The Nonlinear Library

The Nonlinear Library

By The Nonlinear FundEducation
Download on the App Store

All The Nonlinear Library episodes

  • AF - Jesse Hoogland on Developmental Interpretability and Singular Learning Theory by Michaël Trazzi
    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: Jesse Hoogland on Developmental Interpretability and Singular Learning Theory, published by Michaël Trazzi on July 6, 2023 on The AI Alignment Forum.
    Jesse Hoogland is a research assistant at David Krueger's lab in Cambridge studying AI Safety who has recently been publishing on LessWrong about how to apply Singular Learning Theory to Alignment, and even organized some workshop in Berkeley last week around this.
    I thought it made sense to interview him to have some high-level overview of Singular Learning Theory (and other more general approaches like developmental interpretability).
    Below are some highlighted quotes from our conversation (available on Youtube, Spotify, Google Podcast, Apple Podcast). For the full context for each of these quotes, you can find the accompanying transcript.
    Interpreting Neural Networks: The Phase Transition View
    Studying Phase Transitions Could Help Detect Deception
    "We want to be able to know when these dangerous capabilities are first acquired because it might be too late. They might become sort of stuck and crystallized and hard to get rid of. And so we want to understand how dangerous capabilities, how misaligned values develop over the course of training. Phase transitions seem particularly relevant for that because they represent kind of the most important structural changes, the qualitative changes in the shape of these models internals.
    Now, beyond that, another reason we’re interested in phase transitions is that phase transitions in physics are understood to be a kind of point of contact between the microscopic world and the macroscopic world. So it’s a point where you have more control over the behavior of a system than you normally do. That seems relevant to us from a safety engineering perspective. Why do you have more control in a physical system during phase transitions?" (context)
    A Concrete Example of Phase Transition In Physics and an analogous example inside of neural networks
    "Jesse: If you heat a magnet to a high enough temperature, then it’s no longer a magnet. It no longer has an overall magnetization. And so if you bring another magnet to it, they won’t stick. But if you cool it down, at some point it reaches this Curie temperature. If you push it lower, then it will become magnetized. So the entire thing will all of a sudden get a direction. It’ll have a north pole and a south pole. So the thing is though, like, which direction will that north pole or south pole be? And so it turns out that you only need an infinitesimally small perturbation to that system in order to point it in a certain direction. And so that’s the kind of sensitivity you see, where the microscopic structure becomes very sensitive to tiny external perturbations.
    Michaël: And so if we bring this back to neural networks, if the weights are slightly different, the overall model could be deceptive or not. Is it something similar?
    Jesse: This is speculative. There are more concrete examples. So there are these toy models of superposition studied by Anthropic. And that’s a case where you can see that it’s learning some embedding and unembeddings. So it’s trying to compress data. You can see that the way it compresses data involves this kind of symmetry breaking, this sensitivity, where it selects one solution at a phase transition. So that’s a very concrete example of this." (context)
    Developmental Interpretability
    "Suppose it’s possible to understand what’s going on inside of neural networks, largely understand them. First assumption. Well then, it’s still going to be very difficult to do that at one specific moment in time. I think intractable. The only way you’re actually going to build up an exhaustive idea of what structure the model has internally, is to look at how it forms over the course of training. You want to look at each moment, where...
    6 min
  • LW - AI labs' statements on governance by Zach Stein-Perlman
    Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: AI labs' statements on governance, published by Zach Stein-Perlman on July 4, 2023 on LessWrong.
    This is a collection of statements on government policy, regulation, and standards from leading AI labs and their leadership.
    I believe this post has all of the relevant announcements/blogposts from the three labs it covers, but I expect it is missing a couple relevant speeches/interviews with lab leadership. I may not keep this post up to date. Suggestions are welcome.
    My quotes tend to focus on AI safety rather than other governance goals.
    Within sections, sources are roughly sorted by priority.
