The Nonlinear Library

The Nonlinear Library

By The Nonlinear FundEducation
Download on the App Store

All The Nonlinear Library episodes

  • EA - Forecasts on Moore v Harper from Samotsvety by gregjustice
    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: Forecasts on Moore v Harper from Samotsvety, published by gregjustice on March 20, 2023 on The Effective Altruism Forum.
    [edited to include full text]
    Disclaimers
    The probabilities listed are contingent on SCOTUS issuing a ruling on this case. An updated numerical forecast on that happening, particularly in light of the NC Supreme Court’s decision to rehear Harper v Hall, may be forthcoming.
    The author of this report, Greg Justice, is an excellent forecaster, not a lawyer. This post should not be interpreted as legal advice. This writeup is still in progress, and the author is looking for a good venue to publish it in.
    You can subscribe to these posts here.
    Introduction
    The Moore v. Harper case before SCOTUS asks to what degree state courts can interfere with state legislatures in the drawing of congressional district maps. Versions of the legal theory they’re being asked to rule on were invoked as part of the attempts to overthrow the 2020 election, leading to widespread media coverage of the case. The ruling here will have implications for myriad state-level efforts to curb partisan gerrymandering.
    Below, we first discuss the Independent State Legislature theory and Moore v. Harper. We then offer a survey of how the justices have ruled in related cases, what some notable conservative sources have written, and what the justices said in oral arguments. Finally, we offer our own thoughts about some potential outcomes of this case and their consequences for the future.
    Background
    What is the independent state legislature theory?
    Independent State Legislature theory or doctrine (ISL) generally holds that state legislatures have unique power to determine the rules around elections. There are a range of views that fall under the term ISL, ranging from the idea that state courts' freedom to interpret legislation is more limited than it is with other laws, to the idea that state courts and other state bodies lack any authority on issues of federal election law altogether. However, “[t]hese possible corollaries of the doctrine are largely independent of each other, supported by somewhat different lines of reasoning and authority. Although these theories arise from the same constitutional principle, each may be assessed separately from the others; the doctrine need not be accepted or repudiated wholesale.”1
    The doctrine is rooted in a narrow reading of Article I Section 4 Clause 1 (the Elections Clause) of the Constitution, which states, “The Times, Places and Manner of holding Elections for Senators and Representatives, shall be prescribed in each State by the Legislature thereof.”2 According to the Brennan Center, this interpretation is at odds with a more traditional reading:
    The dispute hinges on how to understand the word “legislature.” The long-running understanding is that it refers to each state’s general lawmaking processes, including all the normal procedures and limitations. So if a state constitution subjects legislation to being blocked by a governor’s veto or citizen referendum, election laws can be blocked via the same means. And state courts must ensure that laws for federal elections, like all laws, comply with their state constitutions.
    Proponents of the independent state legislature theory reject this traditional reading, insisting that these clauses give state legislatures exclusive and near-absolute power to regulate federal elections. The result? When it comes to federal elections, legislators would be free to violate the state constitution and state courts couldn’t stop them.
    Extreme versions of the theory would block legislatures from delegating their authority to officials like governors, secretaries of state, or election commissioners, who currently play important roles in administering elections.3
    The doctrine, which governs the actions of state cou...
    50 min
  • LW - A tension between two prosaic alignment subgoals by Alex Lawsen
    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: A tension between two prosaic alignment subgoals, published by Alex Lawsen on March 19, 2023 on LessWrong.
    Written quickly rather than not at all, as I've described this idea a few times and wanted to have something to point at when talking to people. 'Quickly' here means I was heavily aided by a language model while writing, which I want to be up-front about given recent discussion.
    BLUF
    In alignment research, two seemingly conflicting objectives arise: eliciting honest behavior from AI systems, and ensuring that AI systems do not produce harmful outputs. This tension is not simply a matter of contradictory training objectives; it runs deeper, creating potential risks even when models are perfectly trained never to utter harmful information.
