The Nonlinear Library

The Nonlinear Library

By The Nonlinear FundEducation
Download on the App Store

All The Nonlinear Library episodes

  • EA - Who owns "Effective Altruism"? by River
    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: Who owns "Effective Altruism"?, published by River on February 1, 2023 on The Effective Altruism Forum.
    I've noticed a bit of confusion among EAs about whether CEA (or technically EVF) owns the phrase "effective altruism". I've heard a couple of EAs claim they have the trademark, or asserted some kind of ownership right against a scammer once. And I do have the impression that CEA tries to manage the EA brand in a way that suggests that they feel like they own it.However, I checked the US Patent and Trademark Office's database, it shows trademarks on "Centre for Effective Altruism" and "Effective Altruism Global", but not "effective altruism". Is it even possible to have a trademark on the name of a philosophy? Is there some other area of law that is relevant?
    Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org.
    1 min
  • AF - Trends in the dollar training cost of machine learning systems by Ben Cottier
    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: Trends in the dollar training cost of machine learning systems, published by Ben Cottier on February 1, 2023 on The AI Alignment Forum.
    Summary
    Using a dataset of 124 machine learning (ML) systems published between 2009 and 2022, I estimate that the cost of compute in US dollars for the final training run of ML systems has grown by 0.49 orders of magnitude (OOM) per year (90% CI: 0.37 to 0.56). See Table 1 for more detailed results, indicated by "All systems."
    By contrast, I estimate that the cost of compute used to train "large-scale" systems since September 2015 (systems that used a relatively large amount of compute) has grown more slowly compared to the full sample, at a rate of 0.2 OOMs/year (90% CI: 0.1 to 0.4 OOMs/year). See Table 1 for more detailed results, indicated by "Large-scale."
    Based on the historical trends, and reviewing some prior work (Lohn & Musser, 2022 and Cotra, 2020), I estimated my best guess for how quickly costs will grow in the future. Here, I'm assuming a model like the one used by Cotra (2020), where this growth rate is sustained up until spending hits a limit at some non-trivial fraction of gross world product. The below estimates are much less robust than the historical trends. (more)
    My independent impression: 0.3 OOMs/year (90% CI: 0.1 to 0.4 OOMs/year)
    My all-things-considered view: 0.2 OOMs/year (90% CI: 0.1 to 0.3 OOMs/year)
    For future work, I recommend the following:
    Incorporate systems trained on Google TPUs, and TPU price-performance data, into Method 2.
    Estimate more reliable bounds on training compute costs, rather than just point estimates. For example, research the profit margin of NVIDIA and adjust retail prices by that margin to get a lower bound on hardware cost.
    As a broader topic, investigate trends in investment, spending allocation, and AI revenue.
    DataPeriodScale (start to end)Growth rate in dollar cost for final training runs(1) Using the overall GPU price-performance trend (go to results)All systems (n=124)$0.02 to $80KLarge-scale (n=25)$30K to $1M(2) Using the peak price-performance of the actual NVIDIA GPUs used to train ML systems (go to results)All systems (n=48)$0.10 to $80KLarge-scale (n=6)$200 to $70KWeighted mixture of growth ratesAll systemsN/A
    Estimation method
    (go to explanation)
    Jun 2009–
    Jul 2022
    0.51 OOMs/year
    90% CI: 0.45 to 0.57
    Oct 2015–
    Jun 2022
    0.2 OOMs/year
    90% CI: 0.1 to 0.4
    Jun 2009–
    Jul 2022
    0.44 OOMs/year
    90% CI: 0.34 to 0.52
    Sep 2016–
    May 2022
    0.2 OOMs/year
    90% CI: 0.1 to 0.4
    Jun 2009–
    Jul 2022
    0.49 OOMs/year
    90% CI: 0.37 to 0.56
    Table 1: Estimated growth rate in the dollar cost of compute to train ML systems over time, based on a log-linear regression. OOM = order of magnitude (10x). See the section Summary of regression results for expanded result tables.
    Figure 1: estimated cost of compute in US dollars for the final training run of ML systems. The costs here are estimated based on the trend in price-performance for all GPUs in Hobbhahn & Besiroglu (2022) (known as "Method 1" in this report).