    OpenAI
    Governance of superintelligence (May 2023)
    First, we need some degree of coordination among the leading development efforts to ensure that the development of superintelligence occurs in a manner that allows us to both maintain safety and help smooth integration of these systems with society. There are many ways this could be implemented; major governments around the world could set up a project that many current efforts become part of, or we could collectively agree (with the backing power of a new organization like the one suggested below) that the rate of growth in AI capability at the frontier is limited to a certain rate per year.
    And of course, individual companies should be held to an extremely high standard of acting responsibly.
    Second, we are likely to eventually need something like an IAEA for superintelligence efforts; any effort above a certain capability (or resources like compute) threshold will need to be subject to an international authority that can inspect systems, require audits, test for compliance with safety standards, place restrictions on degrees of deployment and levels of security, etc. Tracking compute and energy usage could go a long way, and give us some hope this idea could actually be implementable. As a first step, companies could voluntarily agree to begin implementing elements of what such an agency might one day require, and as a second, individual countries could implement it. It would be important that such an agency focus on reducing existential risk and not issues that should be left to individual countries, such as defining what an AI should be allowed to say.
    Planning for AGI and beyond (Feb 2023)
    We think it's important that efforts like ours submit to independent audits before releasing new systems; we will talk about this in more detail later this year. At some point, it may be important to get independent review before starting to train future systems, and for the most advanced efforts to agree to limit the rate of growth of compute used for creating new models. We think public standards about when an AGI effort should stop a training run, decide a model is safe to release, or pull a model from production use are important. Finally, we think it's important that major world governments have insight about training runs above a certain scale.
    Altman Senate testimony (May 2023)
    Written testimony (before the hearing)
    There are several areas I would like to flag where I believe that AI companies and governments can partner productively.
    First, it is vital that AI companies-especially those working on the most powerful models-adhere to an appropriate set of safety requirements, including internal and external testing prior to release and publication of evaluation results. To ensure this, the U.S. government should consider a combination of licensing or registration requirements for development and release of AI models above a crucial threshold of capabilities, alongside incentives for full compliance with these requirements.
    Second, AI is a complex and rapidly evolving field. It is essential that the safety requirements that AI companies must meet have a governance regime flexible enough to adapt to new technical developments. The U.S...
    47 min
  • LW - Optimized for Something other than Winning or: How Cricket Resists Moloch and Goodhart's Law by A.H.
    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: Optimized for Something other than Winning or: How Cricket Resists Moloch and Goodhart's Law, published by A.H. on July 5, 2023 on LessWrong.
    Recently, there has been some controversy in the world of cricket. The full details are described well in this piece but the gist of it is as follows. During a game between Australia and England, after the ball had been bowled, the English batsman Jonny Bairstow stepped forward thinking that the game was no longer in play. This allowed the Australian wicketkeeper Alex Carey to throw the ball at the stumps and get Bairstow out. As a result, Bairstow had to leave the field and later that day, Australia won the game. Would England have won were it not for this piece of quick thinking from Carey? It's hard to say, but the game certainly would have been a lot closer.
    So why is this controversial? After all, Carey's actions were completely within the rules of the game and all of the umpires and players agreed that Bairstow indeed out. The controversy stems from a nebulous concept known as 'the spirit of cricket'. The 'spirit of cricket' is mentioned in the preamble to the official Laws of Cricket and broadly encompasses good sportsmanship and fair play. But the fact that it is not properly defined means that different people will have different views on what kind of actions are within the spirit of the game. Pat Cummins, the Australian captain believes that Carey's actions were within the spirit of the game, and the Ben Stokes, the English captain believes that they were not. Various media outlets have taken sides on this issue.
    To be clear, the 'spirit of cricket' is not just 'good sportsmanship applied when convenient'. Players often make choices that are within the spirit of the game which put their team at a severe disadvantage. Indian player MS Dhoni famously withdrew his appeal against Ian Bell. Within the rules of the game Bell had been run out in a moment of confusion but Dhoni believed that it was not within the spirit of the game, and let Bell come back to bat. The spirit of the game is something that spectators and players genuinely care about. Certain historical cricketing episodes, such as the underarm incident, and the bodyline tour were within the rules but still evoke strong emotions to this day.
