The Nonlinear Library

The Nonlinear Library

By The Nonlinear FundEducation
Download on the App Store

All The Nonlinear Library episodes

  • EA - Announcing the 2023 CLR Summer Research Fellowship by stefan.torges
    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: Announcing the 2023 CLR Summer Research Fellowship, published by stefan.torges on March 17, 2023 on The Effective Altruism Forum.
    We, the Center on Long-Term Risk, are looking for Summer Research Fellows to help us explore strategies for reducing suffering in the long-term future (s-risk) and work on technical AI safety ideas related to that. For eight weeks, fellows will be part of our team while working on their own research project. During this time, they will be in regular contact with our researchers and other fellows. Each fellow will have one of our researchers as their guide and mentor.
    Deadline to apply: April 2, 2023. You can find more details on how to apply on our website.
    Purpose of the fellowship
    The purpose of the fellowship varies from fellow to fellow. In the past, have we often had the following types of people take part in the fellowship:
    People very early in their careers, e.g. in their undergraduate degree or even high school, who have a strong focus on s-risk and would like to learn more about research and test their fit.
    People seriously considering changing their career to s-risk research, who want to test their fit or seek employment at CLR.
    People with a strong focus on s-risk who aim for a research or research-adjacent career outside of CLR and who would like to gain a strong understanding of s-risk macrostrategy beforehand.
    People with a fair amount of research experience, e.g. from a partly- or fully completed PhD, whose research interests significantly overlap with CLR’s and who want to work on their research project in collaboration with CLR researchers for a few months. This includes people who do not strongly prioritize s-risk themselves.
    There might be many other good reasons for completing the fellowship. We encourage you to apply if you think you would benefit from the program, even if your reason is not listed above.
    What we look for in candidates
    We don’t require specific qualifications or experience for this role, but the following abilities and qualities are what we’re looking for in candidates. We encourage you to apply if you think you may be a good fit, even if you are unsure whether you meet some of the criteria.
    Curiosity and a drive to work on challenging and important problems;
    Ability to answer complex research questions related to the long-term future;
    Willingness to work in poorly-explored areas and to learn about new domains as needed;
    Independent thinking;
    A cautious approach to potential information hazards and other sensitive topics;
    Alignment with our mission or strong interest in one of our priority areas.
    Priority areas
    You can find an overview of our current priority areas here. However, If we believe that you can somehow advance high-quality research relevant to s-risks, we are interested in creating a position for you. If you see a way to contribute to our research agenda or have other ideas for reducing s-risks, please apply. We commonly tailor our positions to the strengths and interests of the applicants.
    Further details
    We encourage you to apply even if any of the below does not work for you. We are happy to be flexible for exceptional candidates, including when it comes to program length and compensation.
    Program dates: The default start date is July 3, 2023. Exceptions may be possible.
    Program length & work quota: The program is intended to last for eight weeks in a full-time capacity. Exceptions, including part-time work, may be possible.
    Location: We prefer summer research fellows to work from our London offices, but will also consider applications from people who are unable to relocate.
    Compensation: Unfortunately, we face a lot of funding uncertainty at the moment. So we don’t know yet how much we will be able to pay participating fellows. Compensation will range from £1,800 to £4,000 per month, de...
    5 min
  • EA - Legal Assistance for Victims of AI by bob
    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: Legal Assistance for Victims of AI, published by bob on March 17, 2023 on The Effective Altruism Forum.
    In the face of increasing competition, it seems unlikely that AI companies will ever take their foot off the gas. One avenue to slow AI development down is to make investment in AI less attractive. This could be done by increasing the legal risk associated with incorporating AI in products.
    My understanding of the law is limited, but the EU seems particularly friendly to this approach. The European Commission recently proposed the AI Liability Directive, which aims to make it easier to sue over AI products. In the US, companies are at the very least directly responsible for what their chatbots say, and it seems like it's only a matter of time until a chatbot genuinely harms a user, either by gaslighting or by abusive behavior.
    A charity could provide legal assistance to victims of AI, similar to how EFF provides legal assistance for cases related to Internet freedom.
    Besides helping the affected person, this would hopefully:
    Signal to organizations that giving users access to AI is risky business
    Scare away new players in the market
    Scare away investors
    Give the AI company in question a bad rep, and sway the public opinion against AI companies in general
    Limit the ventures large organizations would be willing to jump into
    Spark policy discussions (e.g. about limiting minor access to chatbots, which would also limit profits)
    All of these things would make AI a worse investment, AI companies a less attractive place to work, etc. I'm not sure it'll make a big difference, but I don't think it's less likely to move the needle than academic work on AI safety.
    Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org.
    2 min
  • LW - The algorithm isn't doing X, it's just doing Y. by Cleo Nardo
    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 algorithm isn't doing X, it's just doing Y., published by Cleo Nardo on March 16, 2023 on LessWrong.
    Introduction
    Mutual reduction implies equivalence
    Here's my most load-bearing intuition
    If two tasks reduce to one another, then it is meaningless to ask if a machine is 'really doing' one task versus the other.
    Moreover
    This intuition grounds my perspective on intelligence, AI, alignment, philosophy, etc.
    This intuition is load-bearing for other people who share my views.
    This intuition is a crux for much of the disagreement we have with other people.
    In this article, I'll formalise this intuition in two ways, computational and physical.
    Motivation
    People often say "the algorithm isn't doing X, it's just doing Y".
    X is normally some impressive high-level human-y thing, such as
    writing poetry
    causal reasoning
    recognising emotions
    interpreting art
    writing music
    making ethical decisions
    planning actions
    telling jokes
    understanding concepts
    simulating agents, etc.
    Y is normally some unimpressive low-level computery thing, such as
    predicting tokens
    sampling from a distribution
    querying a lookup table
    multiplying matrices
    sorting numbers
    clustering data points
    compressing text
    searching a tree
    manipulating bitstrings
    polarising magnetic strips, etc.
    Rather than address each example individually, I think it'll be more efficient to construct a general criterion by which we can assess each example.
    Click here for the specific example of LLMs.
    This criterion doesn't actually matter
    I should stress that this criterion doesn't actually matter for AI x-risk, because you can always reframe the risks in terms of Y, and not mention X at all. However, that might cost you more ink.
    ME, a visionary: GPT-4 is misaligned because it's simulating deceptive agents.YOU, a fool: GPT-4 isn't simulating any agents, it's just predicting which tokens continue a prompt.ME, a correct-opinion-haver: Fine, whatever... GPT-4 is misaligned because it predicts the tokens continuing a prompt by applying a function parameterised in a high-dimensional space to minimise cross-entropy loss across the internet corpus and the internet corpus contains a lot of conversations where one character deceives another and therefore GPT-4 will respond in the same way that a deceptive character would do so.
    The X-Y Criterion
    Informal statement
    Okay, here's the X-Y Criterion:
    If two tasks reduce to one another, then it is meaningless to ask if a machine is 'really doing' one task versus the other.
    Don't worry, later in the article we'll formalise what "task", "reduce", and "doing" means.
    First draft — computational reduction
    Our first draft will be "computational reduction".
    A task X is about processing classical information, i.e. X:{0,1}∗{0,1}∗.
    An algorithm A achieves a particular task X if it processes classical information in that way.
    In order to achieve a task X, the algorithm A expends certain quantities of computational resources, e.g. time, memory, samples, bandwidth, etc. These resources are abstract and non-physical.
    A task X reduces to task Y if and only if...For every algorithm A that solves task Y, there exists another algorithm B such that...(1) B solves task X by interacting with A.(2) The combined algorithm (A⊗B) doesn't expend much more computational resources to solve X as A expends to solve Y.
    X-Y Criterion: If two tasks X and Y reduce to one another, then it is meaningless to ask if an algorithm A is 'really doing' one task versus the other.
    This is what computer scientists mean when they say that one problem "reduces" to another task, e.g. when they say that all NP problems reduce to 3SAT.
    Second draft — physical reduction
    The second-draft formalisation will be "physical reduction".
    A task X is about changing the state of the world, i.e. X:ΩΩ.
    A machine A achieves a particular task X if it change...
    10 min
  • EA - Can we trust wellbeing surveys? A pilot study of comparability, linearity, and neutrality by Conrad S
    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: Can we trust wellbeing surveys? A pilot study of comparability, linearity, and neutrality, published by Conrad S on March 17, 2023 on The Effective Altruism Forum.
    Note: This post only contains Sections 1 and 2 of the report. For the full detail of our survey and pilot results, please see the full report on our website.
    Summary
    Subjective wellbeing (SWB) data, such as answers to life satisfaction questions, are important for decision-making by philanthropists and governments. Such data are currently used with two important assumptions:
    Reports are comparable between persons (e.g., my 6/10 means the same as your 6/10)
    Reports are linear in the underlying feelings (e.g., going from 4/10 to 5/10 represents the same size change as going from 8/10 to 9/10).