    Tension
    Eliciting honest behavior in this context means developing techniques to extract AI systems' "beliefs", to the extent that they are well-described as having them. In other words, honest models should, if they have an internal world model, accurately report predictions or features of that world model. Incentivizing honesty in AI systems seems important in order to avoid and detect deceptive behavior. Additionally, something like this seems necessary for aiding with alignment research - we want to extract valuable predictions of genuine research breakthroughs, as opposed to mere imaginative or fictional content.
    On the other hand, avoiding harmful outputs entails training AI systems never to produce information that might lead to dangerous consequences, such as instructions for creating weapons that could cause global catastrophes.
    The tension arises not just because "say true stuff" and "sometimes don't say stuff" seem like objectives which will occasionally end up in direct opposition, but also because methods that successfully elicit honest behavior could potentially be used to extract harmful information from AI systems, even when they have been perfectly trained not to share such content. In this situation, the very techniques that promote honest behavior might also provide a gateway to accessing dangerous knowledge.
    Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org.
    3 min
  • EA - Some Comments on the Recent FTX TIME Article by Ben West
    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: Some Comments on the Recent FTX TIME Article, published by Ben West on March 20, 2023 on The Effective Altruism Forum.
    Background
    Alameda Research (AR) was a cryptocurrency hedge fund started in late 2017.
    In early 2018, approximately half the employees quit, including myself and Naia Bouscal, the main person mentioned in the TIME article. At the time, I had considered AR to have failed, and I think even the people who stayed would have agreed that it had not achieved what it had wanted to.
    Later in 2018, some of the remaining AR staff started working on a cryptocurrency exchange named FTX. FTX grew to become a multibillion-dollar company.
    In late 2022, FTX collapsed. It has since been alleged that FTX defrauded their investors by misrepresenting the relationship between AR and FTX, and that this effectively led to them stealing customer deposits.
    The recent TIME article doesn’t make a very precise argument; here is my attempt at steelmanning/clarifying a major argument made in that article, which I will then respond to:
    Some EAs worked at AR before FTX started
    Even though those EAs (including myself) quit before FTX was founded and therefore could not have had any first-hand knowledge of this improper relationship between AR and FTX, they knew things (like information about Sam’s character) which would have enabled them to predict that something bad would happen
    This information was passed on to “EA leaders”, who did not take enough preventative action and are therefore (partly) responsible for FTX’s collapse
    Personal Background
    I worked at Alameda Research (AR) for about three months in early 2018. I was not involved in stealing FTX customer funds, and hopefully people trust me about that claim, if only because I quit before FTX was founded.
    To make my COI clear: I left the company I founded to join AR; doing so was very costly to me; AR crashed and burned within a few months of me joining; I blamed this crashing and burning largely on Sam.
    People who know I had a bad experience at AR are sometimes surprised that I’m not on the “obviously Sam was obviously 100% evil” bandwagon. I’ve been wanting to write something but found it hard because there weren’t specific things I could react to, it was just some vague difference in vibes.
    So I appreciate the TIME article sharing some specific things that “EA Leaders” allegedly knew which the author suggests should have caused them to predict FTX’s fraud.
    My Experience at AR at a High Level
    I thought Sam was a bad CEO. I think he literally never prepared for a single one-on-one we had, his habit of playing video games instead of talking to you was “quirky” when he was a billionaire but aggravating when he was my manager, and my recollection is that Alameda made less money in the time I was there than if it had just simply bought and held bitcoin.
    But my opinion of Sam overall was more positive than the sense I get from the statements in the TIME article. (This is not very surprising, given that the TIME article consists of statements that were probably intentionally selected to be the worst possible thing the journalist could find someone to say about Sam.)
    It's hard to convey nuance in these posts, and I'm sure someone is going to interpret me as trying to defend Sam here. This is not what I’m trying to do, but I do think it’s worth trying to share my reflections to help others refine their models.
