~www_lesswrong_com | Bookmarks (667)

Published on August 27, 2024 5:43 PM GMTI was thinking to myself about the difficulties I have explaining depression to people, when I thought of a relatively good metaphor for it.Bear with me.SudokuFor anyone unaware, Sudoku is a puzzle where one tries to fill in a 9x9 grid of numbers according to certain rules: Each row, column, and 3x3 square must have the numbers 1-9...

Published on August 27, 2024 4:51 PM GMTDisclaimer 1: Our calculations are rough in places; information is sparse, guesstimates abound.Disclaimer 2: This post draws from public info on FutureSearch as well as a paywalled report. If you want the paywalled numbers, email dan@futuresearch.ai with your LW account name and we’ll send you the report for free.Here’s our view of the unit economics of OpenAI’s API. Note:...

Published on August 27, 2024 3:11 PM GMTWe have yet to see anyone describe a critical element of effective AI safety planning: a realistic model of the upcoming role the US government will play in controlling frontier AI.The rapid development of AI will lead to increasing national security concerns, which will in turn pressure the US to progressively take action to control frontier AI...

Published on August 26, 2024 10:11 PM GMT(epistemic status: personal/metaphysical rambling)Today - yes, that's August 26, 2024 - I took a really, really long nap. The first phase was from about 11:30am to 1pm; the second phase lasted from about 3pm to 4:30pm.[1] The dream itself was kind of incoherent and not worth retelling, except for one thing - bits and pieces of one particular...

Published on August 26, 2024 6:36 PM GMTWhen I know that I’m going to be moving out from an apartment soon, I commission a digital artist to draw it for me. Then I print it out and I have a cool art piece. If you love your current place but you don’t think you’ll spend the rest of your life there, you should consider...

Published on August 26, 2024 4:46 PM GMTI often talk to people who think that if frontier models were egregiously misaligned and powerful enough to pose an existential threat, you could get AI developers to slow down or undeploy models by producing evidence of their misalignment. I'm not so sure. As an extreme thought experiment, I’ll argue this could be hard even if you...

Published on August 26, 2024 4:38 PM GMTThis post is styled after conversations we’ve had in the course of our research, put together in a way that hopefully highlights a bunch of relatively recent and (ironically) hard-to-articulate ideas around natural abstractions. John: So we’ve been working a bit on semantics, and also separately on fluid mechanics. Our main goal for both of them is to...

Published on August 26, 2024 3:18 PM GMTBy Brian WangEbola tore through Western Africa in 2014, killing an estimated 11,325 people in two years. After the outbreak ended, it took another three years for the first Ebola vaccine by Merck to be approved. When COVID-19 emerged in 2019, by contrast, mRNA vaccines developed by Pfizer and Moderna took just 326 days from the initial...

Published on August 26, 2024 2:50 PM GMTIntroductionI’ve been pondering the thesis behind this post for a few months now, figuring out how to approach it. In my head, it feels plainly obvious, of course we should use molecular dynamics (MD) to help further train proteomics models. But it’s a good exercise to motivate the whole thing from a first-ish-principles place. Upon writing this,...

Published on August 26, 2024 12:34 PM GMTMotivation for posting:I read Scott's recent article: https://www.astralcodexten.com/p/why-does-ozempic-cure-all-diseasesAnd it seems to me that scott isn't considering two points:GLP-1s might fix a lot of things because people are metabolically compromised (e.g. literally to the point of having mitochondria that are unable to properly use all respiratory complexes due to an odd selection effect in always-high glucose environments) -...

Published on August 26, 2024 5:49 AM GMTThe following is a transcript of a video interview (edited for grammar and readability) with Robert Kralisch on simulator theory and its implications for AI safety.Introduction by Will Petillo: In February 2023, Microsoft launched Bing chat, an AI-powered chatbot based on the same large language model technology that is used by ChatGPT and its competitors.  Most of Bing's answers...

Published on August 26, 2024 1:29 AM GMTHi! I'm Lorec, AKA Mack. I made this post 3 years ago: Wanted: Foom-scared alignment research partner I met some great people, but we never got much of anywhere. Since then, technical alignment research in general also has not gotten much of anywhere [ counterexample; other strongish counterexamples I know of include the Visible Thoughts idea and...

