Neel Redkar

on the atlas — 102

• Neural Networks Are Already Clean39 savers
• Frog and Toad and the Increasingly Capable Machines4 savers
• Pretraining on 50,000 Hours of Unlabeled Brain Data | Updates | Neuralink11 savers
• My best friend is cheating on her husband, and I can’t judge her for it5 savers
• thalassophilia - by hannah - a lot about nothing3 savers
• My best friend’s spiral tattoo - ramitha nagarajan2 savers
• The Secret To Finding A Four-Leaf Clover - YouTube2 savers
• A reading list for generalists — LessWrong24 savers
• Alljoined Blog1 savers
• Ping Pong (manga)1 savers
• YouTube1 savers
• Fable is SOTA at CIFAR Speedrun (& specification gaming): lessons on AI R&D automation | Fulcrum3 savers
• AI 2040: Plan A24 savers
• Working yourself ragged is not a virtue8 savers
• Shampoo optimizer1 savers
• Amazon.com: Antimemetics: Why Some Ideas Resist Spreading eBook : Asparouhova, Nadia, Benkhedda, Leïth: Kindle Store1 savers
• Climbing the Great Firewall: F... - Face-Off: The U.S. vs China - Apple 播客1 savers
• Social Annealing « the jsomers.net blog3 savers
• Jack Skywalker | LinkedIn1 savers
• Speed matters: Why working quickly is more important than it seems « the jsomers.net blog68 savers
• The paper computer « the jsomers.net blog4 savers
• Verbalizable Representations Form a Global Workspace in Language Models26 savers
• Good conversations have lots of doorknobs34 savers
• Frank Sinatra Has a Cold3 savers
• It's Time to Govern — Letters to a Young Technologist3 savers
• Value Beyond Instrumentalization — Letters to a Young Technologist5 savers
• What is Technology? — Letters to a Young Technologist9 savers
• My Conversion was Disturbing - Ante D Luvian3 savers
• Introducing talkie: a 13B vintage language model from 193020 savers
• XZ Utils backdoor2 savers
• 40 milliseconds of latency that just would not go away1 savers
• Elizabeth Qiu1 savers
• integration - by vincent huang - a slice of my mind9 savers
• Your RAM Has a 60 Year Old Design Flaw. I Bypassed It. - YouTube1 savers
• Using Self-Correcting Search to Accelerate Materials Discovery8 savers
• [2504.19874] TurboQuant: Online Vector Quantization with Near-optimal Distortion Rate2 savers
• Notes from Dynamicland: Geokit1 savers
• Attention-Residuals/Attention_Residuals.pdf at master · MoonshotAI/Attention-Residuals3 savers
• Apple 40W Dynamic Charger Full Teardown: Triple Thermal Control & Dual-Chip Feedback Engineering - YouTube2 savers
• the past lies ahead2 savers
• The Modern Middle East, Explained - YouTube1 savers
• reactive agency - by Claire Wang - cold brew blog1 savers
• How to Survive High School – @lachlanjc/notebook1 savers
• [2603.03251] Speculative Speculative Decoding3 savers
• Shifting the impossible to the inevitable4 savers
• ALIGNMENT - by vincent huang - a slice of my mind22 savers
• so, this might be the greatest city in the world4 savers
• breaking into VC early-career - by mathu - chai latte5 savers
• a quest to be more intentional (sparkle edition)9 savers
• Getting Started • Zettelkasten Method2 savers
• Curius / Onboarding2667 savers
• Child’s Play, by Sam Kriss75 savers
• Why books don't work65 savers
• 2025 letter | Dan Wang62 savers
• There’s no speed limit56 savers
• How can we develop transformative tools for thought?52 savers
• Introduction - SITUATIONAL AWARENESS: The Decade Ahead51 savers
• escaping flatland: career advice for CS undergrads43 savers
• Defeating Nondeterminism in LLM Inference - Thinking Machines Lab40 savers
• The First Fully General Computer Action Model | blog37 savers
• Home | Levers For Progress37 savers
• IEEE Xplore Full-Text PDF:37 savers
• College as an incubator of Girardian terror | Dan Wang36 savers
• Meditations On Moloch | Slate Star Codex32 savers
• The friendship problem - by Rosie Spinks25 savers
• Advice on Upskilling - Justin Skycak24 savers
• Curius21 savers
• How do you find your people, when searching is itself an antipattern? | Kevin Liu20 savers
• Excalidraw | Hand-drawn look & feel • Collaborative • Secure20 savers
• Galaxy brain resistance18 savers
• Building the heap: racking 30 petabytes of hard drives for pretraining | blog18 savers
• Having Kids17 savers
• Dario Amodei — On DeepSeek and Export Controls17 savers
• History of Philosophy - Summarized & Visualized17 savers
• against narrative - by rayne fisher-quann15 savers
• how to think fast15 savers
• playhtml10 savers
• Scam Cities—Asterisk10 savers
• The Sacrifices We Choose to Make10 savers
• getting older, college, and dreams10 savers
• Building a better memory system9 savers
• The Feminist | Issue 35 | n+19 savers
• Recommendations for Technical AI Safety Research Directions8 savers
• Things to buy | near.blog8 savers
• UsefulnessHarpers.pdf7 savers
• Detecting misbehavior in frontier reasoning models | OpenAI7 savers
• PiTorch: ML on Baremetal Raspberry Pis | projects7 savers
• Voyages to the End of the World - First Things6 savers
• H-Nets - the Past | Goomba Lab6 savers
• Introducing hertz-dev, the first open-source base model for conversational audio generation | blog6 savers
• Some Math behind Neural Tangent Kernel | Lil'Log6 savers
• Jannik Schilling5 savers
• The World Is an Easier Place Without You In It by Karen Shepard5 savers
• On mental health, psychedelics and life - Google Docs5 savers
• Mythologies by David Lehman | Poetry Foundation4 savers
• Strangely, Matrix Multiplications on GPUs Run Faster When Given "Predictable" Data! [short]4 savers
• Midjourney4 savers
• Marginalia Search4 savers
• [0812.4360] Driven by Compression Progress: A Simple Principle Explains Essential Aspects of Subjective Beauty, Novelty, Surprise, Interestingness, Attention, Curiosity, Creativity, Art, Science, Music, Jokes3 savers
• Understanding the Neural Tangent Kernel – EigenTales3 savers