    Adding my personal experience to supplement some statements from the article
    But one of the people who did warn others about Bankman-Fried says that he openly wielded this power when challenged. “It was like, ‘I could destroy you,’” this person says. “Will and Holden would believe me over you. No one is going to believe you.”
    I don’t want to speak for this person, but my own experience was pretty different. For example: Sam was f...
    9 min
  • EA - Save the Date April 1st 2023 EAGatherTown: UnUnConference by Vaidehi Agarwalla
    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: Save the Date April 1st 2023 EAGatherTown: UnUnConference, published by Vaidehi Agarwalla on March 20, 2023 on The Effective Altruism Forum.
    We're excited to officially announce the very first EA UnUnConference! APPLY HERE.
    Naming What We Can, the most impactful post ever published on April 1st, have already volunteered to host a Q&A. We’re calling in the producers of the TV hit Impact Island, and would like to invite Peter Barnett to launch his new book What The Future Owes Us. The X-risk-Men Incubation Program is running an enlightening talks session.
    Location: Mars in Gathertown
    Date: April 1st, 2023, 24 hours and 37 minutes starting at 12:00pm UTC (or “lunch time” for british people)
    The case for impact
    Over the years, humanity realized that Unconferences are a great twist of traditional conferences, since the independence gives room for more unexpected benefits to happen.
    For the reason, we’re experimenting with the format of an UnUnconference. This means we’ll actively try not to organize anything, therefore (in expectancy) achieving even more unexpected benefits.
    We encourage you to critique our (relatively solid, in our opinion) theory of change in the comments!
    We understand this is not the most ambitious we could be. Although we fall short of the dream of EAGxMars, we believe this Ununconference is a proof-of-concept that will help validate the model of novel, experimental conferences and possibly redefine what impact means for EA events for years to come.
    This team is well placed to unorganize this event because we have previously successfully not organized 10^10 possible events.
    What to expect
    All beings welcomed, that includes infants, face mice, gut microbiome, etc.
    Expect to have the most impactful time
    Make more impact than everyone on earth could ever do combined
    Network with the best minds in ultra-near-termist research
    Never meet your connections again after the event
    Certificates of £20 worth of impact just for £10!
    No success metrics
    No theory of change
    No food, no wine, no suffering
    Check out our official event poster!
    Pixelated lightbulb that looks like mars as a logo for an unconference (DALL-E)
    Get involved
    Take a look at the confernce agenda and add sessions to your calendar
    Comment on this post with content suggestions and anti-suggestions
    Sign up for an enlightning talk
    Unvolunteer for the event
    UnUnvolunteer for the event (your goal will be to actively unorganize stuff)
    UnUnUnvolunteer for the event (your goal will be to actively ununorganize stuff)
    .. And so on. We think at least 5 levels of volunteers will be necessary for this event to be a complete success, to minimize risk of not falling into the well known meta trap.
    Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org.
    4 min
  • LW - RLHF does not appear to differentially cause mode-collapse by Arthur Conmy
    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: RLHF does not appear to differentially cause mode-collapse, published by Arthur Conmy on March 20, 2023 on LessWrong.
    Epistemic status: confident but not certain. This post is part of the work done at Conjecture.
    TL;DR: the results in Mysteries of mode collapse do not reproduce in text-davinci-003, a model trained with RLHF. In fact, there are cases where RLHF models exhibit higher entropy outputs than base models. We observe that the mode collapse phenomenon occurs more for the public OpenAI GPT-3 model trained with supervised finetuning (text-davinci-002) than RLHF, and present early experiments and theory to support this.
    Background
    Mysteries of mode collapse details how "mode collapse" (which we operationalize as a large increase in model output confidence and decreases in entropy of output distribution) arises more in text-davinci-002 than the base model davinci, and speculates about how this connects to RLHF training. At the time, OpenAI was very unclear on the training process for this model, and later (as @janus points out in the edited introduction to the post) it was revealed that this model was finetuned on highly-rated samples rather than trained with RLHF. However, the connection between RLHF and mode collapse has stuck, and several posts written since assume a connection.