    Other tactics, like Mankading (where a bowler hits the stumps during his run-up to 'run out' an over-eager batsman who has started running) are allowed by the rules but rarely occur, though they still happen occasionally, generating controversy when they do and exist in a grey area with respect to 'the spirit'.
    Isn't this stupid? To have an extra set of 'rules' which aren't really rules and everyone disagrees on what they actually are and you can choose to ignore them and still win the game? I would argue that, in fact, it is this strange, poorly-defined concept of 'the spirit of cricket' which makes cricket entertaining to watch and helps it avoid pitfalls present in many other sports.
    Take football (soccer) which is arguably the most popular spectator sport worldwide. Presently, one of the key ways that players can get an advantage for their team is to be 'fouled' by a member of the opposing team. Examples of 'fouls' include tripping or pushing. In order to convince the referee that they have been fouled, football players will often dramatically fall to the floor, screaming and clutching some body part, only to get up moments later, seemingly unhurt. While 'diving' (attempting to deceive the referee by pretending to be fouled) is against the rules of football, it is not against the rules to draw attention to a potential foul by exaggerating your injury. As a result, many high-level games become competitions as to who can get the referee to award them a penalty or free kick and players end up spending significant amounts of ...
    8 min
  • LW - (tentatively) Found 600+ Monosemantic Features in a Small LM Using Sparse Autoencoders by Logan Riggs
    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: (tentatively) Found 600+ Monosemantic Features in a Small LM Using Sparse Autoencoders, published by Logan Riggs on July 5, 2023 on LessWrong.
    Using a sparse autoencoder, I present evidence that the resulting decoder (aka "dictionary") learned 600+ features for Pythia-70M layer_2's MLP (after the GeLU), although I expect around 8k-16k features to be learnable.
    Dictionary Learning: Short Explanation
    Good explanation here & original here, but in short: a good dictionary means that you could give me any input & I can reconstruct it using a linear combination of dictionary elements. For example, signals can be reconstructed as a linear combination of frequencies:
    In the same way, the neuron activations in a large language model (LLM) can be reconstructed as a linear combination of features. e.g.
    neuron activations = 4([duplicate token] feature) + 7(bigram " check please" feature).
    Big Picture: If we learn all the atomic features that make up all model behavior, then we can pick & choose the features we want (e.g. honesty)
    To look at the autoencoder:
    So for every neuron activation (ie a 2048-sized vector), the autoencoder is trained to encode a sparse set of feature activations/magnitudes (sparse as in only a few features "activate" ie have non-zero magnitudes), which are then multiplied by their respective feature vector (ie a row in the decoder/"dictionary") in order to reconstruct the original neuron activations.
    As an example, an input is " Let u be f(8). Let w", and the decomposed linear combination of features are:
    [" Let u be f(8). Let w"] = 4(letter "w" feature) + 7(bigram "Let [x/w/n/p/etc]" feature)
    Note: The activation here is only for the last token " w" given the previous context. In general you get 2048 neuron activations for every token, but I'm just focusing on the last token in this example.
    For the post, it's important to understand that:
    Features have both a feature vector (ie the 2048-sized vector that is a row in the decoder) & a magnitude (ie a real number calculated on an input-by-input basis). Please ask questions in the comments if this doesn't make sense, especially after reading the rest of the post.
    I calculate Max cosine similarity (MCS) between the feature vectors in two separately trained dictionaries. So if feature0 in dictionary 0 is "duplicate tokens", and feature98 in dictionary 1 has high cosine similarity, then I expect both to be representing "duplicate tokens" and for this to be a "real" feature [Intuition: there are many ways to be wrong & only one way to be right]
    Feature Case Study
    Top-activating Examples for feature #52
    I ran through 500k tokens & found the ones that activated each feature. In this case, I chose feature52 which had an MCS of ~0.99. We can then look at the datapoints that maximally activate this feature. (To be clear, I am running datapoints through Pythia-70M, grabbing the activations mid-way through at layer 2's MLP after the GeLU, & running that through the autoencoder, grabbing the feature magnitudes ie latent activations).