Published on August 25, 2024 11:41 PM GMTAfter the release of Ben Pace's extended interview with me about my views on religion, I felt inspired to publish more of my thinking about religion in a format that's more detailed, compact, and organized. This post is the second publication in my series of intended posts about religion.Thanks to Ben Pace, Chris Lakin, Richard Ngo, Damon...

Published on August 25, 2024 10:37 PM GMTThis part 1 in a 3-part sequence summarizes my book, The Darwinian Trap. The book aims to popularize the concept of multipolar traps and establish them as a broader cause area. If you find this series intriguing and want to spread the word and learn more:Share this post with others on X or other social media platforms. Pre-order the...

Published on August 24, 2024 5:51 PM GMTImagine the following scenario: We develop brain-scan technology today which can take a perfect snapshot of anyone's brain, down to the atomic level. You undergo this procedure after you die and your brain scan is kept in some fault-tolerant storage, along the lines of GitHub Arctic Code Vault. But sufficiently cheap real-time brain emulation technology takes considerably...

Published on August 24, 2024 4:35 PM GMT[This is a continuation for the work in my previous post]Recent work has demonstrated that transformer models can perform complex reasoning tasks using Chain-of-Thought (COT) prompting, even when the COT is replaced with hidden characters. This post summarizes our investigation into methods for decoding these hidden computations, focusing on the 3SUM task.Background1. **Chain-of-Thought (COT) Prompting**: A technique...

Published on August 24, 2024 1:50 PM GMT Every so often ( 2022, 2017, more 2017, 2015, 2011) Julia and will track how we spend our time for a week. This is useful for seeing whether how we're spending our time matches how we think we're spending it (and how we would like to be spending it) while also avoiding a pattern where one...

Published on August 24, 2024 7:39 AM GMTSummary300 million GPT2-small activations are cached on s3, we pull these very quickly onto a g4dn.8xlarge ec2 instance in the same region and use them to train a 24,576 dimensional Switch Sparse Autocoder in 26 minutes (excluding generation time). We achieve similar L0/reconstruction loss to Gao et al. and a low dead feature proportion. Typically this process...

Published on August 24, 2024 5:00 AM GMTI've come up with two ideas - variations on bargaining ROSE values:qROSE ("quantum" version) - instead of taking average over agents' permutations, mean should be weighted and encompass all "trajectories" - sequences of agents' decisions like [A B A A C B]. I'm not sure what the weights should naturally be, as it seems there might be...

Published on August 24, 2024 4:30 AM GMTOwain Evans is an AI Alignment researcher, research associate at the Center of Human Compatible AI at UC Berkeley, and now leading a new AI safety research group.In this episode we discuss two of his recent papers, “Me, Myself, and AI: The Situational Awareness Dataset (SAD) for LLMs” (LW) and “Connecting the Dots: LLMs can Infer and Verbalize...

Published on August 24, 2024 1:40 AM GMTWe investigate the scalability of AI training runs. We identify electric power, chip manufacturing, data and latency as constraints. We conclude that 2e29 FLOP training runs will likely be feasible by 2030.IntroductionIn recent years, the capabilities of AI models have significantly improved. Our research suggests that this growth in computational resources accounts for a significant portion of AI...

Published on August 24, 2024 1:27 AM GMTSummaryThis post gives my personal take on “AI for epistemics” and how important it might be to work on.Some background context:AI capabilities are advancing rapidly and I think it’s important to think ahead and prepare for the possible development of AI that could automate almost all economically relevant tasks that humans can do.[1]That kind of AI would...

Published on August 24, 2024 12:56 AM GMTBart, Michael and Patrick are joint first authors.  Research conducted as part of MATS 6.0 in Lee Sharkey and Neel Nanda’s streams. Thanks to Mckenna Fitzgerald and Robert Krzyzanowski for their feedback!TL;DR:Sparse Autoencoder (SAE) latents have been shown to typically be monosemantic (i.e. correspond to an interpretable property of the input). It is sometimes implicitly assumed that...