    Read the rest of the report here
    These are "milestone" systems selected from the database Parameter, Compute and Data Trends in Machine Learning, using the same criteria as described in Sevilla et al. (2022, p.16): "All models in our dataset are mainly chosen from papers that meet a series of necessary criteria (has an explicit learning component, showcases experimental results, and advances the state-of-the-art) and at least one notability criterion (>1000 citations, historical importance, important SotA advance, or deployed in a notable context). For new models (from 2020 onward) it is harder to assess these criteria, so we fall back to a subjective selection. We refer to models meeting our selection criteria as milestone models."
    This growth rate is about 0.2 OOM/year lower than the growth of t...
    8 min
  • EA - Nelson Mandela’s organization, The Elders, backing x risk prevention and longtermism by krohmal5
    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: Nelson Mandela’s organization, The Elders, backing x risk prevention and longtermism, published by krohmal5 on February 1, 2023 on The Effective Altruism Forum.
    Excerpt from the web posting below. Strong endorsement of EA priorities from a decidedly not EA source—The Elders—an organization founded by Nelson Mandela in 2007. Although their reference to “approaching a precipice” and imagery that looks a lot like the Toby Ord book cover seems unlikely to be a coincidence? Also, thanks to Nathan for encouraging me to post this on the forum, a first for me.
    “For the next five years, our focus will be on the climate crisis, nuclear weapons, pandemics, and the ongoing scourge of conflict. The impact of these threats is already being seen on lives and livelihoods: a rapid rise in extreme weather events, a pandemic that killed millions and cost trillions, a war in which the use of nuclear weapons has been openly raised. Some of these threats jeopardise the very existence of human life on our planet, yet nations lack the ability or will to manage these risks.
    We are approaching a precipice.
    The urgency of the interconnected existential threats we face requires a crisis mindset from our leaders – one that puts our shared humanity centre-stage, leaves no one behind, and recognises the rights of future generations.
    We believe that when nations work together, these threats can all be addressed for the good of the whole world. But to do so, our political leaders must show purpose, moral courage and an urgency of action.
    This is why, over the next five years, we will work on three programmes that address existential threats to humanity requiring a collective response - the climate crisis, pandemics, and nuclear weapons - and also on conflict (a threat in itself, and a risk factor for other threats).”
    Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org.
    2 min
  • LW - Aiming for Convergence Is Like Discouraging Betting by Zack M Davis
    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: Aiming for Convergence Is Like Discouraging Betting, published by Zack M Davis on February 1, 2023 on LessWrong.
    Summary
    In a list of guidelines for rational discourse, Duncan Sabien proposes that one should "[a]im for convergence on truth, and behave as if your interlocutors are also aiming for convergence on truth."
    However, prediction markets illustrate fundamental reasons why rational discourse doesn't particularly look like "aiming for convergence." When market prices converge on the truth, it's because traders can only make money by looking for divergences where their beliefs are more accurate than the market's. Similarly, when discussions converge on the truth, it's because interlocutors can only advance the discussion by making points where the discussion-so-far has been wrong or incomplete. Convergence on the truth, if it happens, happens as a side-effect of correctly ironing out all existing mispricings/disagreements; it seems wrong to describe this as "aiming for convergence" (even if convergence would be the end result if everyone were reasoning perfectly).
    Sabien's detailed discussion of the "aim for convergence on truth" guideline concerns itself with how to determine whether an interlocutor is "present in good faith and genuinely trying to cooperate." I don't think I understand how these terms are being used in this context. More generally, the value of "collaborative truth-seeking" is unclear to me: if I can evaluate arguments on their merits, the question of whether the speaker is "collaborative" with me does not seem intellectually substantive.
    Mostly, I don't expect to disagree with heavily-traded prediction markets. If the market says it's going to rain on Saturday with 85% probability, then I (lacking any special meteorology knowledge) basically think it's going to rain on Saturday at 85% probability.
    Why is this? Why do I defer to the market, instead of tarot cards, or divination sticks, or my friend Maddie the meteorology enthusiast?
    Well, I don't expect the tarot cards to tell me anything about whether it will rain on Saturday, because there's no plausible physical mechanism by which information about the weather could influence the cards. Shuffling and dealing the cards should work the same in worlds where it will rain and worlds where it won't rain. Even if there is some influence (because whether it will rain affects the moisture and atmospheric pressure in the air, which affects my grip on the cards, which affects my shuffling motion?), it's not something I can detect from which cards are drawn.