    Blue here means this feature is activating highly for that (token,context)-pair. The top line has activation 3.5 for the first " $" & 5 for the second. Note that it doesn't activate for closing $ and is generally sparse.
    Ablate Context
    Ablate the context one token at a time & see the effect on the feature magnitudes on last token. Red means the feature activated less when ablating that token.
    To be clear, I am literally removing the token & running it through the model again to see the change in feature activation at the last position. In the top line, removing the last " $" makes the new last token to be " all" which has an activation of 0, so the difference is 0-5 = -5, which is value assigned to the dark red value on " $" (removing " all" before it had no effect, so it's white which mean...
    12 min
  • LW - [Linkpost] Introducing Superalignment by beren
    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: [Linkpost] Introducing Superalignment, published by beren on July 5, 2023 on LessWrong.
    Some quotes:
    Our approach
    Our goal is to build a roughly human-level automated alignment researcher. We can then use vast amounts of compute to scale our efforts, and iteratively align superintelligence.
    To align the first automated alignment researcher, we will need to 1) develop a scalable training method, 2) validate the resulting model, and 3) stress test our entire alignment pipeline:
    To provide a training signal on tasks that are difficult for humans to evaluate, we can leverage AI systems to assist evaluation of other AI systems (scalable oversight). In addition, we want to understand and control how our models generalize our oversight to tasks we can’t supervise (generalization).
    To validate the alignment of our systems, we automate search for problematic behavior (robustness) and problematic internals (automated interpretability).
    Finally, we can test our entire pipeline by deliberately training misaligned models, and confirming that our techniques detect the worst kinds of misalignments (adversarial testing).
    We expect our research priorities will evolve substantially as we learn more about the problem and we’ll likely add entirely new research areas. We are planning to share more on our roadmap in the future.
    While this is an incredibly ambitious goal and we’re not guaranteed to succeed, we are optimistic that a focused, concerted effort can solve this problem:C
    There are many ideas that have shown promise in preliminary experiments, we have increasingly useful metrics for progress, and we can use today’s models to study many of these problems empirically.
    Ilya Sutskever (cofounder and Chief Scientist of OpenAI) has made this his core research focus, and will be co-leading the team with Jan Leike (Head of Alignment). Joining the team are researchers and engineers from our previous alignment team, as well as researchers from other teams across the company.
    Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org.
    3 min
  • AF - [Linkpost] Introducing Superalignment by Beren Millidge
    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: [Linkpost] Introducing Superalignment, published by Beren Millidge on July 5, 2023 on The AI Alignment Forum.
    Some quotes:
    Our approach
    Our goal is to build a roughly human-level automated alignment researcher. We can then use vast amounts of compute to scale our efforts, and iteratively align superintelligence.
    To align the first automated alignment researcher, we will need to 1) develop a scalable training method, 2) validate the resulting model, and 3) stress test our entire alignment pipeline:
    To provide a training signal on tasks that are difficult for humans to evaluate, we can leverage AI systems to assist evaluation of other AI systems (scalable oversight). In addition, we want to understand and control how our models generalize our oversight to tasks we can’t supervise (generalization).
    To validate the alignment of our systems, we automate search for problematic behavior (robustness) and problematic internals (automated interpretability).
    Finally, we can test our entire pipeline by deliberately training misaligned models, and confirming that our techniques detect the worst kinds of misalignments (adversarial testing).
    We expect our research priorities will evolve substantially as we learn more about the problem and we’ll likely add entirely new research areas. We are planning to share more on our roadmap in the future.
    While this is an incredibly ambitious goal and we’re not guaranteed to succeed, we are optimistic that a focused, concerted effort can solve this problem:C
    There are many ideas that have shown promise in preliminary experiments, we have increasingly useful metrics for progress, and we can use today’s models to study many of these problems empirically.
    Ilya Sutskever (cofounder and Chief Scientist of OpenAI) has made this his core research focus, and will be co-leading the team with Jan Leike (Head of Alignment). Joining the team are researchers and engineers from our previous alignment team, as well as researchers from other teams across the company.
    Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org.
    3 min
  • AF - (tentatively) Found 600+ Monosemantic Features in a Small LM Using Sparse Autoencoders by Logan Riggs Smith
    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: (tentatively) Found 600+ Monosemantic Features in a Small LM Using Sparse Autoencoders, published by Logan Riggs Smith on July 5, 2023 on The AI Alignment Forum.
    Using a sparse autoencoder, I present evidence that the resulting decoder (aka "dictionary") learned 600+ features for Pythia-70M layer_2's MLP (after the GeLU), although I expect around 8k-16k features to be learnable.
    Dictionary Learning: Short Explanation
    Good explanation here & original here, but in short: a good dictionary means that you could give me any input & I can reconstruct it using a linear combination of dictionary elements. For example, signals can be reconstructed as a linear combination of frequencies:
    In the same way, the neuron activations in a large language model (LLM) can be reconstructed as a linear combination of features. e.g.
    neuron activations = 4([duplicate token] feature) + 7(bigram " check please" feature).
    Big Picture: If we learn all the atomic features that make up all model behavior, then we can pick & choose the features we want (e.g. honesty)
    To look at the autoencoder:
    So for every neuron activation (ie a 2048-sized vector), the autoencoder is trained to encode a sparse set of feature activations/magnitudes (sparse as in only a few features "activate" ie have non-zero magnitudes), which are then multiplied by their respective feature vector (ie a row in the decoder/"dictionary") in order to reconstruct the original neuron activations.
    As an example, an input is " Let u be f(8). Let w", and the decomposed linear combination of features are:
    [" Let u be f(8). Let w"] = 4(letter "w" feature) + 7(bigram "Let [x/w/n/p/etc]" feature)
    Note: The activation here is only for the last token " w" given the previous context. In general you get 2048 neuron activations for every token, but I'm just focusing on the last token in this example.
    For the post, it's important to understand that:
    Features have both a feature vector (ie the 2048-sized vector that is a row in the decoder) & a magnitude (ie a real number calculated on an input-by-input basis). Please ask questions in the comments if this doesn't make sense, especially after reading the rest of the post.
    I calculate Max cosine similarity (MCS) between the feature vectors in two separately trained dictionaries. So if feature0 in dictionary 0 is "duplicate tokens", and feature98 in dictionary 1 has high cosine similarity, then I expect both to be representing "duplicate tokens" and for this to be a "real" feature [Intuition: there are many ways to be wrong & only one way to be right]
    Feature Case Study
    Top-activating Examples for feature #52
    I ran through 500k tokens & found the ones that activated each feature. In this case, I chose feature52 which had an MCS of ~0.99. We can then look at the datapoints that maximally activate this feature. (To be clear, I am running datapoints through Pythia-70M, grabbing the activations mid-way through at layer 2's MLP after the GeLU, & running that through the autoencoder, grabbing the feature magnitudes ie latent activations).
    Blue here means this feature is activating highly for that (token,context)-pair. The top line has activation 3.5 for the first " $" & 5 for the second. Note that it doesn't activate for closing $ and is generally sparse.
    Ablate Context
    Ablate the context one token at a time & see the effect on the feature magnitudes on last token. Red means the feature activated less when ablating that token.
    To be clear, I am literally removing the token & running it through the model again to see the change in feature activation at the last position. In the top line, removing the last " $" makes the new last token to be " all" which has an activation of 0, so the difference is 0-5 = -5, which is value assigned to the dark red value on " $" (removing " all" before it had no effect, so it...
    12 min
  • EA - An Impact Roadmap by Aaron Boddy
    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: An Impact Roadmap, published by Aaron Boddy on July 5, 2023 on The Effective Altruism Forum.
    One of the things that surprised me when trying to develop potential interventions for Shrimp Welfare Project, was trying to understand a clear pathway from A (wanting to help shrimps cost-effectively) to Z (having a demonstrably impactful intervention scaling-up).