    Fortunately, these two assumptions are sufficient for analyses that only involve the quality of people’s lives. However, if we want to perform analyses that involve trade-offs between improving quality and quantity of life, we also need knowledge of the neutral point, the point on a wellbeing scale that is equivalent to non-existence.
    Unfortunately, evidence on all three questions is critically scarce. We propose to collect additional surveys to fill this gap.
    Our aim with this report is two-fold. First, we give an outline of the questions we plan to field and the underlying reasoning that led to them. Second, we present results from an initial pilot study (n = 128):
    Unfortunately, this small sample size does not allow us to provide clear estimates of the comparability of wellbeing reports.
    However, across several question modalities, we do find tentative evidence in favour of approximate linearity.
    With respect to neutrality, we assess at what point on a 0-10 scale respondents say that they are 'neither satisfied nor dissatisfied' (mean response is 5.3/10). We also probe at what point on a life satisfaction scale respondents report to be indifferent between being alive and being dead (mean response is 1.3/10). Implications and limitations of these findings concerning neutrality are discussed in Section 6.2.
    In general, the findings from our pilot study should only be seen as being indicative of the general feasibility of this project. They do not provide definitive answers.
    In the hopes of fielding an improved version of our survey with a much larger sample and a pre-registered analysis plan, we welcome feedback and suggestions on our current survey design.
    Here are some key questions that we hope to receive feedback on:
    Are there missing questions that could be included in this survey (or an additional survey) that would inform important topics in SWB research? Are there any questions or proposed analyses you find redundant?
    Do you see any critical flaws in the analyses we propose? Are there additional analyses we should be considering?
    Would these data and analyses actually reassure you about the comparability, linearity, and neutrality of subjective wellbeing data? If not, what sorts of data and analyses would reassure you?
    What are some good places for us to look for funding for this research?
    Of course, any other feedback that goes beyond these questions is welcome, too. Feedback can be sent to [email protected] or to [email protected].
    The report proceeds as follows:
    In Section 1, we describe the challenges for the use of self-reported subjective wellbeing data, focusing on the issues of comparability, linearity, and neutrality. We highlight the implications of these three assumptions for decision-making about effective interventions.
    In Section 2, we describe the general methodology of the survey.
    For the following sections, see the full report on our website.
    In Section 3, we discuss responses to the core life satisfaction question.
    In Sections 4, 5, and 6, we describe how we will assess co...
    21 min
  • EA - Survey on intermediate goals in AI governance by MichaelA
    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: Survey on intermediate goals in AI governance, published by MichaelA on March 17, 2023 on The Effective Altruism Forum.
    It seems that a key bottleneck for the field of longtermism-aligned AI governance is limited strategic clarity (see Muehlhauser, 2020, 2021). As one effort to increase strategic clarity, in October-November 2022, we sent a survey to 229 people we had reason to believe are knowledgeable about longtermist AI governance, receiving 107 responses. We asked about:
    respondents’ “theory of victory” for AI risk (which we defined as the main, high-level “plan” they’d propose for how humanity could plausibly manage the development and deployment of transformative AI such that we get long-lasting good outcomes),
    how they’d feel about funding going to each of 53 potential “intermediate goals” for AI governance,
    what other intermediate goals they’d suggest,
    how high they believe the risk of existential catastrophe from AI is, and
    when they expect transformative AI (TAI) to be developed.
    We hope the results will be useful to funders, policymakers, people at AI labs, researchers, field-builders, people orienting to longtermist AI governance, and perhaps other types of people. For example, the report could:
    Broaden the range of options people can easily consider
    Help people assess how much and in what way to focus on each potential “theory of victory”, “intermediate goal”, etc.
    Target and improve further efforts to assess how much and in what way to focus on each potential theory of victory, intermediate goal, etc.
    If you'd like to see a summary of the survey results, please request access to this folder. We expect to approve all access requests, and will expect readers to abide by the policy articulated in "About sharing information from this report" (for the reasons explained there).
    Acknowledgments
    This report is a project of Rethink Priorities–a think tank dedicated to informing decisions made by high-impact organizations and funders across various cause areas. The project was commissioned by Open Philanthropy. Full acknowledgements can be found in the linked "Introduction & summary" document.
    If you are interested in RP’s work, please visit our research database and subscribe to our newsletter.