    Results
    In this section, we compare the base model (davinci code-davinci-002, thanks commenters!) with the supervised fine-tuned model (text-davinci-002) and the RLHF model (text-davinci-003). We recommend trying some prompts for yourself in the OpenAI playground. The first result is that the mode collapse to “ 97” for the completion of the first prompt from @janus’ does not occur in the RLHF model:
    In fact, when we try another prompt we get that the base model has the lowest entropy:
    (ETA: this result is somewhat weaker than hoped, since text-davinci-002 seems to not output " 0" - " 100" here. davinci does exhibit collapses on other prompts, but commenters pointed out this is not the base model)
    The finding that mode collapse occurs in finetuned models is not robust. Comparing two of the prompts from the original post and two more, there is no noticeable pattern where the base model has higher entropy than the other models:
    (the uncertainty bars represent the maximum possible entropy if the model had uniform probability on all tokens other than “ 0”, . , “ 100” - the OpenAI API doesn't provide probabilities for all tokens)
    Reproducing the qualitative examples
    What about the other examples from the mode-collapse post? We found that the Blake Lemoine result was reproduced by davinci. On the Blake Lemoine greentext prompt with temperature 0.3, davinci gave completions where anon leaves after at most 5 lines. Most other results quickly led into repetitions of 3-4 sentences, something that occurred more frequently with the base language model.
    Overall, extrapolation from just the responses of one language model risks overstating conclusions, in this case about how unlikely the completion "leaving" was.
    Interpretation
    It appears as if the finetuning used for text-davinci-002 does cause mode collapses on the first two prompts. Arguably this is not surprising; RLHF training has a KL penalty to the base model’s outputs, which constrains the entropy of the RLHF model’s outputs to be close to that of the base model. Directly finetuning on new samples does not have this property since KL penalties to the base model are generally not so ubiquitous in standard finetuning (though lack of training details limits the conclusions that can be made here).
    Inferences about the phenomenon of mode collapse must be compatible with the evidence from both text-davinci-002 and text-davinci-003. For example, the author speculates that FeedME’s reliance on samples from RLHF models may be responsible for text-davi...
    9 min
  • EA - Tensions between different approaches to doing good by James Özden
    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: Tensions between different approaches to doing good, published by James Özden on March 19, 2023 on The Effective Altruism Forum.
    Link-posted from my blog here.
    TLDR: I get the impression that EAs don't always understand where certain critics are coming from e.g. what do people actually mean when they say EAs aren't pursuing "system change" enough? or that we're focusing on the wrong things? I feel like I hear these critiques a lot, so I attempted to steelman them and put them into more EA-friendly jargon. It's almost certainly not a perfect representation of these views, nor exhaustive, but might be interesting anyway. Enjoy!
    I feel lucky that I have fairly diverse groups of friends. On one hand, some of my closest friends are people I know through grassroots climate and animal rights activism, from my days in Extinction Rebellion and Animal Rebellion. On the other hand, I also spend a lot of time with people who have a very different approach to improving the world, such as friends I met through the Charity Entrepreneurship Incubation Program or via effective altruism.
    Both of these somewhat vague and undefined groups, “radical” grassroots activists and empirics-focused charity folks, often critique the other group with various concerns about their methods of doing good. Almost always, I end up defending the group under attack, saying they have some reasonable points and we would do better if we could integrate the best parts of both worldviews.
    To highlight how these conversations usually go (and clarify my own thinking), I thought I would write up the common points into a dialogue between two versions of myself. One version, labelled Quantify Everything James (or QEJ), discusses the importance of supporting highly evidence-based and quantitatively-backed ways of doing good. This is broadly similar to what most effective altruists advocate for. The other part of myself, presented under the label Complexity-inclined James (CIJ), discusses the limitations of this empirical approach, and how else we should consider doing the most good. With this character, I’m trying to capture the objections that my activist friends often have.
