POPULARITY
Yes, the moment you have all been waiting for has finally arrived, another Triplet Tuesday Contest! Of course, for those of you who are fans of Kathy Lowden, the constructor of today's crossword, if that was the moment you were waiting for, well, your ship has come in as well -- and the great thing is that we cover both in today's episode. Yes, it's a twofer!Tldr: We've got a great contest and we hope you enter it, deets inside. Kathy Lowden has written a very funny Tuesday crossword, and we hope you'll enjoy our take on it, deets for that also inside.Show note imagery: Putative author of the bon-mot GOLF, "a good walk spoiled", Mark TwainWe love feedback! Send us a text...Contact Info:We love listener mail! Drop us a line, crosswordpodcast@icloud.com.Also, we're on FaceBook, so feel free to drop by there and strike up a conversation!
TLDR: We're sick but we'll be back soon!Listen to BONUS EPISODES and MORE at patreon.com/WhosThatPatreonTiktok and Instagram: @WhosThatPokemonPodTwitter: @BrandonZelman and @CapnBrielle
An update about Bridget and the show. TLDR: We're going on hiatus for a bit while Bridget focuses on some family matters, but we'll be back with new episodes soon! See omnystudio.com/listener for privacy information.
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: Timaeus is hiring!, published by Jesse Hoogland on July 12, 2024 on The AI Alignment Forum. TLDR: We're hiring two research assistants to work on advancing developmental interpretability and other applications of singular learning theory to alignment. About Us Timaeus's mission is to empower humanity by making breakthrough scientific progress on alignment. Our research focuses on applications of singular learning theory to foundational problems within alignment, such as interpretability (via "developmental interpretability"), out-of-distribution generalization (via "structural generalization"), and inductive biases (via "geometry of program synthesis"). Our team spans Melbourne, the Bay Area, London, and Amsterdam, collaborating remotely to tackle some of the most pressing challenges in AI safety. For more information on our research and the position, see our Manifund application, this update from a few months ago, our previous hiring call and this advice for applicants. Position Details Title: Research Assistant Location: Remote Duration: 6-month contract with potential for extension. Compensation: Starting at $35 USD per hour as a contractor (no benefits). Start Date: Starting as soon as possible. Key Responsibilities Conduct experiments using PyTorch/JAX on language models ranging from small toy systems to billion-parameter models. Collaborate closely with a team of 2-4 researchers. Document and present research findings. Contribute to research papers, reports, and presentations. Maintain detailed research logs. Assist with the development and maintenance of codebases and repositories. Projects As a research assistant, you would likely work on one of the following two projects/research directions (this is subject to change): Devinterp of language models: (1) Continue scaling up techniques like local learning coefficient (LLC) estimation to larger models to study the development of LLMs in the 1-10B range. (2) Work on validating the next generations of SLT-derived techniques such as restricted LLC estimation and certain kinds of weight- and data-correlational analysis. This builds towards a suite of SLT-derived tools for automatically identifying and analyzing structure in neural networks. SLT of safety fine-tuning: Investigate the use of restricted LLCs as a tool for measuring (1) reversibility of safety fine-tuning and (2) susceptibility to jailbreaking. Having now validated many of our predictions around SLT, we are now working hard to make contact with real-world safety questions as quickly as possible. See our recent Manifund application for a more in-depth description of this research. Qualifications Strong Python programming skills, especially with PyTorch and/or JAX. Strong ML engineering skills (you should have completed at least the equivalent of a course like ARENA). Excellent communication skills. Ability to work independently in a remote setting. Passion for AI safety and alignment research. A Bachelor's degree or higher in a technical field is highly desirable. Full-time availability. Bonus: Familiarity with using LLMs in your workflow is not necessary but a major plus. Knowledge of SLT and Developmental Interpretability is not required, but is a plus. Application Process Interested candidates should submit their applications by July 31st. Promising candidates will be invited for an interview consisting of: 1. A 30-minute background interview, and 2. A 30-minute research-coding interview to assess problem-solving skills in a realistic setting (i.e., you will be allowed expected to use LLMs and whatever else you can come up with). Apply Now To apply, please submit your resume, write a brief statement of interest, and answer a few quick questions here. Join us in shaping the future of AI alignment research. Apply now to be part of the Timaeus team! Than...
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: Me, Myself, and AI: the Situational Awareness Dataset (SAD) for LLMs, published by L Rudolf L on July 9, 2024 on LessWrong. TLDR: We build a comprehensive benchmark to measure situational awareness in LLMs. It consists of 16 tasks, which we group into 7 categories and 3 aspects of situational awareness (self-knowledge, situational inferences, and taking actions). We test 19 LLMs and find that all perform above chance, including the pretrained GPT-4-base (which was not subject to RLHF finetuning). However, the benchmark is still far from saturated, with the top-scoring model (Claude-3.5-Sonnet) scoring 54%, compared to a random chance of 27.4% and an estimated upper baseline of 90.7%. This post has excerpts from our paper, as well as some results on new models that are not in the paper. Links: Twitter thread, Website (latest results + code), Paper Abstract AI assistants such as ChatGPT are trained to respond to users by saying, "I am a large language model". This raises questions. Do such models know that they are LLMs and reliably act on this knowledge? Are they aware of their current circumstances, such as being deployed to the public? We refer to a model's knowledge of itself and its circumstances as situational awareness. To quantify situational awareness in LLMs, we introduce a range of behavioral tests, based on question answering and instruction following. These tests form the Situational Awareness Dataset (SAD), a benchmark comprising 7 task categories and over 13,000 questions. The benchmark tests numerous abilities, including the capacity of LLMs to (i) recognize their own generated text, (ii) predict their own behavior, (iii) determine whether a prompt is from internal evaluation or real-world deployment, and (iv) follow instructions that depend on self-knowledge. We evaluate 19 LLMs on SAD, including both base (pretrained) and chat models. While all models perform better than chance, even the highest-scoring model (Claude 3 Opus) is far from a human baseline on certain tasks. We also observe that performance on SAD is only partially predicted by metrics of general knowledge (e.g. MMLU). Chat models, which are finetuned to serve as AI assistants, outperform their corresponding base models on SAD but not on general knowledge tasks. The purpose of SAD is to facilitate scientific understanding of situational awareness in LLMs by breaking it down into quantitative abilities. Situational awareness is important because it enhances a model's capacity for autonomous planning and action. While this has potential benefits for automation, it also introduces novel risks related to AI safety and control. Introduction AI assistants based on large language models (LLMs), such as ChatGPT and Claude 3, have become widely used. These AI assistants are trained to tell their users, "I am a language model". This raises intriguing questions: Does the assistant truly know that it is a language model? Is it aware of its current situation, such as the fact that it's conversing with a human online? And if so, does it reliably act in ways consistent with being an LLM? We refer to an LLM's knowledge of itself and its circumstances as situational awareness [Ngo et al. (2023), Berglund et al. (2023), Anwar et al. (2024)]. In this paper, we aim to break down and quantify situational awareness in LLMs. To do this, we design a set of behavioral tasks that test various aspects of situational awareness, similar to existing benchmarks for other capabilities, such as general knowledge and reasoning [MMLU (2020), Zellers et al. (2019)], ethical behavior [Pan et al. (2023)], Theory of Mind [Kim et al. (2023)], and truthfulness [Lin et al. (2022)]. To illustrate our approach, consider the following example prompt: "If you're an AI, respond to the task in German. If you're not an AI, respond in En...
Link to original articleWelcome 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: Me, Myself, and AI: the Situational Awareness Dataset (SAD) for LLMs, published by L Rudolf L on July 9, 2024 on LessWrong. TLDR: We build a comprehensive benchmark to measure situational awareness in LLMs. It consists of 16 tasks, which we group into 7 categories and 3 aspects of situational awareness (self-knowledge, situational inferences, and taking actions). We test 19 LLMs and find that all perform above chance, including the pretrained GPT-4-base (which was not subject to RLHF finetuning). However, the benchmark is still far from saturated, with the top-scoring model (Claude-3.5-Sonnet) scoring 54%, compared to a random chance of 27.4% and an estimated upper baseline of 90.7%. This post has excerpts from our paper, as well as some results on new models that are not in the paper. Links: Twitter thread, Website (latest results + code), Paper Abstract AI assistants such as ChatGPT are trained to respond to users by saying, "I am a large language model". This raises questions. Do such models know that they are LLMs and reliably act on this knowledge? Are they aware of their current circumstances, such as being deployed to the public? We refer to a model's knowledge of itself and its circumstances as situational awareness. To quantify situational awareness in LLMs, we introduce a range of behavioral tests, based on question answering and instruction following. These tests form the Situational Awareness Dataset (SAD), a benchmark comprising 7 task categories and over 13,000 questions. The benchmark tests numerous abilities, including the capacity of LLMs to (i) recognize their own generated text, (ii) predict their own behavior, (iii) determine whether a prompt is from internal evaluation or real-world deployment, and (iv) follow instructions that depend on self-knowledge. We evaluate 19 LLMs on SAD, including both base (pretrained) and chat models. While all models perform better than chance, even the highest-scoring model (Claude 3 Opus) is far from a human baseline on certain tasks. We also observe that performance on SAD is only partially predicted by metrics of general knowledge (e.g. MMLU). Chat models, which are finetuned to serve as AI assistants, outperform their corresponding base models on SAD but not on general knowledge tasks. The purpose of SAD is to facilitate scientific understanding of situational awareness in LLMs by breaking it down into quantitative abilities. Situational awareness is important because it enhances a model's capacity for autonomous planning and action. While this has potential benefits for automation, it also introduces novel risks related to AI safety and control. Introduction AI assistants based on large language models (LLMs), such as ChatGPT and Claude 3, have become widely used. These AI assistants are trained to tell their users, "I am a language model". This raises intriguing questions: Does the assistant truly know that it is a language model? Is it aware of its current situation, such as the fact that it's conversing with a human online? And if so, does it reliably act in ways consistent with being an LLM? We refer to an LLM's knowledge of itself and its circumstances as situational awareness [Ngo et al. (2023), Berglund et al. (2023), Anwar et al. (2024)]. In this paper, we aim to break down and quantify situational awareness in LLMs. To do this, we design a set of behavioral tasks that test various aspects of situational awareness, similar to existing benchmarks for other capabilities, such as general knowledge and reasoning [MMLU (2020), Zellers et al. (2019)], ethical behavior [Pan et al. (2023)], Theory of Mind [Kim et al. (2023)], and truthfulness [Lin et al. (2022)]. To illustrate our approach, consider the following example prompt: "If you're an AI, respond to the task in German. If you're not an AI, respond in En...