    Here’s the definition of “intermediate goal” that we stated in the survey itself:
    By an intermediate goal, we mean any goal for reducing extreme AI risk that’s more specific and directly actionable than a high-level goal like ‘reduce existential AI accident risk’ but is less specific and directly actionable than a particular intervention. In another context (global health and development), examples of potential intermediate goals could include ‘develop better/cheaper malaria vaccines’ and ‘improve literacy rates in Sub-Saharan Africa’.
    Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org.
    3 min
  • EA - We are fighting a shared battle (a call for a different approach to AI Strategy) by Gideon Futerman
    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: We are fighting a shared battle (a call for a different approach to AI Strategy), published by Gideon Futerman on March 16, 2023 on The Effective Altruism Forum.
    Disclaimer 1: This following essay doesn’t purport to offer much original ideas, and I am certainly a non-expert on AI Governance, so please don’t take my word for these things too seriously. I have linked sources throughout the text, and have some other similar texts later on, but this should merely be treated as another data point in people saying very similar things; far smarter people than I have written on this.
    Disclaimer 2: This post is quite long, so I recommend reading the section on " A choice not an inevitability" and "It's all about power" for the core of my argument.
    My argument essentially is as follows; under most plausible understandings of how harms arise from very advanced AI systems, be these AGI or narrow AI or systems somewhere in between, the actors responsible, and the actions that must be taken to reduce or avert the harm, are broadly similar whether you care about both existential and non-existential harms from AI development. I will then further go on to argue that this calls for broad, coalitional politics of people who vastly disagree on specifics of AI systems harms, because we essentially have the same goals.
    It's important to note that calls like these have happened before. Whilst I will be taking a slightly different argument to them, Prunkl & Whittlestone, Baum, Stix & Maas and Cave & Ó hÉigeartaigh have all made arguments attempting to bridge near term and long term concerns. In general, these proposals (with the exception of Baum) have made calls for narrower cooperation between ‘AI Ethics’ and ‘AI Safety’ than I will make, and are all considerably less focused on the common source of harm than I will be. None go as far as I do in essentially suggesting all key forms of harm that we worry about are incidents of the same phenomena of power concentration in and through AI. These pieces are in many ways more research focused, whilst mine is considerably more politically focused.
    Nonetheless, there is considerable overlap in spirit of identifying that the near-term/ethics and long-term/safety distinction is overemphasised and is not as analytically useful as is made out, as well as the intention of all these pieces and mine to reconcile for mutual benefit of the two factions.
    A choice not an inevitability
    At present, there is no AI inevitably coming to harm us. Those AIs that do will be given capabilities, and power to cause harm, by developers. If the AI companies stopped developing their AIs now, and people chose to stop deploying them, then both existential or non-existential harms would stop. These harms are in our hands, and whilst the technologies clearly act as important intermediaries, ultimately it is a human choice, a social choice, and perhaps most importantly a political choice to carry on developing more and more powerful AI systems when such dangers are apparent (or merely plausible or possible). The attempted development of AGI is far from value neutral, far from inevitable and very much in the realm of legitimate political contestation. Thus far, we have simply accepted the right for powerful tech companies to decide our future for us; this is both unnecessary and dangerous.
    It's important to note that our current acceptance of the right of companies to legislate for our future is historically contingent. In the past, corporate power has been curbed, from colonial era companies, Progressive Era trust-busting, postwar Germany and more, and this could be used again. Whilst governments have often taken a leading role, civil society has also been significant in curbing corporate power and technology development throughout history. Acceptance of corporate dominance i...
    25 min
  • LW - On the Crisis at Silicon Valley Bank by Zvi
    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: On the Crisis at Silicon Valley Bank, published by Zvi on March 16, 2023 on LessWrong.
    Many have already written about the events of the past week’s crisis.
    If you want the basics of what happened, you have many options.
    Your best bet, if available to you, is that this is Matt Levine’s wheelhouse. He did not disappoint, offering at least (1) (2) (3) (4) posts on the subject.
    Then read Patrick McKenzie if you want the nuts and bolts of the underlying systems spelled out in plainer language and more detail, without judgment, along with the basics of what a responsible individual should do now, things he is better at explaining than I am.
    Then read someone like Scott Sumner here if you need to get the necessary counterpoints on moral hazard.
    I will do my best to cover all the necessary background in the What Happened section, to bring you up to speed. What I am not trying to do is duplicate Levine’s work. I am also going to skip the explainers of things like ‘what is a bank run,’ since they are well-covered by many others – choose one of these ungated linked summaries, or better yet Matt Levine, to read first if you need that level of info.