    As it might be apparent, I’m sympathetic to both of these different approaches and I think they both provide some valuable insights. In this piece, I focus more on describing the common critiques of effective altruist-esque ways of doing good, as this seems to be something that isn’t particularly well understood (in my opinion).
    Without further ado:
    Quantify Everything James (QEJ): We should do the most good by finding charities that are very cost-effective, with a strong evidence base, and support them financially! For example, organisations like The Humane League, Clean Air Task Force and Against Malaria Foundation all seem like they provide demonstrably significant benefits on reducing animal suffering, mitigating climate change and saving human lives. For example, external evaluators estimate the Against Malaria Foundation can save a human life for around $5000 and that organisations like The Humane League affect 41 years of chicken life per dollar spent on corporate welfare campaigns.
    It’s crucial we support highly evidence-based organisations such as these, as most well-intentioned charities probably don’t do that much good for their beneficiaries. Additionally, the best charities are likely to be 10-100x more effective than even the average charity! Using an example from this very relevant paper by Toby Ord: If you care about helping people with blindness, one option is to pay $40,000 for someone in the United States to have access to a guide dog (the costs of training the dog & the person). However, you could also pay for surgeries to treat trachoma, a bacterial infection that is the top cause of blindness worldwide. At around $20 per ...
    19 min
  • LW - The Natural State is Goodhart by devansh
    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 Natural State is Goodhart, published by devansh on March 20, 2023 on LessWrong.
    Epistemic Status: Meant to describe a set of beliefs that I have about accidental optimization pressures, and be a reference post for a thing I can refer back to later.
    Why do we live in worlds of bureaucracy and Lost Purpose? Because this is the default state of problem-solving, and everything else is an effortful push against Goodharting. Humans are all problem-solving machines, and if you want to experience inner misalignment inside your own brain, just apply anything less than your full attention to a metric you’re trying to push up.
    People claim to want things like more legroom, or comfier seats, or better service, or smaller chances of delays and cancellations. But when you actually sit down and book a flight, they are ordered by cost, and if you’re not a frequent flier then you generally choose the flight with the lowest sticker cost. This leads to a “race to the bottom” amongst airlines to push everything possible out of the sticker price and nickel-and-dime you—thereby causing the cheapest flights to actually be more expensive and worse.
    I was talking to a mentor of mine / giving her feedback and trying to work out how to best approach a problem. Sometimes I said things that she found helpful, and she noted these out loud. We then realized this disrupted conversation too much, so we changed to having her recognize my helpful sentences with a snap. This might have worked well, had I not immediately noticed my brain Goodharting towards extracting her snaps, instead of actually trying to figure out solutions to the problem and saying true things and improving my own models.
    There is a point that I’m trying to make here, which I think mostly fails to get made by the current writing on Goodhart’s law. It’s not just an explanation for the behavior of [people dumber than you]. Me, you, all of us, are constantly, 24/7. Goodharting towards whatever outcome fits our local incentives.
    This becomes even more true for groups of people and organizations. For example, EAG(x)s have a clear failure mode along this dimension. From reading retrospectives (EAGx Berkeley and EAGx Boston), they sure do seem to focus a lot on making meaningful connections and hyping people up about EA ideas and the community, and a lot of the retrospective is about how much people enjoyed EAG. I don't mean to call EAG out specifically, but instead to highlight a broader point - we’re not a religion trying to spread a specific gospel; we’re a bunch of people trying to figure out how to figure out what's true, and do things in the world that accomplish our goals. It does sure seem like we’re putting a bunch of optimization pressure into things that don’t really track our final goals, and we should step back and be at least concerned about this fact.