    I do expect my friend Maddie the meteorology enthusiast to tell me something about whether it will rain on Saturday. That's because she's always looking at the latest satellite cloud data and tinkering with her computer models, which is a mechanism by which information about the weather can influence her forecasts. The cloud data will be different in worlds where it will rain and worlds where it won't rain. If Maddie is pretty sharp and knows her stuff, maybe she can tell the difference.
    And yet—no offense, Maddie—I expect the market to do even better. It's not just that the market has a lot of other pretty sharp people looking at the cloud data, and that maybe some of them are even sharper than Maddie, even though Maddie is my friend and my friends are the best.
    It's that the market mechanism rewards people for being less wrong than the market. If the rain-on-Saturday market is trading at 85%, and Maddie's rival Kimber buys 100 shares of No, that doesn't mean Kimber thinks it's not going to rain. It means Kimber thinks 85% is too high. If Kimber thinks it's "actually" only going to rain with 80% probability, then she figures that a No share that pays out $1 if it doesn't rain should be worth 20¢. If it's currently trading for 15¢, it's worth buyin...
    18 min
  • AF - The effect of horizon length on scaling laws by Jacob Hilton
    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 effect of horizon length on scaling laws, published by Jacob Hilton on February 1, 2023 on The AI Alignment Forum.
    The scaling of optimal model size with compute is a key input into the biological anchors framework for forecasting transformative AI. In particular, the "effective horizon length" introduces a multiplier into this scaling law that can have a big effect on forecasts.
    This paper studies this scaling law for several RL environments: Procgen, Dota 2 and a toy MNIST-based environment. The last of these is used to study the effect of the task horizon length in a toy setting. There are a number of takeaways for the biological anchors framework, which are summarized in Section 5.4.
    Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org.
    1 min
  • LW - Nice Clothes are Good, Actually by Gordon Seidoh Worley
    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: Nice Clothes are Good, Actually, published by Gordon Seidoh Worley on January 31, 2023 on LessWrong.
    This is just some casual thoughts about how to dress nice. I think dressing nice is a useful skill, but nobody told me in a way that convinced me. So this is written in the spirit of what I wish someone would have told 15 year old me.
    For most of my life I thought dressing nicely was a type of bullshit signaling game I didn't need to play because I was doing Real Work. I'm here to tell you that I was mistaken and that dressing nicely is good and useful to your life.
    My story went something like this:
    I care about doing important, meaningful things in the world, like solving hard technical problems.
    I can solve those problems regardless of what I wear.
    Further, lots of the people who seem best at solving hard technical problems dress poorly.
    Also, I don't really understand fashion that well, and most of the people who do understand it seem to be focused on things other than solving hard technical problems.
    Conclusion: I should not care about dressing nicely. Just wear whatever is comfortable regardless of how it looks.
    Dressing poorly or casually, though, is still a kind of signaling game. My reasoning makes this clear: it shows I'm part of the ingroup of serious people solving serious, real problems rather than part of the outgroup of silly people who care more about how they look than getting things done.
    In some situations and in some contexts, this is a true, useful fact. If you show up to a coding interview in a suit, you've probably hurt your chances to getting the job. Even if your code is great, the person interviewing you is going to worry that you have expectations about the job that don't line up with reality. And, if we're honest, you don't look the part, so unconscious bias comes into play.
    However, unconscious bias is the start of seeing why dressing nicely, or at least intentionally, matters. Another word for unconscious bias is vibes. What vibe do your clothes put off? That is, how do you present yourself to the world? What interfaces and affordances are you giving people based on the clothes you wear?
    Some examples:
    Wearing a pot-leaf hoodie? People will reasonably think you smoke weed and that smoking weed is important to you.
    Wearing a three-piece suit? People will think you have money (because you had enough money at least to afford the suit) and are serious.
    Wearing a poorly fitted suit? People will think you are a low-level employee of a faceless business and you don't have the experience or the money to dress better.
    Wearing a sundress or a tank top and shorts? People will think you're relaxed and out to have a nice day off.
    Wearing yoga pants or sweat pants and graphic tee? You're not especially concerned about your looks and just focused on getting through the day.
