killing (harming) one individual will save (help) two other individuals. So if you help individual A, she will harm individuals B and C, such that these two individuals will no longer harm their victims, namely individuals D, E, F and G. The latter individuals are helped and will as a consequence harm individuals H, I, J, K, L, M, N and O, who will no longer harm individuals P, Q,. And so forth. This is the harm cascade.
    As a concrete example, consider saving a human child's life. That child will not become a new Hitler, but is most likely a meat eater: she will eat more than one animal during her life. And if you help a poor person raising his income, that person is likely to spend some of that extra income on meat, thereby increasing animal farming that causes harm to even more animals. This is the poor meat-eater problem. What if the humans that you helped become vegan and stop eating animal products? That will reduce animal farming and the use of agricultural land. Those vegans may have a preference for turning that unused farm land back into natural habitat. That will increase the population of wild animals. Those wild animals may be better-off than farmed animals, but many of those wild animals will cause harm to other wild animals.
    In nature we see again a harm cascade. One top-predator kills more than one meso-predators. Those harmed meso-predators can no longer kill their many prey. So due to the top-predator harming many animals, many other prey animals are saved. Those saved animals can cause harm to even other animals. Especially in aquatic ecosystems we see long food chains, where a top predator is like a serial killer who kills many other serial killers who kill even more serial killers, and so on.
    If animals are not killed by predators, they might increase in population and hence increase competition for food with other animals. They might use violence to attain food ...
    22 min
  • EA - Visit Mexico City in January & February to interact with the AI Futures Fellowship by AmAristizabal
    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: Visit Mexico City in January & February to interact with the AI Futures Fellowship, published by AmAristizabal on July 29, 2023 on The Effective Altruism Forum.
    TLDR
    Apply via this form if you would like to cowork with other AI researchers in Mexico City in January and/or February 2024 (we will check applications on a rolling basis).See also our general post announcing the AI Futures Fellowship.
    AI Futures Fellowship
    We are planning to open a coworking space in Mexico City in January and February 2024. This office space will be occupied by the fellows and staff from the ITAM AI Futures Fellowship and an engaging intellectual community of visitors and collaborators, and you could be one of them!
    Visitors
    The fellowship will take place in an office space located in La Condesa, a very green and calm area of Mexico City. The office faces Amsterdam street and Parque Mexico, which offer a fantastic mix of nature, cafes and restaurants.
    We are offering extra office spots for the duration of the program. This includes:
    Access to our coworking space for up to two months. Shorter visits are welcome, but longer visits are preferred.
    The default office policy is Hot-Desking. We might be able to accommodate more specific requests for individuals or organizations depending on availability, so let us know if there is a particular arrangement that would make you or your organization more likely to visit (e.g. "a private office for 5 people for x organization").
    We will provide lunch and some snacks throughout that time (Monday through Friday).
    The exact number of spots offered will vary depending on space availability and demand. We are also exploring the possibility of receiving more visitors who are willing to pay for their memberships.
    We are interested in hosting people working on technical AI safety research and AI governance. We will prioritize those who can have the most fruitful interactions with our fellows or those who would benefit the most from interacting with our community.
    If your organization would like to use the office space for a team retreat, do let us know as soon as possible.
    By default we won't cover accommodation or travel costs, but please let us know if this would significantly affect your chances of coming. We might be able to explore some options with you or point you to possible alternatives.
    Consider also joining us as a mentor
    The form also gives you the option of applying as a mentor. This might involve the following responsibilities:
    Help a fellow define a project and support them throughout its execution (This may be a project you've been interested in doing but haven't found the time to work on!)
    Availability to commit ~1 hour weekly for 8 weeks to support the fellow (this can include reviewing drafts and having regular calls).
    Assist the fellow in expanding their professional network, identifying additional resources, and valuable opportunities.
    Offer career guidance and motivation whenever needed. However, our staff will meet regularly with mentees to follow up on their time management, accountability, wellbeing and intrinsic motivation. You are expected to provide more object-level guidance.
    There is a compensation of $1000 USD for the valuable time and effort invested by mentors.
    If you want to join us as a visitor or mentor, fill out this brief form. If you're not sure yet but you're interested in exploring ways of interacting with our fellowship, also let us know!
    Feel free to contact us with any further questions you may have: [email protected].
    Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org
    4 min