    Some parts of the rationality community do a similar thing. I notice a circuit in my own brain that Goodharts towards certain words / ways of speaking because they’re more “rational.” Like, I personally have adopted this language, but actually talking about “priors” and “updates” and appending “or something” to the end of sentences does not make you better at finding the truth. You’re not a better Bayesian reasoner purely because you use words that correspond to Bayesian thinking. (The counterargument here is the Sapir-Whorf hypothesis, which weakens but does not kill this point—I think many of the mannerisms seen as desirable by people in the rationality community and accepted as status or ingroup indicators track something different from truth.)
    By default we follow local incentives, and we should to be quite careful to step back every once in a while and really, properly make sure that we are optimizing for the right purposes. You should expect the autopilot that runs ...
    4 min
  • LW - Tabooing "Frame Control" by Raemon
    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: Tabooing "Frame Control", published by Raemon on March 19, 2023 on LessWrong.
    "Frame Control" is a colloquial term people have used to describe "Someone is doing something rhetorically fishy that somehow relates to frames." I think it's a fairly loaded phrase, and hasn't really been used consistently. I'm not sure we should actually use the phrase – it seems easy to weaponize in unhelpful ways. But it does seem like it's getting at something important that I want to understand and talk about.
    Aella's post on the topic focused on particularly abusive dynamics. I think abusive frame control is an important central example. But I think there are many times when "something rhetorically fishy is going on with frames", and it isn't particularly abusive but still is worth talking about.
    In this post I want to try and taboo frame control, as well as draw more of a distinction between "the cluster of patterns that is 'frame control'", and "the cluster of patterns that is 'abuse' and 'manipulate'."
    in practice, I still needed to refer to "the gestalt cluster of things that feel centrally 'frame control-y'" and I didn't have a better word for that than "frame control" although I tried to mostly put it in quotes.
    First, a quick recap on frames.
    A frame is a colloquial term for "what someone sees as important, what sort of questions they ask or what they're trying to get out of a conversation." I think it's often used in a fuzzy metaphorical way, and there are slightly different metaphors people were unconsciously using, including picture frames, window frames and frameworks.
    John Wentworth explores a more technical approach to frames in his post Shared Frames Are Capital Investments in Coordination. There, he defines a frame as way of conceptualizing a problem or solution space. A frame suggests which types of questions to ask, and which type of answers to look for.
    Previously, I've discussed how sometimes people have different assumptions about what frame they're in. The result can be annoying, confused conversations that take years to resolve. Noticing those different frames is an important communication skill.
    Okay. So what's "Frame Control?"
    People use "Frame control" differently. I assume they all roughly means, well, "someone is trying to control your frame". Possibly unconsciously, possibly deliberately, their actions are shaping what sort of questions you're able to ask and think about, and what you think is important.
    But, just as people had originally used the word "frame" in an ambiguous way that led to some confusion, I think people have used the phrase "frame control" inconsistently. I'm about to share my own ontology of "what concepts 'frame control' breaks down into." If you've experienced something-you-call-frame-control, you may want to take a moment to think through your own conceptions of it.
    (here is you having some space to think through your own experiences and ontology. Feel free to leave your own takes in the comments)
    When I reflect on the times something "frame-control-ish" has happened to me, four distinctions that strike me are:
    Holding a frame, at all. i.e. having a sense of how you're trying to think or communicate, and what sort of questions or goals you're trying to address. This is super normal and reasonable.
    Presenting a strongly held/presented frame, such as by speaking confidently/authoritatively (which many people who don't hold their own frames very strongly sometimes find disorienting)
    Persistently insisting on a frame. such that when someone tries to say/imply 'hey, my frame is X' you're like 'no, the frame is Y'. And if they're like 'no, it's X' you just keep talking in frame Y and make it socially awkward to communicate in frame X.
    Frame manipulation, where you change someone else's frame in a subtle way without them noticing, i.e. pres...
    16 min
  • LW - Probabilistic Payor Lemma? by abramdemski
    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: Probabilistic Payor Lemma?, published by abramdemski on March 19, 2023 on LessWrong.
    Epistemic status: too good to be true? Please check my math.