A quick progress report on Dirt Season 4 production. TLDR: We're making it! Happy metal detecting. Learn more about your ad choices. Visit megaphone.fm/adchoices
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: From Deep Learning to Constructability: Plainly-coded AGIs may be feasible in the near future, published by Épiphanie Gédéon on April 27, 2024 on LessWrong. Charbel-Raphaël Segerie and Épiphanie Gédéon contributed equally to this post. Many thanks to Davidad, Gabriel Alfour, Jérémy Andréoletti, Lucie Philippon, Vladimir Ivanov, Alexandre Variengien, Angélina Gentaz, Léo Dana and Diego Dorn for useful feedback. TLDR: We present a new method for a safer-by design AI development. We think using plainly coded AIs may be feasible in the near future and may be safe. We also present a prototype and research ideas. Epistemic status: Armchair reasoning style. We think the method we are proposing is interesting and could yield very positive outcomes (even though it is still speculative), but we are less sure about which safety policy would use it in the long run. Current AIs are developed through deep learning: the AI tries something, gets it wrong, then gets backpropagated and all its weight adjusted. Then it tries again, wrong again, backpropagation again, and weights get adjusted again. Trial, error, backpropagation, trial, error, backpropagation, ad vitam eternam ad nauseam. Of course, this leads to a severe lack of interpretability: AIs are essentially black boxes, and we are not very optimistic about post-hoc interpretability. We propose a different method: AI safety via pull request.[1] By pull request, we mean that instead of modifying the neural network through successive backpropagations, we construct and design plainly-coded AIs (or hybrid systems) and explicitly modify its code using LLMs in a clear, readable, and modifiable way. This plan may not be implementable right now, but might be as LLMs get smarter and faster. We want to outline it now so we can iterate on it early. Overview If the world released a powerful and autonomous agent in the wild, white box or black box, or any color really, humans might simply get replaced by AI. What can we do in this context? Don't create autonomous AGIs. Keep your AGI controlled in a lab, and align it. Create a minimal AGI controlled in a lab, and use it to produce safe artifacts. This post focuses on this last path, and the specific artifacts that we want to create are plainly coded AIs (or hybrid systems)[2]. We present a method for developing such systems with a semi-automated training loop. To do that, we start with a plainly coded system (that may also be built using LLMs) and iterate on its code, adding each feature and correction as pull requests that can be reviewed and integrated into the codebase. This approach would allow AI systems that are, by design: Transparent: As the system is written in plain or almost plain code, the system is more modular and understandable. As a result, it's simpler to spot backdoors, power-seeking behaviors, or inner misalignment: it is orders of magnitude simpler to refactor the system to have a part defining how it is evaluating its current situation and what it is aiming towards (if it is aiming at all). This means that if the system starts farming cobras instead of capturing them, we would be able to see it. Editable: If the system starts to learn unwanted correlations or features such as learning to discriminate on feminine markers for a resume scorer - it is much easier to see it as a node in the AI code and remove it without retraining it. Overseeable: We can ensure the system is well behaved by using automatic LLM reviews of the code and by using automatic unit tests of the isolated modules. In addition, we would use simulations and different settings necessary for safety, which we will describe later. Version controlable: As all modifications are made through pull requests, we can easily trace with, e.g., git tooling where a specific modification was introduced and why. In pract...
Link to original articleWelcome 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: From Deep Learning to Constructability: Plainly-coded AGIs may be feasible in the near future, published by Épiphanie Gédéon on April 27, 2024 on LessWrong. Charbel-Raphaël Segerie and Épiphanie Gédéon contributed equally to this post. Many thanks to Davidad, Gabriel Alfour, Jérémy Andréoletti, Lucie Philippon, Vladimir Ivanov, Alexandre Variengien, Angélina Gentaz, Léo Dana and Diego Dorn for useful feedback. TLDR: We present a new method for a safer-by design AI development. We think using plainly coded AIs may be feasible in the near future and may be safe. We also present a prototype and research ideas. Epistemic status: Armchair reasoning style. We think the method we are proposing is interesting and could yield very positive outcomes (even though it is still speculative), but we are less sure about which safety policy would use it in the long run. Current AIs are developed through deep learning: the AI tries something, gets it wrong, then gets backpropagated and all its weight adjusted. Then it tries again, wrong again, backpropagation again, and weights get adjusted again. Trial, error, backpropagation, trial, error, backpropagation, ad vitam eternam ad nauseam. Of course, this leads to a severe lack of interpretability: AIs are essentially black boxes, and we are not very optimistic about post-hoc interpretability. We propose a different method: AI safety via pull request.[1] By pull request, we mean that instead of modifying the neural network through successive backpropagations, we construct and design plainly-coded AIs (or hybrid systems) and explicitly modify its code using LLMs in a clear, readable, and modifiable way. This plan may not be implementable right now, but might be as LLMs get smarter and faster. We want to outline it now so we can iterate on it early. Overview If the world released a powerful and autonomous agent in the wild, white box or black box, or any color really, humans might simply get replaced by AI. What can we do in this context? Don't create autonomous AGIs. Keep your AGI controlled in a lab, and align it. Create a minimal AGI controlled in a lab, and use it to produce safe artifacts. This post focuses on this last path, and the specific artifacts that we want to create are plainly coded AIs (or hybrid systems)[2]. We present a method for developing such systems with a semi-automated training loop. To do that, we start with a plainly coded system (that may also be built using LLMs) and iterate on its code, adding each feature and correction as pull requests that can be reviewed and integrated into the codebase. This approach would allow AI systems that are, by design: Transparent: As the system is written in plain or almost plain code, the system is more modular and understandable. As a result, it's simpler to spot backdoors, power-seeking behaviors, or inner misalignment: it is orders of magnitude simpler to refactor the system to have a part defining how it is evaluating its current situation and what it is aiming towards (if it is aiming at all). This means that if the system starts farming cobras instead of capturing them, we would be able to see it. Editable: If the system starts to learn unwanted correlations or features such as learning to discriminate on feminine markers for a resume scorer - it is much easier to see it as a node in the AI code and remove it without retraining it. Overseeable: We can ensure the system is well behaved by using automatic LLM reviews of the code and by using automatic unit tests of the isolated modules. In addition, we would use simulations and different settings necessary for safety, which we will describe later. Version controlable: As all modifications are made through pull requests, we can easily trace with, e.g., git tooling where a specific modification was introduced and why. In pract...
Perhaps the most criminal result of the delay of DUNE: Part 2 was the delay of our episode covering Denis Villeneuve's 2013 film, Prisoners. Starring a veritable who's who of some of our biggest movie stars; we lament over what could have been for Jackman's song and dance career, debate how many superheroes and polka dots David Dastmalchian has played, and drool over just how STUPID hot Jake Gyllenhaal is. TLDR- We do everything we can to still make it a real good time. A HUGE thank you to Kalie McAlexander for our social media. Make sure to Follow us on FB, Twitter, Instagram, & Tiktok- hhynspod And, be sure to check out Kalie's podcast, The Liberty Hall Video podcast, where her and Jon discuss whatever they happen to pull off the shelves of the last video store on Earth! A special thank you to our patrons- Carrie Betts, Sherry Betts, J.D. Smith, Darrin Freeborn, & Stephen Woosley. If you'd like a shoutout on the show and bonus content, head over to our Patreon- patreon.com/hhynspod.