    And on and on. If you don't have the same associations with these clothes as I do, that's fine. One of the tough things about clothes is that not everyone gets the same vibe from the same clothes. This makes the problem of how to dress nice harder. Doesn't mean it's unsolvable, just that the solutions are fuzzy.
    So why should you dress nice, even given this challenge? Because dressing nice makes your vibes better and people treat you better and are more willing to accommodate your requests. This is related to the way being attractive causes people to treat you better. While you can't easily fix your face or age or weight (though you can try), you can easily change your clothes, and clothes go a long way towards making you attractive.
    What does dressing nice mean? A few things:
    picking clothes that fit your body shape rather than whatever is in fashion or available off the rack
    picking clothes with fabrics that look and feel nice and of high quality
    wearing styles that reflect the ima...
    6 min
  • EA - More Is Probably More - Forecasting Accuracy and Number of Forecasters on Metaculus by nikos
    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: More Is Probably More - Forecasting Accuracy and Number of Forecasters on Metaculus, published by nikos on January 31, 2023 on The Effective Altruism Forum.
    TLDR
    An increase in the number of forecasters seems to lead to an improvement of the Metaculus community prediction. I believe this effect is real, but due to confounding effects, the analysis presented here might overestimate the improvement gained.
    That improvement of the Metaculus community prediction seems to be approximately logarithmic, meaning that doubling the number of forecasters seems to lead to a roughly constant (albeit probably diminishing) relative improvement in performance in terms of Brier Score: Going from 100 to 200 would give you a relative improvement in Brier score almost as large as when going from 10 to 20 (e.g. an improvement by x percent). Note though, that it is a bit unclear what "an improvement in Brier score by X" actually means in terms of forecast quality.
    Increasing the number of forecasters on Metaculus seems to not only improve performance on average, but also seems to decrease the variability of predictions, making them more stable and reliable
    This analysis complements another existing one and comes to similar conclusions. Both analyses suffer from potential biases, but they are different ones.
    All code used for this analysis can be found here.
    Introduction
    One of the central wisdoms in forecasting is that an ensemble of forecasts is more than the sum of its parts. Take a crowd of forecasters and average their predictions - the resulting ensemble will usually be more accurate than almost all of the individual forecasts.
    But how does the performance of the ensemble change when you increase the number of forecasters? Are fifty forecasters ten times as good as five? Are five hundred even better? Charles Dillon looked at this a while ago using Metaculus data.
    He broadly found that more forecasters usually means better performance. Specifically, he estimated that doubling the number of forecasters would reduce the average Brier score by 0.012 points. The Brier score is a metric commonly used to evaluate the performance of forecasts with a binary yes/no outcome and equals the squared difference between the outcome (0 or 1) and the forecast. Smaller values are better. Charles concluded that in practice a Metaculus community prediction with only ten forecasters is not a lot less reliable than a community prediction with thirty forecasters.
    Charles' analysis was restricted to aggregated data, which means that he had access to the Metaculus community prediction, but not to individual level data. This makes the analysis susceptible to potential biases. For example, it could be the case that forecasters really like easy questions and that those questions which attracted fewer forecasters were genuinely harder. We would then expect to see worse performance on questions with fewer forecasters even if the number of forecasters had no actual effect on performance. In this post I will try to shed some more light on the question, this time making use of individual level data.
    Methodology
    To examine the effect of the number of forecasters on the performance of the community prediction, we can use a technique called "bootstrapping". The idea is simple. Take a question that has n = 200 forecasters. And let's suppose we are interested in how the community prediction would have performed had there only been n = 5, 10, 20, or 50 instead of the actual 200 forecasters. To find out we can take a random sample of size n, e.g. 5, forecasters, discard all the other ones, compute a community prediction based on just these 5 forecasters and see how it would have performed. One data point isn't all too informative, so we repeat the process a few thousand times and average the results. Now we do that for othe...
    17 min
  • AF - On value in humans, other animals, and AI by Michele Campolo
    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 value in humans, other animals, and AI, published by Michele Campolo on January 31, 2023 on The AI Alignment Forum.
    This will be posted also on the EA Forum, and included in a sequence containing some previous posts and other posts I'll publish this year.
    Introduction
    Humans think critically about values and, to a certain extent, they also act according to their values. To the average human, the difference between increasing world happiness and increasing world suffering is huge and evident, while goals such as collecting coins and collecting stamps are roughly on the same level.