    As a result of this, I spent a lot of time reading up on and learning about Programme Development Methodologies - trying to Frankenstein something together that I thought would make sense for us. I want to share this work in case it’s useful to other NGOs (in particular for NGOs in contexts similar to ours). I hope it’s useful to you
    A quick note - Though this Roadmap is written in such a way as to outline the steps from A to B to C, etc. in reality, some of these steps may run in parallel (or at least, overlap in some ways). For example, one member of your team may be conducting an Evidence Review, while others are preparing to do some on-the-ground Needs Assessment work. In particular, throughout the Roadmap, you’ll likely continue to revisit and reflect on your Theory of Change, as well as updating and refining your Impact Monitoring.
    I also wrote some "Appendix posts" to complement this one that some people may find useful:
    Programme Development Methodologies - Summaries of all the existing methodologies I used as a reference when putting this Roadmap together
    Decision-Making Tools - These can be referred to throughout the Roadmap when a decision needs to be made
    Creative-Thinking Tools - Useful tools to help you brainstorm
    Executive Summary
    1. Key Decisions - The most important decisions we’ll make, and how to make them
    There are some useful Decision-Making and Creative-Thinking that it makes sense to learn early on, as they can be used throughout the roadmap. Not least, to help you with some of the most important decisions you’ll make on the path to impact, such as defining your values and choosing your career area based on personal fit. If starting a nonprofit is the best fit for you, then we can move on to deciding on a charity idea, finding a co-founder and selecting your (initial) implementation country. We can now begin to more concretely understand our path to impact through an (initial) Theory of Change.
    2. Theory of Change - How and why the program will work
    Now that you’ve made your key decisions, you can move on to defining your potential program. You’ll do this by first outlining some Project Objectives, before creating your first (likely somewhat broad at this point) Theory of Change diagram, sketching out the Activities you will undertake which lead to your projects Outputs, resulting in the overall positive Outcomes of your project, which ultimately achieves Impact. This will provide some clarity around the impact you intend to make, as well as the pathway to that impact. It’s likely though that there are some gaps in your knowledge and some uncertainties and assumptions in your Theory of Change, which brings us to an Evidence Review.
    3. Evidence Review - A critical review of secondary sources related to our program
    We now have an idea of how we intend to make an impact. It’s time to review any evidence that could help us more concretely answer some of our uncertainties and clarify our assumptions. To do this, we’ll undertake a literature review to help us answer these questions (i.e. What interventions have been tested or studied in relation to this program? Under what conditions were they implemented? Were they effective? Why or why not?). Still, however, we will need to get a better understanding of our chosen context, so we’ll go on to undertake a Needs Assessment.
    4. Needs Assessment - Seeking out perspectives from the populations we aim to help
    We now have a good overview of the relevant e...
    39 min
  • LW - Dominant Assurance Contract Experiment #2: Berkeley House Dinners by Arjun Panickssery
    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: Dominant Assurance Contract Experiment #2: Berkeley House Dinners, published by Arjun Panickssery on July 5, 2023 on LessWrong.
    As an afterthought at the end of my previous post on dominant assurance contracts, I decided on a whim as I was typing to append a small dominant assurance contract:
    In theory, writers could kickstart posts using dominant assurance contracts. An example (this is a real offer): If you send $20 to arjun.panickssery at Gmail via PayPal by noon New York time on January 21st, I'll send you back $25 if fewer than 10 people sent me money. If 10 or more people send me money, I'll post a review of Steven Pinker's The Sense of Style: The Thinking Person's Guide to Writing in the 21st Century by the end of the month. I'm not sure whether I'm just giving away free money right now.
    I received exactly nine out of the ten funders needed and lost $45.
    Now a second attempt, this time in the real world: at Andromeda House we plan to host large weekly dinners for the local EA/rationality/etc community at our house in Southside Berkeley.
    Here’s how it works:
    You can Venmo me (@Arjun-Panickssery if the link doesn't work) or PayPal me any amount of at least $20 with the subject line "dinner" or similar.
    If I get at least $700 total by noon Pacific time on July 15, I'll host dinners from July 17 till the end of August (seven dinners).