    We've known for a while that Löb's theorem fails when proof is relaxed to probabilistic belief. This has pros and cons. On the pro side, it means there's no Löbian Obstacle to probabilistic self-trust. On the con side, it means that some Löb-derived insights for proof-based decision theory don't translate to probabilistic decision theory, at least not as directly as one might hope. In particular, it appeared to dash hopes for probabilistic generalizations of the "Löbian handshake" for cooperation.
    Recently, Andrew Critch wrote about the Payor Lemma, which allows for a very similar "modal handshake" without Löb's Theorem. The lemma was proved using the same modal assumptions as Löb's, so on the surface it may appear to be just a different method to achieve similar results, whose main advantage is that it is much easier to prove (and therefore explain and understand) than Löb's Theorem.
    But, a natural question arises: does Payor's Lemma have a suitable probabilistic version?
    I'll give an affirmative proof; but I haven't confirmed that the assumptions are reasonable to my satisfaction.
    Setup
    Let L be a language in first-order logic, expressive enough to represent its sentences s∈L as quoted terms ┌s┐, eg, through Gödel numbering; and with a probability function symbol on these terms, p(┌s┐), which can be equated with (some representation of) rational numbers, e.g. p(┌⊤┐)=1, p(┌s┐)=12, etc. I also assume the system can reason about these rational numbers in the basic ways you'd expect.
    For all a,b∈L and all r∈Q, we have:
    If ⊢a, then ⊢p(┌a┐)=1.
    If ⊢ab, then ⊢p(┌a┐)≤p(┌b┐).
    (These assumptions might look pretty minimal, but they aren't going to be true for every theory of self-referential truth; more on this later.)
    Let B(s) abbreviate the sentence p(┌s┐)>c for any s and some globally fixed constant c strictly between 0 and 1. This is our modal operator.
    Some important properties of B:
    Necessitation. If ⊢s, then ⊢B(s), for any s.
    Proof: Since ⊢s implies ⊢p(s)=1, and c∈(0,1), we have ⊢p(┌s┐)>c,, which is to say, ⊢B(s). [End proof.]
    Weak distrubitivity. If ⊢xy, then ⊢B(x)B(y).
    Proof: When ⊢xy, we have ⊢p(y)≥p(x), so ⊢p(x)>cp(y)>c. [End proof.]
    (Regular distributivity would say B(xy) implies B(x)B(y). The assumption ⊢xy is stronger than B(xy), so the above is a weaker form of distributivity.)
    Theorem Statement
    If ⊢B(B(x)x)x, then ⊢x.
    Proof
    ⊢x(B(x)x), by tautology (a(ba)).
    So ⊢B(x)B(B(x)x), from 1 by weak distributivity.
    Suppose ⊢B(B(x)x)x.
    ⊢B(x)x from 2 and 3.
    ⊢B(B(x)x) from 4 by necessitation.
    ⊢x from 4 and 1.[End proof.]
    Discussion
    Comparison to Original Proof
    The proof steps mirror Critch's treatment very closely. The key difference is step 2, IE, how I obtain a statement like ⊢□x□(□xx). Critch uses distributivity, which is not available to me:
    B(ab)(B(a)B(b))?
    Suppose B(ab), ie, p(┌ab┐)>c.
    Rewrite p(┌b∨¬a┐)>c.
    Now suppose B(a), that is, p(┌a┐)>c.
    p(┌¬a┐)<1−c.
    p(┌b∨¬a┐)≤p(┌b┐)+p(┌¬a┐)p(┌b∨¬a┐)−1+c>c−1+c.
    p(┌b┐)>2c−1.
    So we only get:
    Bc(ab)(Bc(a)Bd(b)),
    where Br(s) abbreviates p(┌s┐)>r and we have d=2c−1.
    So in general, attempted applications of distributivity create weakened belief operators, which would get in the way of the proof (very similar to how probabilistic Löb fails).