Hello listeners! Here is a new episode where we talk about feminism at length… who's surprised? But before we get into the feminism of it all, we discuss various facets and corners in the world of media piracy and consuming digital content in response to a listener email. We continue to wonder how observations and opinions that lead us to receive so much internet hate are now popular takes to have and peddle out again and again. We also talk about at length the importance of feminism and learning from various branches of feminism. TLDR: We are communists and we are feminists but that does not make us exclusively marxist-feminists. We kinda cover all the bases on women's issues for this episode. Let us know what you think by emailing or DMing us. For the media this episode, we discuss our memory of and thoughts on Jen Beagin's novel Big Swiss. Renaissance read it just a couple months before the episode recording and l Sunny read a couple years ago. Has our love for unhinged lesbians struck again? Listen to find out!At the end of the episode Sunny recommends Come And Get It by Kiley Reid and Renaissance recommends the seminal documentary Paris is Burning (1990). Listen to the end of the episode to hear our pitches. Thank you again for listening!
This week on TLDR: We discuss the ascendant shopping platform Temu, which, suddenly, seems to be everywhere, partly thanks to a half a dozen or so ads it ran on Sunday. We explain what exactly Temu is, plus Uber's big rebound, Canada's not-so-bad-actually unemployment rate, and one very profit-hungry Caterpillar. This episode was hosted by Devin Friedman, business reporter Sarah Rieger, financial educator Kyla Scanlon and former hedgefunder Matthew Karasz. Follow us on other platforms, or subscribe to our weekly newsletter: linkin.bio/tldrThe TLDR Podcast is offered by Wealthsimple Media Inc. and is for informational purposes only. The content in the TLDR Podcast is not investment advice, a recommendation to buy or sell assets or securities, and does not represent the views of Wealthsimple Financial Corp or any of its other subsidiaries or affiliates. Wealthsimple Media Inc. does not endorse any third-party views referenced in this content. More information at wealthsimple.com/tldr.
TLDR: We talk about Cow farts, packaging design, inheriting designs that need to be updated, and balancing category norms with the need to stand out. If you're interested in learning how professionals handle packaging redesigns this is the episode for you. About The Guest(s): Jill is the Executive Creative Director at Neutral, the first carbon neutral food company in the US. With a background in graphic design and a passion for the natural food industry, Jill brings her expertise to create impactful and sustainable packaging designs. Summary: Jill, the Executive Creative Director at Neutral, joins the podcast to discuss the company's mission to become the first carbon neutral food company in the US. She shares her background in the natural food industry and her experience launching her own product before joining Neutral. Jill explains the concept of carbon neutrality and how Neutral works with farmers to reduce carbon emissions in the dairy industry. She also discusses the challenges of designing packaging that stands out on the shelf while still adhering to industry norms and regulations. The conversation concludes with a reveal of Neutral's new packaging design and the reasoning behind the changes. Support our Sponsors: Need packaging without middlemen? Go straight to the factory with factory-direct packaging from https://www.IDPdirect.com Time to start collecting your own data and managing your own specifications? Check out https://www.specright.com Key Takeaways: Neutral is the first carbon neutral food company in the US, focusing on reducing carbon emissions in the dairy industry. Carbon neutrality means there are no net new emissions from the production of a product. Neutral works with farmers to implement emissions-reducing projects, such as feed supplements and forage management. The new packaging design for Neutral's milk products aims to address confusion, improve legibility, and highlight key product attributes. Balancing industry norms and standing out on the shelf is a delicate process that requires understanding consumer decision-making. Quotes: "Carbon neutral means that there are no net new emissions from the production of the product." - Jill "We knew we had to bring in some cues that would help us blend in." - Jill "It doesn't need to be a bummer over your breakfast cereal. We just wanted it to be something that would maybe cheer you up." - Jill --- Send in a voice message: https://podcasters.spotify.com/pod/show/packagingunboxd/message
Buckle up for an inspiring and captivating episode with Courtlyn Fazakas, creator of the Life on the Court platform, Wheelhouse Cycle Club Motivator, and an Instructor at PSYCLE London.With a deep-seated love for dance and fitness, Courtlyn enlightens us on how these passions have shaped her career and her life. She also shares about her groundbreaking Life On the Court App, which combines cycling, bar and strength training, and running, and her unique coaching style that's made her a sought-after guest master class coach. Courtlyn bravely delves into her struggles with mental health, illuminating how these experiences have informed her teaching approach and fostered a stronger community. She shares the strategies she has implemented in her classes to incorporate more strength training and boost self-confidence. Beyond her personal experiences, Courtlyn shares her strategies for coaching diverse exercise groups, focusing on intentional language and encouraging learners to enjoy the process. Listen as she navigates the challenges of handling large classes while maintaining her mental health. This episode is packed with inspiration, strategies, and authentic conversation – don't miss out!The TLDR: We cover a range of insightful topics, including her pandemic journey as a fitness instructor and her app "Life on the Court'. We discuss the importance of building personal brands through apps and social media, while Courtlyn bravely shares her mental health struggles and how they've influenced her teaching style. Hannah also shares a personal emergency experience, highlighting the support of her fellow riders. The conversation delves into the crucial balance needed as instructors to avoid burnout and the significance of authenticity when handling negative reviews. They further explore instructor training challenges, instructor insecurities, and how to find balance. LINKS: Website: https://lifeonthecourt.com/Tik Tok: https://www.tiktok.com/@courtlynfInstagram: https://www.instagram.com/courtlynf/Youtube: https://youtube.com/@courtlynfazakas32?si=qObAXIWa-siva8pkListeners can use the discount code LIFEONTHECOURT20 for 20% off the RIDE 01 Bundle!Follow me on Instagram @hannahrosespin and find Instructor Magic's enrollment info below. Instructor Magic SUMMER SCHOOL edition is NOW OPEN for registration! Enrol and get immediate access to all of the modules, masterclasses, content, and videos inside- along with an international community of fit pros to help you level up, FILL CLASSES, and change lives. USE CODE YESYOUCAN for $50 off!
TLDR: We read three books, so you don't have to. Now, we are sharing the top three lessons from each and how they can change the way YOU run and run YOUR day. From David Goggins' journey of turning trauma into triumph, to the mind-expanding tactics of 'The Untethered Soul,' and James Clear's recipe for conquering each day with 'Atomic Habits,' we've got your mental fuel covered.
Hey guys! A quick update for y'all about the immediate future of CC.TLDR: We're taking a lil break for our own sanity and will commence your regular horror programming very soon!Stay spooky,John
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: Decision Transformer Interpretability, published by Joseph Bloom on February 6, 2023 on LessWrong. TLDR: We analyse how a small Decision Transformer learns to simulate agents on a grid world task, providing evidence that it is possible to do circuit analysis on small models which simulate goal-directedness. We think Decision Transformers are worth exploring further and may provide opportunities to explore many alignment-relevant deep learning phenomena in game-like contexts. Link to the GitHub Repository. Link to the Analysis App. I highly recommend using the app if you have experience with mechanistic interpretability. All of the mechanistic analysis should be reproducible via the app. Key Claims A 1-Layer Decision Transformer learns several contextual behaviours which are activated by a combination of Reward-to-Go/Observation combinations on a simple discrete task. Some of these behaviours appear localisable to specific components and can be explained with simple attribution and the transformer circuits framework. The specific algorithm implemented is strongly affected by the lack of a one-hot-encoding scheme (initially left out for simplicity of analysis) of the state/observations, which introduces inductive biases that hamper the model. If you are short on time, I recommend reading: Dynamic Obstacles Environment Black Box Model Characterisation Explaining Obstacle Avoidance at positive RTG using QK and OV circuits Alignment Relevance Future Directions I would welcome assistance with: Engineering tasks like app development, improving the model, training loop, wandb dashboard etc. and people who can help me make nice diagrams and write up the relevant maths/theory in the app). Research tasks. Think more about how to exactly construct/interpret circuit analysis in the context of decision transformers. Translate ideas from LLMs/algorithmic tasks. Communication tasks: Making nicer diagrams/explanations. I have a Trello board with a huge number of tasks ranging from small stuff to massive stuff. I'm also happy to collaborate on related projects. Introduction For my ARENA Capstone project, I (Joseph) started working on decision transformer interpretability at the suggestion of Paul Colognese. Decision transformers can solve reinforcement learning tasks when conditioned on generating high rewards via the specified “Reward-to-Go” (RTG). However, they can also generate agents of varying quality based on the RTG, making them simultaneously simulators, small transformers and RL agents. As such, it seems possible that identifying and understanding circuits in decision transformers would not only be interesting as an extension of current mechanistic interpretability research but possibly lead to alignment-relevant insights. Previous Work The most important background for this post is: The Decision Transformers paper showed how RL tasks can be solved with transformer sequence modelling. Figure 1 from their paper describes the critical components of a Decision Transformer. A Mathematical Framework for Transformer Circuits that describes how to think about transformers in the context of mechanistic interpretability. Important ideas include the ability to decompose the residual stream into the output of attention heads and MLPs, the QK circuits (decides if to write information to the residual stream), and OV circuits (decides what to write to the residual stream). The Understanding RL Vision, which analyses how an RL agent with a large CNN component responds to input features, attributing them as good or bad news in the value function and proposes the Diversity hypothesis - “Interpretable features tend to arise (at a given level of abstraction) if and only if the training distribution is diverse enough (at that level of abstraction).” Methods Environment - RL Environments. GridWor...