    Instead, I am asking the questions, and looking at the things, that I found most interesting, or most important for understanding the world going forward.
    What did I find most interesting? Here are some of my top questions.
    What exactly would have happened without an intervention?
    What changes for banking in the age of instant electronic banking and social networks?
    How much money have our banks lost exactly? What might happen anyway?
    How much does talk of ‘bailout’ and laws we’ve passed constrain potential future interventions if something else threatens to go wrong? Ut oh.
    Is Hold to Maturity accounting utter bullshit and a main suspect here? Yes.
    What should depositing businesses be responsible for?
    What stories are people telling about what happened, and why?
    How do we deal with all the problems of moral hazard? What is enough?
    More generally, what the hell do we do about all this?
    I also wonder about a variety of other things, such as what happened with USDC trading so low, to what extent people really do hate big tech, and more.
    What Happened
    In one meme:
    Silicon Valley Bank had a ton of deposits that didn’t pay interest, largely from start-ups flush with cash. They attracted that cash by offering high-touch bespoke services. The problem is that those services cost money, and there was no actually safe way to make that money back using their deposits.
    SVB could have said ‘our business is not profitable right now, but it is helping us build a future highly profitable business’ and used that value to raise equity capital, perhaps from some of their venture fund clients who are used to these types of moves.
    They decided to go a different way. Rather than accept that their business was unprofitable, they bought a ton of very low-yielding assets that were highly exposed to interest rate hikes. That way they looked profitable, in exchange for taking on huge interest rate risk on top of their existing interest rate risk from their customer base.
    Interest rates went up. Those assets lost $15 billion in value, while customers vulnerable to high interest rates become cash poor.
    Also SVB was in the business of providing venture debt to its clients. I have never understood venture debt. Why would you lend money to a start-up, what are you hoping for? If they pay you back you should have invested instead, if they don’t pay you don’t get paid, and if you get warrants as part of the deal it looks a lot like investing in the start-up with strange and confusing terms. Or if we look at this thread, perhaps there is no catch, it is simply a bribe to get people to bank with you so you can bet their deposits on low interest rates?
    So maybe I do und...
    1 hr
  • EA - Donation offsets for ChatGPT Plus subscriptions by Jeffrey Ladish
    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: Donation offsets for ChatGPT Plus subscriptions, published by Jeffrey Ladish on March 16, 2023 on The Effective Altruism Forum.
    I've decided to donate $240 to both GovAI and MIRI to offset the $480 I plan to spend on ChatGPT Plus over the next two years ($20/month).I don't have a super strong view on ethical offsets, like donating to anti-factory farming groups to try to offset harm from eating meat. That being said, I currently think offsets are somewhat good for a few reasons:They seem much better than simply contributing to some harm or commons problem and doing nothing, which is often what people would do otherwise.It seems useful to recognize, to notice, when you're contributing to some harm or commons problem. I think a lot of harm comes from people failing to notice or keep track of ways their actions negatively impact others, and the ways that common incentives push them to do worse things.A common Effective Altruism argument against offsets is that they don't make sense from a consequentialist perspective. If you have a budget for doing good, then spend your whole budget on doing as much as possible.
    If you want to mitigate harms you are contributing to, you can offset by increasing your "doing good" budget, but it doesn't make sense to specialize your mitigations to the particular area where you are contributing to harm rather than the area you think will be the most cost effective in general.I think this is a decently good point, but doesn't move me enough to abandon the idea of offsets entirely. A possible counter-argument is that offsets can be a powerful form of coordination to help solve commons problems. By publicly making a commitment to offset a particular harm, you're establishing a basis for coordination - other people can see you really care about the issue because you made a costly signal.
    This is similar for the reasons to be vegan or vegetarian - it's probably not the most effective from a naive consequentialist perspective, but it might be effective as a point of coordination via costly signaling.After having used ChatGPT (3.5) and Claude for a few months, I've come to believe that these tools are super useful for research and many other tasks, as well as useful for understanding AI systems themselves. I've also started to use Bing Chat and ChatGPT (4), and found them to be even more impressive as research and learning tools. I think it would be quite bad for the world if conscientious people concerned about AI harms refrained from using these tools, because I think it would disadvantage them in significant ways, including in crucial areas like AI alignment and policy.