    If I get less than $700 total, I'll give you a 25% return (e.g., if you sent me $100, I'll send you back $125).
    Example: If you live in town and this service would be worth more than $20 per dinner to you and you'd expect to come to three before September, you should pay at least $60, since you either get a service worth more than that amount or you make a 25% return. You could try to free-ride but the fact that my previous contract received just one less than the target number suggests that I'm well-calibrated.
    If you don't live in town—or even if you do—you can also idly speculate if you think that I'm unlikely to hit my target and likely to pay out.
    You could also just pseudo-donate money without regard to whether you think I'm calibrated just because you think that DACs are cool and should proliferate.
    Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org.
    3 min
  • LW - Twitter Twitches 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: Twitter Twitches, published by Zvi on July 4, 2023 on LessWrong.
    The situation is evolving rapidly. Here’s where we stand as of the morning of July 4th.
    Well You See What Happened Was.
    Oh no! To be clear, by twitches, I mean ‘Elon refused to pay the cloud bill.’
    As a result, Twitter has been forced to rate limit users.
    This started out as 600 posts per day for most accounts, 300 posts per day for new accounts and 6,000 posts per day for those who pay.
    This is now up to 1k/500/10k according to one Musk tweet.
    If you are not logged in, you get nothing. Even direct links will break.
    Tweetdeck has been forced into a new worse version, but now works again. In 30 days, this will be for paid accounts only, which seems fair.
    That fourth one hurts my process. Navigation is somewhat slower and more annoying. In particular, forced threading breaks chronological order assumptions and one’s ability to use duplication to locate one’s place, and zooming in to move around twisting Twitter threads is so bad you need to jump to Twitter itself. Navigation to zoom back requires clicking in annoying places. I was unable to configure the column order without deleting them all and then creating them again, although this was quick. Column width and monitor real estate use is screwed up in subtle ways. Oh, and now its settings are linked to Twitter’s even though I want them to be different. Sheesh.
    Another little thing is that the tab icon is now identical to Twitter’s. So annoying.
    This is still vastly better than the period where Tweetdeck stopped working.
    The third is brutal for some of my readers. Many report they can’t view any links.
    What to do, if this doesn’t end soon?
    The Plan
    Three parts: How I will deal with processing info, how I will change how I present info, and how you can adjust to the new situation.
    The efficiency hit on my end is unavoidable. I’ll make three adjustments.
    I’ll check Twitter less often, rely more on other sources.
    I’ll raise the bar somewhat for what is worth including or investigating.
    I’ll shift some people from lists to follows and eliminate duplicates.
    The inability for viewers to see Twitter links means moving away from them.
    I’ll more often quote more of the relevant Twitter threads.
    I’ll link directly to the primary source first whenever possible.
    If you don’t have a Twitter account, you have a choice to make.
    Having an account with zero connections fixes links. Consider that.
    If you simply can’t do that responsibly, so be it, and I’ll write the posts such that this does not cause any showstopper issues.
    Most of you almost never clicked the links anyway.
    What I will not do is join BlueSky or another Twitter alternative. We are still far from where that would make sense.
    Also, clarifying some policies on how Twitter threads work here.
    I will silently correct capitalization and spelling and other obvious errors.
    I will reformat as makes sense for the new context, again silently, if the words themselves need to change I will use [brackets].
    I use your real name if I know it, your handle if that’s all I have.
    I will use judgment on ordering and when to provide detail on structure.
    For a quote tweet, I use the order that seems right to me – sometimes the intention is that the reader will see the quoted tweet only after, despite it happening earlier in time.
    For multiple different replies, I will make that clear if it is ambiguous.
    A link generally indicates a new thread of some kind.
    Oh No, That’s Not the Reason, Except It Is
    This Maggie Johnson-Pint thread offers a cool explanation of what else might have caused this situation, written before we knew the answer.
    Twitter owner Musk explains, with a different justification than the real one:
    Twitter Daily News: NEWS: Twitter’s web version no longer allows users to browse without logging in. All urls redirec...
    11 min