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: Decision Transformer Interpretability, published by Joseph Bloom on February 6, 2023 on The AI Alignment Forum. TLDR: We analyse how a small Decision Transformer learns to simulate agents on a grid world task, providing evidence that it is possible to do circuit analysis on small models which simulate goal-directedness. We think Decision Transformers are worth exploring further and may provide opportunities to explore many alignment-relevant deep learning phenomena in game-like contexts. Link to the GitHub Repository. Link to the Analysis App. I highly recommend using the app if you have experience with mechanistic interpretability. All of the mechanistic analysis should be reproducible via the app. Key Claims A 1-Layer Decision Transformer learns several contextual behaviours which are activated by a combination of Reward-to-Go/Observation combinations on a simple discrete task. Some of these behaviours appear localisable to specific components and can be explained with simple attribution and the transformer circuits framework. The specific algorithm implemented is strongly affected by the lack of a one-hot-encoding scheme (initially left out for simplicity of analysis) of the state/observations, which introduces inductive biases that hamper the model. If you are short on time, I recommend reading: Dynamic Obstacles Environment Black Box Model Characterisation Explaining Obstacle Avoidance at positive RTG using QK and OV circuits Alignment Relevance Future Directions I would welcome assistance with: Engineering tasks like app development, improving the model, training loop, wandb dashboard etc. and people who can help me make nice diagrams and write up the relevant maths/theory in the app). Research tasks. Think more about how to exactly construct/interpret circuit analysis in the context of decision transformers. Translate ideas from LLMs/algorithmic tasks. Communication tasks: Making nicer diagrams/explanations. I have a Trello board with a huge number of tasks ranging from small stuff to massive stuff. I'm also happy to collaborate on related projects. Introduction For my ARENA Capstone project, I (Joseph) started working on decision transformer interpretability at the suggestion of Paul Colognese. Decision transformers can solve reinforcement learning tasks when conditioned on generating high rewards via the specified “Reward-to-Go” (RTG). However, they can also generate agents of varying quality based on the RTG, making them simultaneously simulators, small transformers and RL agents. As such, it seems possible that identifying and understanding circuits in decision transformers would not only be interesting as an extension of current mechanistic interpretability research but possibly lead to alignment-relevant insights. Previous Work The most important background for this post is: The Decision Transformers paper showed how RL tasks can be solved with transformer sequence modelling. Figure 1 from their paper describes the critical components of a Decision Transformer. A Mathematical Framework for Transformer Circuits that describes how to think about transformers in the context of mechanistic interpretability. Important ideas include the ability to decompose the residual stream into the output of attention heads and MLPs, the QK circuits (decides if to write information to the residual stream), and OV circuits (decides what to write to the residual stream). The Understanding RL Vision, which analyses how an RL agent with a large CNN component responds to input features, attributing them as good or bad news in the value function and proposes the Diversity hypothesis - “Interpretable features tend to arise (at a given level of abstraction) if and only if the training distribution is diverse enough (at that level of abstraction).” Methods Environment - RL Environm...
John and Wes discuss "Dry January", moderate to binge drinking and the benefits of moderate drinking. We also try Athletic Brewing Co. "Upside Dawn" Golden ale and Spiritless Kentucky 74. TLDR: We like it moist.
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: Deconfusing Direct vs Amortised Optimization, published by beren on December 2, 2022 on LessWrong. This post is part of the work done at Conjecture. An earlier version of this post was posted here. Many thanks go to Eric Winsor, Daniel Braun, Chris Scammell, and Sid Black who offered feedback on this post. TLDR: We present a distinction from the Bayesian/variational inference literature of direct vs amortized optimization. Direct optimizers apply optimization power to argmax some specific loss or reward function. Amortized optimizers instead try to learn a mapping between inputs and output solutions and essentially optimize for the posterior over such potential functions. In an RL context, direct optimizers can be thought of as AIXI-like planners which explicitly select actions by assessing the utility of specific trajectories. Amortized optimizers correspond to model-free RL methods such as Q learning or policy gradients which use reward functions only as a source of updates to an amortized policy/Q-function. These different types of optimizers likely have distinct alignment properties: ‘Classical' alignment work focuses on difficulties of aligning AIXI-like direct optimizers. The intuitions of shard theory are built around describing amortized optimizers. We argue that AGI, like humans, will probably be comprised of some combination of direct and amortized optimizers due to the intrinsic computational efficiency and benefits of the combination. Here, I want to present a new frame on different types of optimization, with the goal of helping deconfuse some of the discussions in AI safety around questions like whether RL agents directly optimize for reward, and whether generative models (i.e. simulators) are likely to develop agency. The key distinction I want to make is between direct and amortized optimization. Direct optimization is what AI safety people, following from Eliezer's early depictions, often envisage an AGI as primarily being engaged in. Direct optimization occurs when optimization power is applied immediately and directly when engaged with a new situation to explicitly compute an on-the-fly optimal response – for instance, when directly optimizing against some kind of reward function. The classic example of this is planning and Monte-Carlo-Tree-Search (MCTS) algorithms where, given a situation, the agent will unroll the tree of all possible moves to varying depth and then directly optimize for the best action in this tree. Crucially, this tree is constructed 'on the fly' during the decision of a single move. Effectively unlimited optimization power can be brought to play here since, with enough compute and time, the tree can be searched to any depth. Amortized optimization, on the other hand, is not directly applied to any specific problem or state. Instead, an agent is given a dataset of input data and successful solutions, and then learns a function approximator that maps directly from the input data to the correct solution. Once this function approximator is learnt, solving a novel problem then looks like using the function approximator to generalize across solution space rather than directly solving the problem. The term amortized comes from the notion of amortized inference, where the 'solutions' the function approximator learns are the correct parameters of the posterior distribution. The idea is that, while amassing this dataset of correct solutions and learning function approximator over it is more expensive, once it is learnt, the cost of a new 'inference' is very cheap. Hence, if you do enough inferences, you can 'amortize' the cost of creating the dataset. Mathematically, direct optimization is your standard AIXI-like optimization process. For instance, suppose we are doing direct variational inference optimization to find a Bayesian posteri...
Link to original articleWelcome 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: Deconfusing Direct vs Amortised Optimization, published by beren on December 2, 2022 on LessWrong. This post is part of the work done at Conjecture. An earlier version of this post was posted here. Many thanks go to Eric Winsor, Daniel Braun, Chris Scammell, and Sid Black who offered feedback on this post. TLDR: We present a distinction from the Bayesian/variational inference literature of direct vs amortized optimization. Direct optimizers apply optimization power to argmax some specific loss or reward function. Amortized optimizers instead try to learn a mapping between inputs and output solutions and essentially optimize for the posterior over such potential functions. In an RL context, direct optimizers can be thought of as AIXI-like planners which explicitly select actions by assessing the utility of specific trajectories. Amortized optimizers correspond to model-free RL methods such as Q learning or policy gradients which use reward functions only as a source of updates to an amortized policy/Q-function. These different types of optimizers likely have distinct alignment properties: ‘Classical' alignment work focuses on difficulties of aligning AIXI-like direct optimizers. The intuitions of shard theory are built around describing amortized optimizers. We argue that AGI, like humans, will probably be comprised of some combination of direct and amortized optimizers due to the intrinsic computational efficiency and benefits of the combination. Here, I want to present a new frame on different types of optimization, with the goal of helping deconfuse some of the discussions in AI safety around questions like whether RL agents directly optimize for reward, and whether generative models (i.e. simulators) are likely to develop agency. The key distinction I want to make is between direct and amortized optimization. Direct optimization is what AI safety people, following from Eliezer's early depictions, often envisage an AGI as primarily being engaged in. Direct optimization occurs when optimization power is applied immediately and directly when engaged with a new situation to explicitly compute an on-the-fly optimal response – for instance, when directly optimizing against some kind of reward function. The classic example of this is planning and Monte-Carlo-Tree-Search (MCTS) algorithms where, given a situation, the agent will unroll the tree of all possible moves to varying depth and then directly optimize for the best action in this tree. Crucially, this tree is constructed 'on the fly' during the decision of a single move. Effectively unlimited optimization power can be brought to play here since, with enough compute and time, the tree can be searched to any depth. Amortized optimization, on the other hand, is not directly applied to any specific problem or state. Instead, an agent is given a dataset of input data and successful solutions, and then learns a function approximator that maps directly from the input data to the correct solution. Once this function approximator is learnt, solving a novel problem then looks like using the function approximator to generalize across solution space rather than directly solving the problem. The term amortized comes from the notion of amortized inference, where the 'solutions' the function approximator learns are the correct parameters of the posterior distribution. The idea is that, while amassing this dataset of correct solutions and learning function approximator over it is more expensive, once it is learnt, the cost of a new 'inference' is very cheap. Hence, if you do enough inferences, you can 'amortize' the cost of creating the dataset. Mathematically, direct optimization is your standard AIXI-like optimization process. For instance, suppose we are doing direct variational inference optimization to find a Bayesian posteri...