    However, the specific application we want happens to go through, due to a logical relationship between a and b; namely, that b is a weaker statement than a.
    This reveals a way in which the assumptions for Payor's Lemma are importantly weaker than those required for Löb to go through.
    So, the key observation I'm making is that weak distributility is all that's needed for Payor, and seems much more plausible for self-...
    7 min
  • EA - Scale of the welfare of various animal populations by Vasco Grilo
    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: Scale of the welfare of various animal populations, published by Vasco Grilo on March 19, 2023 on The Effective Altruism Forum.
    Summary
    I Fermi-estimated the scale of the welfare of various animal populations from the relative intensity of their experiences, moral weight, and population size.
    Based on my results, I would be very surprised if the scale of the welfare of:
    Wild animals ended up being smaller than that of farmed animals.
    Farmed animals turned out to be smaller than that of humans.
    Introduction
    If it is worth doing, it is worth doing with made-up statistics?
    Methods
    I Fermi-estimated the scale of the welfare of various animal populations from the absolute value of the expected total hedonistic utility (ETHU). I computed this from the product between:
    Intensity of the mean experience as a fraction of that of the worst possible experience.
    Mean moral weight.
    Population size.
    The data and calculations are here.
    Intensity of experience
    I defined the intensity of the mean experience as a fraction of that of the worst possible experience based on the types of pain defined by the Welfare Footprint Project (WFP) here (search for “definitions”). I assumed:
    The following correspondence between the various types of pain (I encourage you to check this post from algekalipso, and this from Ren Springlea to get a sense of why I think the intensity can vary so much):
    Excruciating pain, which I consider the worst possible experience, is 1 k times as bad as disabling pain.
    Disabling pain is 100 times as bad as hurtful pain, which together with the above implies excruciating pain being 100 k times as bad as hurtful pain.
    Hurtful pain is 10 times as bad as annoying pain, which together with the above implies excruciating pain being 1 M times as bad as annoying pain.
    The intensity of the mean experience of:
    Farmed animal populations is as high as that of broiler chickens in reformed scenarios. I assessed this from the time broilers experience each type of pain according to these data from WFP (search for “pain-tracks”), and supposing:
    The rest of their time is neutral.
    Their lifespan is 42 days, in agreement with section “Conventional and Reformed Scenarios” of Chapter 1 of Quantifying pain in broiler chickens by Cynthia Schuck-Paim and Wladimir Alonso.
    Humans and other non-farmed animal populations is as high as 2/3 of that of hurtful pain. 2/3 (= 16/24) such that 1 day (24 h) of such intensity is equivalent to 16 h spent in hurtful pain plus 8 h in neutral sleeping.
    Ideally, I would have used empirical data for the animal populations besides farmed chickens too. However, I do not think they are readily available, so I had to make some assumptions.
    In general, I believe the sign of the mean experience is:
    For farmed animal populations, negative, judging from the research of WFP on chickens.
    For humans, positive (see here).
    For other non-farmed animal populations, positive or negative (see this preprint from Heather Browning and Walter Weit).
    Moral weight
    I defined the mean moral weight from Rethink Priorities’ median estimates for mature individuals provided here by Bob Fischer. For the populations I studied with animals of different species, I used those of:
    For wild mammals, pigs.
    For farmed fish, salmon.
    For wild fish, salmon.
    For farmed insects, silkworms.
    For wild terrestrial arthropods, silkworms.
    For farmed crayfish, crabs and lobsters, mean between crayfish and crabs.
    For farmed shrimps and prawns, shrimps.
    For wild marine arthropods, silkworms.
    For nematodes, silkworms multiplied by 0.1.
    Population size
    I defined the population size from:
    For humans, these data from Our World in Data (OWID) (for 2021).
    For wild mammals, the mean of the lower and upper bounds provided in section 3.1.5.2 of Carlier 2020.
    For farmed chickens and pigs, these data from OWID (for 2014).
    F...
    11 min