    Unfortunately both can be true:1) Language models are really useful and can help people learn, write, and research more effectively2) The rapid development of huge models is extremely dangerous and a huge contributor to AI existential riskI think OpenAI, and to varying extent other scaling labs, are engaged in reckless behavior scaling up and deploying these systems before we understand how they work enough to be confident in our safety and alignment approaches. And also, I do not recommend people in the "concerned about AI x-risk" reference class refrain from paying for these tools, even if they do not decide to offset these harms. The $20/month to OpenAI for GPT-4 access right now is not a lot of money for a company spending hundreds of millions training new models.
    But it is something, and I want to recognize that I'm contributing to this rapid scaling and deployment in some way.Weighing all this together, I've decided offsets are the right call for me, and I suspect they might be right for many others, which is why I wanted to share my reasoning here. To be clear, I think concrete actions aimed at quality alignment research or AI policy aimed at buying more time are much mo...
    5 min
  • LW - Donation offsets for ChatGPT Plus subscriptions by Jeffrey Ladish
    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: Donation offsets for ChatGPT Plus subscriptions, published by Jeffrey Ladish on March 16, 2023 on LessWrong.
    I've decided to donate $240 to both GovAI and MIRI to offset the $480 I plan to spend on ChatGPT Plus over the next two years ($20/month).I don't have a super strong view on ethical offsets, like donating to anti-factory farming groups to try to offset harm from eating meat. That being said, I currently think offsets are somewhat good for a few reasons:They seem much better than simply contributing to some harm or commons problem and doing nothing, which is often what people would do otherwise.It seems useful to recognize, to notice, when you're contributing to some harm or commons problem. I think a lot of harm comes from people failing to notice or keep track of ways their actions negatively impact others, and the ways that common incentives push them to do worse things.A common Effective Altruism argument against offsets is that they don't make sense from a consequentialist perspective. If you have a budget for doing good, then spend your whole budget on doing as much as possible.
    If you want to mitigate harms you are contributing to, you can offset by increasing your "doing good" budget, but it doesn't make sense to specialize your mitigations to the particular area where you are contributing to harm rather than the area you think will be the most cost effective in general.I think this is a decently good point, but doesn't move me enough to abandon the idea of offsets entirely. A possible counter-argument is that offsets can be a powerful form of coordination to help solve commons problems. By publicly making a commitment to offset a particular harm, you're establishing a basis for coordination - other people can see you really care about the issue because you made a costly signal.
    This is similar for the reasons to be vegan or vegetarian - it's probably not the most effective from a naive consequentialist perspective, but it might be effective as a point of coordination via costly signaling.After having used ChatGPT (3.5) and Claude for a few months, I've come to believe that these tools are super useful for research and many other tasks, as well as useful for understanding AI systems themselves. I've also started to use Bing Chat and ChatGPT (4), and found them to be even more impressive as research and learning tools. I think it would be quite bad for the world if conscientious people concerned about AI harms refrained from using these tools, because I think it would disadvantage them in significant ways, including in crucial areas like AI alignment and policy.
    Unfortunately both can be true:1) Language models are really useful and can help people learn, write, and research more effectively2) The rapid development of huge models is extremely dangerous and a huge contributor to AI existential riskI think OpenAI, and to varying extent other scaling labs, are engaged in reckless behavior scaling up and deploying these systems before we understand how they work enough to be confident in our safety and alignment approaches. And also, I do not recommend people in the "concerned about AI x-risk" reference class refrain from paying for these tools, even if they do not decide to offset these harms. The $20/month to OpenAI for GPT-4 access right now is not a lot of money for a company spending hundreds of millions training new models.
    But it is something, and I want to recognize that I'm contributing to this rapid scaling and deployment in some way.Weighing all this together, I've decided offsets are the right call for me, and I suspect they might be right for many others, which is why I wanted to share my reasoning here. To be clear, I think concrete actions aimed at quality alignment research or AI policy aimed at buying more time are much more important than o...
    5 min
  • LW - Conceding a short timelines bet early by Matthew Barnett
    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: Conceding a short timelines bet early, published by Matthew Barnett on March 16, 2023 on LessWrong.
    Last year I bet some people about short AI timelines. While I don't think I've lost the bet yet, I think it's clear at this point that I will lose with high probability. I've outlined the reasons why I think that in a retrospective here. Even if I end up winning, I think it will likely be the result of a technicality, and that wouldn't be very interesting.
    Because of my personal preference for settling this matter now without delay, I have decided to take the step of conceding the bet now. Note however that I am not asking Tamay to do the same. I have messaged the relevant parties and asked them to send me details on how to pay them.
    I congratulate Nathan Helm-Burger and Tomás B. for taking the other side of the bet.
    Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org.
    2 min