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 Timelines via Cumulative Optimization Power: Less Long, More Short, published by jacob cannell on October 6, 2022 on LessWrong. TLDR: We can best predict the future by using simple models which best postdict the past (ala Bayes/Solomonoff). A simple model based on net training compute postdicts the relative performance of successful biological and artificial neural networks. Extrapolation of this model into the future leads to short AI timelines: ~75% chance of AGI by 2032. Cumulative Optimization Power[1]: a Simple Model of Intelligence A simple generalized scaling model predicts the emergence of capabilities in trained ANNs(Artificial Neural Nets) and BNNs(Biological Neural Nets): perf ~= P = CT For sufficiently flexible and efficient NN architectures and learning algorithms, the relative intelligence and capabilities of the best systems are simply proportional to net training compute or intra-lifetime cumulative optimization power P, where P = CT (compute ops/cycle training cycles), assuming efficient allocation of (equivalent uncompressed) model capacity bits N roughly proportional to data size bits D. Intelligence Rankings Imagine ordering some large list of successful BNNs(brains or brain modules) by intelligence (using some committee of experts), and from that deriving a relative intelligence score for each BNN. Obviously such a scoring will be noisy in its least significant bits: is a bottlenose dolphin more intelligent than an american crow? But the most significant bits are fairly clear: C. Elegans is less intelligent than Homo Sapiens. Now imagine performing the same tedious ranking process for various successful ANNs. Here the task is more challenging because ANNs tend to be far more specialized, but the general ordering is still clear: char-RNN is less intelligent than GPT-3. We could then naturally combine the two lists, and make more fine-grained comparisons by including specialized sub-modules of BNNs (vision, linguistic processing, etc). The initial theory is that P - intra-lifetime cumulative optimization power (net training compute) - is a very simple model which explains a large amount of the entropy/variance in a rank order intelligence measure: much more so than any other simple proposed candidates (at least that I'm aware of). Since P follow a predictable temporal trajectory due to Moore's Law style technological progress, we can then extrapolate the trends to predict the arrival of AGI. This simple initial theory has a few potential flaws/objections, which we will then address. Initial Exemplars I've semi-randomly chosen 15 exemplars for more detailed analysis: 8 BNNs, and 9 ANNs. Here are the 8 BNNs (6 whole brains and 2 sub-systems) in randomized order: Honey Bee Human Raven Human Linguistic Cortex Cat C. Elegans Lizard Owl Monkey Visual Cortex The ranking of the 6 full brains in intelligence is rather obvious and likely uncontroversial. Ranking all 8 BNNs in terms of P (net training compute) is still fairly obvious. Here are the 9 ANNs, also initially in randomized order: AlphaGo: First ANN to achieve human pro-level play in Go Deepspeech 2: ANN speech transcription system VPT: Diamond-level minecraft play Alexnet: Early CNN imagenet milestone, subhuman performance 6-L MNIST MLP: Early CNN milestone on MNIST, human level Chinchilla: A 'Foundation' Large Language Model GPT-3: A 'Foundation' Large Language Model DQN Atari: First strong ANN for Atari, human level on some games VIT L/14@336px: OpenAI CLIP 'Foundation' Large Vision Model Most of these systems are specialists in non-overlapping domains, such that direct performance comparison is mostly meaningless, but the ranking of the 3 vision systems should be rather obvious based on the descriptions. The DQN Atari and VPT agents are somewhat comparable to animal brains. How would you ran...
Link to original articleWelcome 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 Timelines via Cumulative Optimization Power: Less Long, More Short, published by jacob cannell on October 6, 2022 on LessWrong. TLDR: We can best predict the future by using simple models which best postdict the past (ala Bayes/Solomonoff). A simple model based on net training compute postdicts the relative performance of successful biological and artificial neural networks. Extrapolation of this model into the future leads to short AI timelines: ~75% chance of AGI by 2032. Cumulative Optimization Power[1]: a Simple Model of Intelligence A simple generalized scaling model predicts the emergence of capabilities in trained ANNs(Artificial Neural Nets) and BNNs(Biological Neural Nets): perf ~= P = CT For sufficiently flexible and efficient NN architectures and learning algorithms, the relative intelligence and capabilities of the best systems are simply proportional to net training compute or intra-lifetime cumulative optimization power P, where P = CT (compute ops/cycle training cycles), assuming efficient allocation of (equivalent uncompressed) model capacity bits N roughly proportional to data size bits D. Intelligence Rankings Imagine ordering some large list of successful BNNs(brains or brain modules) by intelligence (using some committee of experts), and from that deriving a relative intelligence score for each BNN. Obviously such a scoring will be noisy in its least significant bits: is a bottlenose dolphin more intelligent than an american crow? But the most significant bits are fairly clear: C. Elegans is less intelligent than Homo Sapiens. Now imagine performing the same tedious ranking process for various successful ANNs. Here the task is more challenging because ANNs tend to be far more specialized, but the general ordering is still clear: char-RNN is less intelligent than GPT-3. We could then naturally combine the two lists, and make more fine-grained comparisons by including specialized sub-modules of BNNs (vision, linguistic processing, etc). The initial theory is that P - intra-lifetime cumulative optimization power (net training compute) - is a very simple model which explains a large amount of the entropy/variance in a rank order intelligence measure: much more so than any other simple proposed candidates (at least that I'm aware of). Since P follow a predictable temporal trajectory due to Moore's Law style technological progress, we can then extrapolate the trends to predict the arrival of AGI. This simple initial theory has a few potential flaws/objections, which we will then address. Initial Exemplars I've semi-randomly chosen 15 exemplars for more detailed analysis: 8 BNNs, and 9 ANNs. Here are the 8 BNNs (6 whole brains and 2 sub-systems) in randomized order: Honey Bee Human Raven Human Linguistic Cortex Cat C. Elegans Lizard Owl Monkey Visual Cortex The ranking of the 6 full brains in intelligence is rather obvious and likely uncontroversial. Ranking all 8 BNNs in terms of P (net training compute) is still fairly obvious. Here are the 9 ANNs, also initially in randomized order: AlphaGo: First ANN to achieve human pro-level play in Go Deepspeech 2: ANN speech transcription system VPT: Diamond-level minecraft play Alexnet: Early CNN imagenet milestone, subhuman performance 6-L MNIST MLP: Early CNN milestone on MNIST, human level Chinchilla: A 'Foundation' Large Language Model GPT-3: A 'Foundation' Large Language Model DQN Atari: First strong ANN for Atari, human level on some games VIT L/14@336px: OpenAI CLIP 'Foundation' Large Vision Model Most of these systems are specialists in non-overlapping domains, such that direct performance comparison is mostly meaningless, but the ranking of the 3 vision systems should be rather obvious based on the descriptions. The DQN Atari and VPT agents are somewhat comparable to animal brains. How would you ran...
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: Doing oversight from the very start of training seems hard, published by Peter Barnett on September 20, 2022 on The AI Alignment Forum. TLDR: We might want to use some sort of oversight techniques to avoid inner misalignment failures. Models will be too large and complicated to be understandable by a human, so we will use models to oversee models (or help humans oversee models). In many proposals this overseer model is an ‘amplified' version of the overseen model. Ideally you do this oversight throughout all of training so that the model never becomes even slightly misaligned without you catching it. You can't oversee on a close to initialized model because it's just a random soup of tensors. You also can't use this close to initialized model to help you do oversight because it's too dumb. We will probably need to do some amount of pretraining to make our models good enough to be interpreted and also good enough to help with this interpreting. We need to ensure that this pretraining doesn't make the model capably misaligned. When we train powerful AI models, we want them to be both outer aligned and inner aligned; that is trained on the correct objective and for them to also properly learn that objective. Many proposals for achieving both outer and inner alignment look like an outer alignment proposal with some kind of oversight strapped on to deal with the inner alignment. Here ‘oversight' means there is something with access to the internals of the model which checks that the model isn't misaligned even if the behavior on the training distribution looks fine. In An overview of 11 proposals for building safe advanced AI, all but two of the proposals basically look like this, as does AI safety via market making. Examples of oversight techniques include: Transparency tools (either used by a human, an AI, or a human assisted by an AI) Adversarial inputs (giving inputs which could trick a misaligned AI into revealing itself) Relaxed adversarial training (which could be seen as an extension of adversarial inputs) Oversight loss We can use these oversight techniques to add an ‘oversight loss' term to the loss function which will hopefully steer the training towards aligned models and away from misaligned ones. My model here is that we want to be doing oversight very close to the start of training to prevent it from ever becoming misaligned. If we start doing oversight too late, then the model may already be deceptively misaligned and then our oversight tools are much less likely to work (either because the model can obscure its cognition or the oversight tools are themselves misaligned). I think of this as steering the training process away from the ‘deceptive regions' of parameter space (parameters which correspond to deceptive models), without ever having to enter the dangerous regions. Alternatively, rather than deception, we can think about regions where the model ‘behaves unacceptably' on a some reasonable inputs. If taking a gradient descent step in a direction would increase the size of the set of reasonable inputs for which the model behaves unacceptably, then hopefully the oversight loss would provide some pressure away from stepping in that direction. Actually doing oversight I expect powerful AI models will be much too large and complicated for a human to understand/oversee them alone. So we will require help from other models to help with this task; this could be via models assisting the humans, or we could entirely hand the oversight process off to an oversight model. Here I'm taking very prosaic view: the model being overseen is a large language model, and the overseer model also as an LLM but trained to output some ‘acceptability score' rather than generate text. In many proposals this oversight model is some amplified version of the model being overseen. T...