    It would be nice to make these differences obvious to AI as they are to us. Even though exactly copying what happens in the human mind is probably not the best strategy to design an AI that understands ethics, having an idea of how value works in humans is a good starting point.
    So, how do humans reason about values and act accordingly?
    Key points
    Let’s take a step back and start from sensation. Through the senses, information goes from the body and the external environment to our mind.
    After some brain processing — assuming we’ve had enough experiences of the appropriate kind — we perceive the world as made of objects. A rock is perceived as distinct from its surrounding environment because of its edges, its colour, its weight, the fact that my body can move through air but not through rocks, and so on.
    Objects in our mind can be combined with each other to form new objects. After seeing various rocks in different contexts, I can imagine a scene in which all these rocks are in front of me, even though I haven’t actually seen that scene before.
    We are also able to apply our general intelligence — think of skills such as categorisation, abstraction, induction — to our mental content.
    Other intelligent animals do something similar. They probably understand that, to satisfy thirst, water in a small pond is not that different from water flowing in a river. However, an important difference is that animals’ mental content is more constrained than our mental mental content: we are less limited by what we perceive in the present moment, and we are also better at combining mental objects with each other.
    For example, to a dog, its owner works as an object in the dog’s mind, while many of its owner’s beliefs do not. Some animals can attribute simple intentions and perception, e.g. they understand what a similar animal can and cannot see, but it seems they have trouble attributing more complex beliefs.
    The ability to compose mental content in many different ways is what allows us to form abstract ideas such as mathematics, religion, and ethics, just to name a few.
    Key point 1:
    In humans, mental content can be abstract.
    Now notice that some mental content drives immediate action and planning. If I feel very hungry, I will do something about it, in most cases.
    This process from mental content to action doesn’t have to be entirely conscious. I can instinctively reach for the glass of water in front of me as a response to an internal sensation, even without moving my attention to the sensation nor realising it is thirst.
    Key point 2:
    Some mental content drives behaviour.
    Not all mental content drives action and planning. The perception of an obstacle in front of me might change how I carry out my plans and actions, but it is unlikely to change what I plan and act for. Conversely, being very hungry directly influences what I’m going to do — not just how I do it — and can temporarily override other drives. It is in this latter sense that some mental content drives behaviour.
    In humans, the mental content that does drive behaviour can be roughly split in two categories.
    The first one groups what we often call evolutionary or innate drives, like hunger and thirst in the examples above, and works similarly i...
    9 min
  • EA - How to use AI speech transcription and analysis to accelerate social science research by AlexanderSaeri
    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 use AI speech transcription and analysis to accelerate social science research, published by AlexanderSaeri on January 31, 2023 on The Effective Altruism Forum.
    Summary
    AI tools like OpenAI Whisper and GPT-3 can be used to improve social science research workflows by helping to collect and analyse speech and text data.
    In this article, I describe two worked examples where I applied AI tools to (1) transcribe and (2) conduct basic thematic analysis of a research interview, and provide enough detail for readers to replicate and adapt my approach.
    OpenAI Whisper (example) created a high quality English transcription of a 30 minute research interview at a ~70x cost saving compared to a human transcriber.
    GPT-3 (text-davinci-003; example) answered a research question and identified relevant themes from a transcribed research interview, after providing a structured prompt and one example.
    These tools, when chained together with human oversight, can be considered an early, weak example of PASTA (Process for Automating Scientific and Technological Advancement).
    Social science research workflows involve a lot of speech and text data that is laborious to collect and analyse
    The daily practice of social science research involves a lot of talking, reading and writing. In my applied behaviour science research consulting role at Monash University and through Ready Research, I generate or participate in the generation of a huge amount of speech and text data. This includes highly structured research activities such as interviews, surveys, observation and experiments; but also less structured research activities like workshops and meetings.