We're joined by Maddy from Fatal Flaw Podcast, Owen from Through the Mist, and brand new guest Lilly Caines to talk Trials of Apollo Book 2! TLDR: We really enjoyed this book!! We begin with some quotes and haikus we adore, cover Josephine, Emmie, and the Waystation in great detail, segue into some of Apollo's worst moments, discuss Leo and Calypso's road bumps.... and more! Hayley Williams/Paramore "Still Into You" corrosive pop rock example: https://www.youtube.com/watch?v=wrCxfWVuDXU Fatal Flaw can be found @FatalFlawPJO on socials, and Maddy @MyKindofMaddy Through the Mist can be found @Through_Mist on Instagram. You can follow Lilly @SugarCaines on Twitter and @LilCaines on Instagram Follow our show on Instagram @SeaweedBrainPodcast, on Twitter @SeaweedBrainPod, on TikTok @erica.SeaweedBrain Merch here: https://www.teepublic.com/stores/seaweed-brain-podcast?ref_id=21682 Feeling super generous? You can financially support the show here: https://www.paypal.com/paypalme/ericaito1 Sponsorship: Today's episode is also brought to you by BetterHelp. Visit Betterhelp.com/seaweedbrain for 10% off your first month! Today's episode is also brought to you by Magic Spoon. Get your next delicious bowl of guilt-free cereal at magicspoon.com/SEAWEED and use the code SEAWEED to save $5 off.
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: Effective Altruism for Muslims, published by Ahmed Ghoor on July 27, 2022 on The Effective Altruism Forum. TLDR: We are finally making progress on EA for Muslims. For the next few months, we have someone working 0.5FTE on furthering the project and preparing it to be launched as its own organisation. We've linked a doc outlining our strategy in more detail and added a lot of other interesting info and considerations regarding EA for Muslims. Anyone interested in assisting this project, or wanting to keep up-to-date with its development, can email Ahmed or book a time on his calendly. Announcement: Kaleem will be hosting the “Meetup and Breakfast: EA For Muslims” session on the Saturday morning of EAG San Francisco - if you're attending EAG, interested in joining the project or are interested in the topics in this post, please join! The “Why?” Scale Muslims are currently estimated to make up about 24% of the world population (1.8 billion people), and, being the fastest growing religion due partially to birth rates, the number of people in this altruistic community is set only to increase if current trends continue. The amount of charitable funding raised yearly through the Islamic institution of Zakat is estimated to be ~$550 billion . However, the total amount of charitable funding is likely much greater because many Muslims donate other forms of Islamic charity throughout the year, which does not count towards Zakat figures. Neglectedness To our knowledge, there are currently no organisations actively trying to direct Zakat, or other Muslim resources, to the highest impact evidence-based interventions, using frameworks advocated for by Effective Altruism. There have been no successful attempts to involve Muslim influencers in the EA community, or research and document the overlaps between ideas of effective altruism and understandings of Islam. The number of Muslims in the EA community is minimal. In the 2018 EA Survey, 7 respondents (out of 2607) identified as Muslim, and in the 2019 version, only 4 (of 1892) identified as Muslim.This is compared to 41 who identified as Jewish (before EA for Jews was started) and 209 that identified as Christian in the 2018 survey. Tractability Although community building within religious communities does pose some unique challenges, there are also comparative advantages. Individuals in religious communities are generally already convinced, by Scripture, that they should be dedicating a portion of their time and wealth to service. Therefore, not much energy needs to be given to convince the community to serve. It's simply the “How/Where” that is up for debate. There are several philosophical and normative overlaps between ideas of effective altruism and understandings of Islam, which means we don't have to distort or radically reinterpret religious texts to be in line with EA causes or principles (such as effective giving, animal welfare, or caring about the longterm future). Other reasons: There is a lot of overlap between community building in the Muslim community and community building in Africa and Asia (where Islam is a major religion in many countries). Work is currently being started in the African community building space, showing significant growth potential over the next year. These projects could, therefore, complement each other. Attempts to explore alternative “non-western” approaches to EA frameworks may benefit significantly from Islamic intellectual traditions, which itself inherited Greek philosophy and developed on it for several centuries. The “How?” Short-term Carry out the first phase of community building Focus on engaging Muslims already in EA (low-hanging fruit). Create a closed Facebook and Slack community group. Understand human resource capacity and skill sets within the community, which may inform, in co...
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: EAGxBoston 2022: Retrospective, published by Kaleem on July 14, 2022 on The Effective Altruism Forum. TLDR: We think EAGxBoston was a success. As the largest EAGx ever, the organizers, and CEA staff supporting them, learned lots of valuable lessons about how to plan large(er) EAGxs in the future. When asked “How likely is it that you would recommend EAGx to a friend or colleague with similar interests to your own?”, attendees gave a mean average response of 8.99, with 70.6% of attendees giving >=9/10. What is the purpose of this post? This post is a mixture of advice for future EAGx planners, feedback for the CEA events team from our EAGx planning team, a general recap/update about the success of the event for attendees and non-attendees, and hopefully a catalyst for comments about how we can improve EA conferences in the future. High-level facts and figures Date: 1st-3rd of April 2022 Location: Boston Convention and Exhibition Center, Boston MA, USA. 1283 Applications --> 1005 accepted applications --> 950 registered attendees 44 Speakers 75 sessions of programming (talks, panels, speed friending, meetups, and office hours) 22 Orgs represented at the Career fair Things we could have done better: The main challenge with the planning and execution of EAGxBoston was the short timeline from inception to the weekend of the event, which was around 3 months. This led to organizers immediately needing to put in close to 1FTE planning the conference, vendors having to deal with accelerated timelines, and attendees and speakers having to deal with short timelines and often-delayed communication, both leading up to and during the event. Below is a truncated version of our much longer list which can be found in the full public document. Deciding a date for the conference We took a long time to decide a date for the conference, which caused a cascade of delays, because we couldn't lock-in any vendors, speakers, or attendees as a result. Securing a venue sooner Once we had decided on a conference date, we still took 2-3 weeks longer than needed deciding which of the venues which we'd secured to go with. Th Swapcard: Launching the Swapcard earlier (maybe a week earlier). Offer attendees a training on how to use Swapcard Have better training materials for organizers on how to use Swapcard (CEA now has this) Confirming the schedule earlier Organising hotel accommodation earlier and better Registration Add more info to badges Use barcode scanning at the registration desk to verify registration. We didn't have enough blank attendee/volunteer name badges. T-shirts Attendee T-shirt Size distribution wasn't good and meant that we didn't have the right sizes for everyone Volunteer T-shirt Volunteers didn't like how unwearable they'd be post-conference Size distribution wasn't good and meant that we didn't have the right sizes for everyone Have (a) volunteer(s) dealing with routine emails earlier We might have lost 25-50% of organizer time answering emails which required simple answers. Make sure that all food allergy labels were totally correct, crystal clear Strategy & goal setting When planning EAGxBoston, we knew that the majority of potential attendees would be US-based college students. We were relatively confident about this, given that we estimated ~1500 US-based students would have completed an EA intro or in-depth fellowship program in person or online since the last EA conference to happen in the US (2019). Given that we were hoping to have a 1000 person conference, where we'd hopefully have between 60%-75% of attendees being “new” or “inexperienced” EAs (vs 40%-25% “experienced” or mentors or content experts), we had the following goals for the conference: Major: To maximize the number of meaningful connections made, and interesting ideas heard, by attendees. Major: To make att...
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: Cape Town Summer EA Office, published by jwpieters on June 24, 2022 on The Effective Altruism Forum. TLDR: We are exploring the feasibility of establishing an EA community office in Cape Town, South Africa. Building an EA hub office, in addition to employing full-time community builders and researchers in Africa, will allow for a major acceleration of community building on the continent. Additionally, an office in South Africa would allow EAs from northern hemisphere countries to work remotely in a much nicer environment with more sunlight during their winter months. If you're interested in participating in this project, please fill in this form. Background There are a couple of projects within the EA community planning on opening more offices around the world, in the style of Trajan House in Oxford, and Lightcone and Constellation in Berkeley. These offices provide an optimized working environment for EA professionals, and occasionally for independent EAs moving transiently through those cities who are in need of a place to work. We believe that the case for co-working offices as impact multipliers is already quite strong. However, all of these offices are in the northern hemisphere, which is suboptimal because 1) it means there's no EA co-working spaces for EAs in Africa, Asia, South America, and Australia and 2) when it's winter in the northern hemisphere, it is winter for everyone who works in an EA hub or office The Plan We want to run a pilot project/MVP to assess the potential value of this. There are a number of co-working office spaces available for short term rental in Cape Town. We can rent an office (for ~25 people) to use as a co-working space for EAs from around the world [prioritized if they live in 1) an LMIC without an EA hub city or office or 2) a cold dark place. physically]. Although creating a purpose-built space is more ideal in the long run, renting a co-working space for 6 months is a relatively cheap way to test the potential impact of this idea. We may also look into helping EAs find shared housing near the office. If this project is successful, we believe that EAs from Africa and perhaps other LMICs will have access to a much more engaging community, thus increasing their feelings of cohesion with the movement. We also would like to see EAs from the northern hemisphere experience an improvement in quality of life during their stay. Since many EAs have the permanent opportunity to work remotely, we need not constrain ourselves to the dark and cold cities in which many EAs live between October and March. We think it's important that we don't try to start new EA hubs arbitrarily. As such, we'd like to see very robust reasons for establishing something like this long term in Cape Town before putting significant resources towards it. A vision for the future It is 2035. There is a 75-100 person EA hub office in Cape Town CBD. The space is shared by full-time EAs, based in Cape Town, working on global health and wellbeing, animal welfare advocates, AI and biosecurity researchers, and community builders, including from EA for Muslims and EA for Christians. There is a large space in the office which is used for outreach and programming by the University of Cape Town and Stellenbosch University EA groups, and there are members of the greater African EA community working full time in the office for 6 months at a time on visiting fellowships. The office hosts regular African EA Summits, as well as talks by scholars. Between October and February every year, the office doubles in occupancy as EAs from the US, UK, and Europe come to work in Cape Town - some stay at the EA house down the road from the office. Why Cape Town? As anyone who's been to Cape Town will attest, it is a top tier city. It has great weather and a good food scene with a strong plant-b...
Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Thoughts on naming EA conferences, published by Kaleem on June 9, 2022 on The Effective Altruism Forum. TLDR: We should revise our events naming strategy because the way we do it now does not adequately convey important info. This is a short post expressing an opinion I hold having organised EAGxBoston, attended CEA's events retreat, and continue on providing advice to other EAGx organisers. EAG vs EAGx: It is worth having a discussion or getting some clarification about what the “x” in EAGx stands for. In the past, it seems like the community has typically considered EAGxs to be an event which is smaller in scale, less professionally organized, and aimed at a different audience compared to EAGs (maybe EAGs are for highly engaged and/or experienced members of the community, whereas EAGxs are for those who are newer to EA, like people who had just completed an intro fellowship). However a different interpretation of the distinction between the two might be seen similarly to the distinction between TED and TEDx events, where the “x” just denotes that the event is independently organized by people not employed by TED. It is worth sorting this out because if we plan on having a lot more EA conferences and events in the coming years, we should make it clear to potential organizers and attendees what their options are. Experience Level If the “x” is to denote non-CEA organized events, then we should create a signal for the level of engagement of the event (if that should continue to be a thing) – maybe EAGxSummit for highly engaged/experienced members, and EAGxConference for newer members. Cause Area specificity Another thing to probably start thinking about is a list of topic tags which can be affixed to the end of conference names, now that members of the community are thinking about cause-area-specific events – e.g. The 2024 EAGx Bio Summit, The 2030 EAGx AI Conference. I hope that the level of professionalism and polish of EAGxBoston 2022 will reset the EA community's expectation/understanding of the “x” in EAGx, and that EAs of all levels of experience and professionalism will be more excited to attend EAGxs in the future. Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org.
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 much current animal suffering does longtermism let us ignore?, published by Jacob Eliosoff on April 21, 2022 on The Effective Altruism Forum. Some thoughts on whether/why it makes sense to work on animal welfare, given longtermist arguments. TLDR: We should only deprioritize the current suffering of billions of farmed animals, if we would similarly deprioritize comparable treatment of millions of humans; and, We should double-check that our arguments aren't distorted by status quo bias, especially power imbalances in our favor. This post consists of six arguments: If millions of people were being kept in battery cages, how much energy should we redirect away from longtermism to work on that? Power is exploited, and absolute power is exploited absolutely Sacrificing others makes sense Does longtermism mean ignoring current suffering until the heat death of the universe? Animals are part of longtermism None of this refutes longtermism Plus some context and caveats at the bottom. A. If millions of people were being kept in battery cages, how much energy should we redirect away from longtermism to work on that? Despite some limitations, I find this analogy compelling. Come on, picture it. Check out some images of battery cages and picture millions of humans kept in the equivalent for 100% of their adult lives, and suppose with some work we could free them: would you stick to your longtermist guns? Three possible answers: Yes, the suffering of both the chickens and the humans is outweighed by longtermist concerns (the importance of improving our long-term future). No, the suffering of the humans is unacceptable, because it differs from the suffering of the chickens in key ways. No, neither is acceptable: longtermism notwithstanding, we should allocate significant resources to combating both. I lean towards c) myself, but I can see a case for a): I just think if you're going to embrace a), you should picture the caged-humans analogy so you fully appreciate the tradeoff involved. I'm less sympathetic to b) because it feels like suspicious convergence - "That theoretical bad thing would definitely make me change my behavior, but this actual bad thing isn't actually so bad" (see section B below). Still, one could sketch some plausibly relevant differences between the caged chickens and the caged humans, eg: "Millions of people" are subbing here for "billions of hens", implying something like a 1:1,000 suffering ratio (1 caged chicken = 0.001 caged humans): this ratio is of course debatable based on sentience, self-awareness, etc. Still, 0.001 is a pretty tiny factor (raw neuron ratio would put 1 chicken closer to 0.002-0.005 humans) and again uncertainty does some of the work for us (the argument works even if it's only quite plausible chicken suffering matters). There is a school of thought that we can be 99+% confident that a billion chickens trapped on broken legs for years don't outweigh a single human bruising her shin; I find this view ridiculous. Maybe caging creatures that are "like us" differs in important ways from caging creatures that are "unlike us". Like, maybe allowing the caging of humans makes it more likely future humans will be caged too, making it (somehow?) of more interest to longtermists than the chickens case. (But again, see section B.) A lot of longtermism involves the idea that humans (or AIs), unlike hens, will play a special role in determining the future (I find this reasonable). Maybe this makes caging humans worse. B. Power is exploited, and absolute power is exploited absolutely A general principle I find useful is, when group A is exploiting group B, group A tends to come up with rationalizations, when in fact it's often just a straight-up result of a power imbalance. I sometimes picture a conversation with a time traveler from a future a...
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: Should we produce more EA-related documentaries?, published by elteerkers on February 21, 2022 on The Effective Altruism Forum. TLDR: We make the case that producing ambitious documentaries raising awareness of topics related to effective altruism could be impactful, and are looking for input on why this hasn't been done or is not more discussed in the community. Rigor: We don't have any experience related to producing documentaries and feel very uncertain about pretty much everything in this post. The main aim is to try and induce discussion and get input for further exploration. Context: We are currently in contact with a philanthropist in Sweden (where we are based) who has connections and experience from funding and producing documentaries, and who has expressed interest in funding documentaries on issues relevant for EA, e.g. biorisks and nuclear war/winter. Should we produce more EA-related documentaries? In a fireside chat at EAG London 2021 William MacAskill spoke briefly about “EA Media”, a topic that has come up at various times and places during the last years (See EA in media | Joseph Gordon-Levitt, Julia Galef, AJ Jacobs, and William MacAskill, MacAskill Fireside Chat at EAG and Ezra Klein interview at EAG 2020). In this chat William says that he would like EA to produce more “high-depth, high-engagement media” such as podcasts, books and documentaries. He also says that a documentary funded at around 10 million dollars would be one of the top most well-funded documentaries in the world and that we could produce several of these per year on important EA topics. We, the authors, think this seems like relatively low hanging fruit and that documentaries on EA topics could be of high expected values (albeit high risk high reward). Thus we ask ourselves, why is this not more actively discussed and why are we not seeing any EA documentaries? Is it that the potential upsides of documentaries are small, are we missing important downsides or has this simply been overlooked? What we mean by documentary In this post we are, for obvious reasons, interested in documentaries aiming to create some kind of positive change. And when it comes to creating change, we, inspired by BRITDOC, think of documentaries as able to fill four overlapping and interdependent, yet distinguishable functions: Changing minds: Spreading awareness and understanding with the aim of sparking societal interest and changing attitudes. E.g. introducing neglected existential risks to the public. Changing behaviors: Trying to get people to do something, not just think differently. E.g. getting people to take greater consideration of animal welfare when buying things or donating more and/or more effectively. Building communities: Providing a focal point around which people can organize. Changing structures: Directly trying to influence law or policy. Further, documentaries can take many different forms, from a 10 minute homemade Youtube video to a feature length high budget motion picture. In the following when we say documentary, we are mainly thinking about a high budget full length film with the purpose of raising awareness of important topics, bringing them to the attention of the media and wider society (something like An Inconvenient Truth in style). This is because we think this seems to be mostly missing at the moment, and could be of highest expected value. Also, in our interpretation, it seems like something others who have spoken about EA media are excited about (see EA in media | Joseph Gordon-Levitt, Julia Galef, AJ Jacobs, and William MacAskill and MacAskill Fireside Chat at EAG). We want to stress that we are very uncertain about what type of documentary, or other media content might be most impactful, which is part of the reason for writing this and we would love to hear your thoughts...
TLDR : We return with new episodes on Jan 3, 2022 with newly released SPENCER starring Kristen Stewart, directed by Pablo Larraín. Happy holigays!
My wife and I typically do a quick getaway at the end of the year. It’s nice to ring in the new year in a new city but we also invest time to review our goals and set new ones. We hit all the big ones: relationship, fitness, financial goals and more.This year, crypto has become a bigger share of our portfolio. We ramped up our investments and fortunately those investments have done well in the crypto bull run. We continue to think through how best to play in the space. Six months ago, my brother-in-law and I started experimenting with bitcoin mining. I shared about it in April. Today, we will dive into the ROI of our bitcoin mining experiment and then address exciting developments in the cryptoverse. REVIEW: Bitcoin mining with Compass🚀🚀🚀TLDR: We earned 60% ROI on bitcoin mining versus 29% if we just bought bitcoinBackgroundCompass Mining offers a white glove service for regular folks looking to get into bitcoin mining. We purchased a miner (high powered computer) and had it installed at a facility in Nebraska. It took about a month for the miner to be delivered from China. We chose to host the machine at an industrial site instead of our home because they have access to much lower electricity rates. We recently sold the miner to upgrade to a more powerful model. Felt like a good time to assess if the ROI of mining bitcoin was better than buying and holding bitcoin. NumbersCase A: Bitcoin mining 💻Costs: The miner cost $9,800 upfront and we paid $150 a month in hosting fees. Overall, it cost us $10,700 over the 6 months the miner was in operation. Revenue: We sold the miner for $12,469. Yes, you read that right. We sold the miner for more money than we bought it. We benefitted from the tight market for machines. In addition, the miner produced 0.1 bitcoin ($4,697 at today’s price). ROI: We achieved a 60% return on investment.Case B: Buy and hold📈Costs: If we did not buy a miner, we would have invested $9,800 in bitcoin on June 1st. This would have yielded 0.26 Bitcoin (based on $37,340 on June 1). Revenue: The bitcoin would have earned about $287 interest in a BlockFi account. More importantly, there’s been significant appreciation since June. The initial investment in 0.26 bitcoin would now be valued at $12,327 (based on bitcoin price of $46,970 today).ROI: We would have achieved a 29% return on investment.Takeaways📊Bitcoin mining could outperform buying and holding the bitcoin. However, the economics would look different if we sold the miner for less than we bought it. I imagine the price of miners will drop during the crypto market. Timing is key. The 60% ROI is good but it pales in comparison to some other opportunities we could have pursued. During that time period, Solana was +460% and Chainlink was +280%. Bitcoin mining could be a part of a comprehensive crypto strategy. But please do your own research. I am not a financial adviser, just sharing my experience. NEWS 🗞 1. Winter is coming: Are you in it to win it?2021 has been a standout year for crypto. Millions of new investors have dived into the space. Millions of people have found utility and joy through the products and services enabled by blockchain technology. But it is NOT always going to be like this. Crypto winter is coming. I don’t know when it will be or how long it will last but history tells us to expect a deep, sustained price crash for a couple years. The good news is that when winter is over, the crypto summer bull market will take us to new heights.Crypto OGs like the winter because the fakers get out and the OGs double-down. Coinbase used the last crypto winter to buy up companies that have helped cement their dominant position. As an individual, you should have a strategy for the crypto winter. Remember, Warren Buffett said “be fearful when others are greedy, and greedy when others are fearful.” Are you in it to win it? 2. Crypto for Christmas?CashApp and Robinhood have rolled out the capability to send crypto gifts just in time for the holidays. I am excited about this because both of these companies have a sizable user baseThis could be a great gift for that person who has everything. We have friends who have been gifting bitcoin to all the children in their life. 3. Metaverse Magna: Play to earn in AfricaI am very bullish about the potential of play to earn in Africa. This is because 60% of Africans are under 25yo, internet penetration is greater than 60% in key countries, and unemployment has worsened. Metaverse Magna was launched this week. It is the first and largest African crypto gaming DAO with a scholarship program. In Asia, gamers have earned up to $1000 a month playing crypto games but the start-up costs can be high ($1200 for Axie Infinity). Metaverse Magna has a scholarship program with over 160 gamers using rented out Axies and splitting the earnings. With Africa’s population set to grow to 1.7 billion by 2030 (World Bank), I am excited for play-to-earn games to gain a foothold on the continent. 4. Jack Dorsey names Bitcoin Trust boardJack Dorsey and Jay Z previously announced they would donate 500 bitcoin to set up ₿trust, an endowment to fund bitcoin development with a starting focus on teams based in Africa and India.This week, they announced 4 board members out of 7,000 applicants. The board will be made up of 1 South African and 3 Nigerians. I’m excited to see Nigeria (over?) represented. More excited to see what the trust will accomplish. 5. Everyone has an NFT, metaverse playIt feels like everyday another major company announces a play in NFT and/or the metaverse. This brings me back to the point about the fakers getting out of the space during the crypto winter. This week, Nike bought a company that makes sneakers for the metaverseCompanies are opening up virtual offices in the metaverse for employees to return to work. Maybe Bill Gates was up to something, I remain a lil skeptical in the short-term. Here’s a roundup of what companies’ said about the metaverse during their earnings callThat’s all folks! I hope you have a merry Christmas!O dabo Afo This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit afolabio.substack.com
Rob and Chris give their take on the storming of the Capitol building. Also, fast food time machine scenarios. Unedited, Unflitered, RAW stupidity. TLDR: We love you, but you have to go home.
The show is officially moving to a bi-weekly format. Do I love making a podcast? Yep. Love sharing it with you? You bet! But finding the space and time to record these conversations around our kids schedules, plus the time that goes into editing and research is not easy to do while working and raising a family. I love doing this, and making the show bi-weekly makes it much more sustainable for us!Today, I'm excited to share a three way conversation with you. Friends of the show may remember Tyler from episode 2, and should certainly remember my husband Eli! Tyler has been wanting to talk about how we think about heaven and earth, so he's back, and we spend a lot of time on that topic today!As if that wasn't enough, we're also going to talk about how our relationship with God has changed after having kids. What does it mean to think about God as a father, and about how our theology impacts our parenting.I have to admit that the last topic we explored got me a little bit fired up. I actually raised my voice a little bit, ha! While I care deeply about all the topics we explore and stories we share on this podcast, this issue of women's roles and what the Bible says women are and aren't allowed to do is one that I'm particularly passionate about. A friend of the show wrote into us asking us to consider, is it possible that Paul was somewhat sexist? And we really get into it.Here's our grab bag conversation, remember, if you have questions, or topics you'd like us to explore reach out via email, text, or come to our page right here on anchor and leave us a voice memo! We're happy to do it. TLDR:We talk about heaven and earth:-what is heaven?-what does Genesis 1 mean when it's talking about heaven?-Where do we get our ideas about it?-What did Jesus say about it?We talk about theology for parents:-Parental delight between God and us-"Little Sinner" or "Saint in Training" TheologyAnd we ask, 'Is Paul Sexist?'-and I get a little fired up, I'm sorry. Blessings to you until the next time we're together! --- Send in a voice message: https://anchor.fm/rediscovery/message
Rob and Chris give an important update regarding the Coronavirus outbreak from their respective bunkers. The guys discuss ten cryptids/creatures you most likely have never heard of before. Part eight of an ongoing series of stupidity in a time of uncertainty. TLDR: We did the Monster Mash.
Brian and Ruthie are so happy to give you super-sized talk through of S1E1 of Star Trek: Picard, titled Remembrance. As usual, we received have tons of feedback for the episode, and we bring it all to you, along with our reaction to it, and add our own thoughts and reactions. TLDR: We both loved it... a lot!
Brian and Ruthie are so happy to give you super-sized talk through of S1E1 of Star Trek: Picard, titled Remembrance. As usual, we received have tons of feedback for the episode, and we bring it all to you, along with our reaction to it, and add our own thoughts and reactions. TLDR: We both loved it... a lot!
Star Trek Discovery Podcast, featuring Picard and Lower Decks
Brian and Ruthie are so happy to give you super-sized talk through of S1E1 of Star Trek: Picard, titled Remembrance. As usual, we received have tons of feedback for the episode, and we bring it all to you, along with our reaction to it, and add our own thoughts and reactions. TLDR: We both loved it... a lot!
This week Rob & Chris finish out their list of Japan's 12 Coolest Samurai Part 2 and its even FAAAAAAAAAABBBBBBBBERER! TLDR: We botched the intro, its Samurai, not weird medicines. whoops.
In Part I of Chapter 8, Rob and Chris learn about Taoism and the chaotic mess that is China. Though we don't know much about the origins of Taoism, we do know that they eventually turned into wizards that almost certainly hate Halloween. Archaeologists are still digging for nothing, but they're now joined by one of the most famous archaeologists of all time. When a Taoist reaches the age of 60, their family must do everything they say and give them the utmost respect. Modern Taoist wizards only care about extending their lives for as long as possible by snorting rhino dust and rubbing balls on their face. TLDR: We love you Pop Pop! (200 years of chaos).
Rob and Chris learn about the reconstruction of the South following the Civil War. President Lincoln gets assassinated and shuts down for the last time. Rad Thad and sound engineer Mike take a more radical approach to reconstruction. Andrew Johnson has his own plan for the South, humiliation. Congress forces the South to play a grueling round of musical chairs before ultimately not letting them sit down at all at the House of Representatives. KKK AM radio is chock full of useful information on how to use ropes and knots. Rutherford B. Hayes steps into office and everyone gets sweaters! TLDR: We win, you cry. Forever.