    Some fictionalised examples of work I’ve done in the past year:
    Research interviews with 20 regular city commuters to understand what influences their commuting behaviour post-COVID, to assist a public transit authority in planning and operating its services efficiently
    Practitioner interviews with staff from city, regional and rural local governments to assess organisational readiness for flood preparation and response
    Workshop of 5-10 people involved in hospital sepsis care, each representing a different interest (e.g., patients, clinicians, researchers, funders) to identify priority areas to direct $5M research funding
    Survey of 5,000 Australians to understand the impacts and experiences of living under lockdown in Melbourne, Australia during COVID-19
    Evaluation interviews with 4 participants in the AGI Safety Fundamentals course to understand the most significant change in their knowledge, skills, or behaviours as a result of their participation
    To make this data useful it needs to be collected, processed, organised, structured and analysed. The typical workflow for these kinds of activities involves taking written notes during the research activity, or recording the audio / video research activity and reviewing the recording later. Interviews are sometimes transcribed by a paid service for later analysis. Other times they are transcribed by the researcher.
    The amount of speech and text data generated during research activity is large - each research activity yields thousands of words. The sheer volume of data can be overwhelming and daunting, making it difficult to carry out analysis in any meaningful way. In addition, sometimes data just isn’t collected (e.g., during an interview or workshop) because the researcher is busy facilitating / listening / processing / connecting with research participants.
    Even for data that is collected, managing and analysing it is a challenge. Specialised programs such as nVivo are used in social science to manage and analyse text data, but less structured research activities would almost never be managed or analysed through this kind of program, because of the time and skills required. Open text data i...
    21 min
  • AF - Criticism of the main framework in AI alignment by Michele Campolo
    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: Criticism of the main framework in AI alignment, published by Michele Campolo on January 31, 2023 on The AI Alignment Forum.
    Originally posted on the EA Forum for the Criticism and Red Teaming Contest. Will be included in a sequence containing some previous posts and other posts I'll publish this year.
    0. Summary
    AI alignment research centred around the control problem works well for futures shaped by out-of-control misaligned AI, but not that well for futures shaped by bad actors using AI. Section 1 contains a step-by-step argument for that claim. In section 2 I propose an alternative which aims at moral progress instead of direct risk reduction, and I reply to some objections. I will give technical details about the alternative at some point in the future, in section 3.
    The appendix clarifies some minor ambiguities with terminology and links to other stuff.
    1. Criticism of the main framework in AI alignment
    1.1 What I mean by main framework
    In short, it’s the rationale behind most work in AI alignment: solving the control problem to reduce existential risk. I am not talking about AI governance, nor about AI safety that has nothing to do with existential risk (e.g. safety of self-driving cars).
    Here are the details, presented as a step-by-step argument.
    At some point in the future, we'll be able to design AIs that are very good at achieving their goals. (Capabilities premise)
    These AIs might have goals that are different from their designers' goals. (Misalignment premise)
    Therefore, very bad futures caused by out-of-control misaligned AI are possible. (From previous two premises)
    AI alignment research that is motivated by the previous argument often aims at making misalignment between AI and designer, or loss of control, less likely to happen or less severe. (Alignment research premise).
    Common approaches are ensuring that the goals of the AI are well specified and aligned with what the designer originally wanted, or making the AI learn our values by observing our behaviour. In case you are new to these ideas, two accessible books on the subject are [1,2].
    5. Therefore, AI alignment research improves the expected value of bad futures caused by out-of-control misaligned AI. (From 3 and 4).
    By expected value I mean a measure of value that takes likelihood of events into account, and follows some intuitive rules such as "5% chance of extinction is worse than 1% chance of extinction". It need not be an explicit calculation, especially because it might be difficult to compare possible futures quantitatively, e.g. extinction vs dystopia.
    I don't claim that all AI alignment research follows this framework; just that this is what motivates a decent amount (I would guess more than half) of work in AI alignment.
    1.2 Response
    I call this a response, and not a strict objection, because none of the points or inferences in the previous argument is rejected. Rather, some extra information is taken into account.
    6. Bad actors can use powerful controllable AI to bring about very bad futures and/or lock-in their values (Bad actors premise)
    For more information about value lock-in, see chapter 4 of What We Owe The Future [3].
    7. Recall that alignment research motivated by the above points makes it easier to design AI that is controllable and whose goals are aligned with its designers' goals. As a consequence, bad actors might have an easier time using powerful controllable AI to achieve their goals. (From 4 and 6)
    8. Thus, even though AI alignment research improves the expected value of futures caused by uncontrolled AI, it reduces the expected value of futures caused by bad human actors using controlled AI to achieve their ends. (From 5 and 7)
    This conclusion will seem more, or less, relevant depending on the beliefs you have about its different components.
    An example: if you think t...
    13 min