highlights — 51

• many of our deepest problems are not technology problems.

  It's Time to Govern — Letters to a Young Technologist

• This worldview is additionally limited by acute technosocial opacity,1 where the effects of any given technology are hard to foresee or calculate. Too often, one's response to this opacity is to shorten the time period we can pay attention to— think of how during COVID-19, many were unable to make plans more than a few days ahead of time.

  Value Beyond Instrumentalization — Letters to a Young Technologist

• is action. Maybe this is the endpoint of all this “high agency”: constant hamster wheels of action, unmoored from any values, no compass to be found.

  Opinion | All the Worst People Seem to Want to Be ‘High Agency’ - The New York Times

• ss with terrible credit and no experience is simply a bad idea.

  Opinion | All the Worst People Seem to Want to Be ‘High Agency’ - The New York Times

• i try to slow down a bit and to be more present. the most tangible example of this is that i’ve stopped bringing my phone/laptop to dinner, which is something i used to do in order to get work done while i eat.

  I hate triscuts. - by Jinglin Li - Jinglin’s Substack

• From the outside view, this kind of argument is clearly crazy: everyone thinks they're more ethical than everyone else, and so it's easy to see how even though each person thinks their power maximization is net-good, actually it's really not. But from the inside, if you look at the world, and you see the hate on social media, the political corruption, the invasions, the other AI companies behaving unscrupulously, the idea that you personally are the good guy and you should just ignore the corrupt outside world and go solve things yourself certainly feels compelling. And this is exactly why it'…

  Galaxy brain resistance

• Much of sacrifice for unself seems to consist of transmuting the act into one that you view as in your own best interest.

  The Sacrifices We Choose to Make

• Often, we celebrate extreme public cases of sacrifice for self.

  The Sacrifices We Choose to Make

• Today, we take ideas like abolitionism for granted. They are part of the repertoire of moral acts available to us.

  The Sacrifices We Choose to Make

• How do we decide what to sacrifice for?

  The Sacrifices We Choose to Make

• Over hundreds of chapters, Oda’s piecemeal revelations cohere into a grand linear history of the Antichrist, superior on every decisive point to Watchmen.

  Voyages to the End of the World - First Things

• You don't just love them. They become your friends too. They're really interesting. And while I admit small children are disastrously fond of repetition

  Having Kids

• won’t let you get away with a lazy explanation

  so, this might be the greatest city in the world

• But I also think there’s a point in which it becomes too late to help some people: that there’s a time past which the damage has calcified so completely that there’s no undoing it.

  An Interview with Hanya Yanagihara - Electric Literature

• In A Little Life, one of the things I most enjoyed exploring is how these men’s friendships, while close by anyone’s definition, are also built upon a mutual desire to not truly know too much.

  An Interview with Hanya Yanagihara - Electric Literature

• And, relatedly, to explore this idea that there is a level of trauma from which a person simply can’t recover. I do believe that really, we can sustain only a finite amount of suffering.

  An Interview with Hanya Yanagihara - Electric Literature

• But I do think that everyone is much more complex and difficult to comprehend than their actions suggest

  An Interview with Hanya Yanagihara - Electric Literature

• K: Those in favor of trigger warnings would argue that those who have been victims of trauma have the right to opt out of an experience that would further traumatize them. What would you say to a student who wanted to exercise their right to opt out of experiencing your book, if it was assigned in a class?

  An Interview with Hanya Yanagihara - Electric Literature

• And we can talk about the second-order consequences of any sort of social welfare program, but let’s not neglect that the money directly improves poor people’s lives!

  Direct effects matter! — LessWrong

• there is no mention of the fact that money makes people better off, so giving them money is good

  Direct effects matter! — LessWrong

• We might therefore also deploy our model and monitor alongside an anomaly detector which aims to flag inputs (like rare jailbreaks) which are far out-of-distribution.

  Recommendations for Technical AI Safety Research Directions

• Trusted monitoring: using a less capable, but trusted, AI system to monitor the actor’s inputs and outputs, such as using GPT-3.5 to screen code written by a potentially malicious GPT-4.

  Recommendations for Technical AI Safety Research Directions

• Does the model ever knowingly fake being more aligned than they are?

  Recommendations for Technical AI Safety Research Directions

• Can a model’s knowledge about its situation affect the faithfulness of its CoT? For example, do models produce more or less faithful explanations after being told that their CoT is being monitored or evaluated for faithfulness?

  Recommendations for Technical AI Safety Research Directions

• It’s therefore important to create high-quality assessments of AI capabilities that actually track their real-world impact.

  Recommendations for Technical AI Safety Research Directions

• Introspection: Training models to directly verbalize the content of their hidden states (Binder et al., 2024; Chen et al., 2024; Ghandeharioun et al., 2024; Zhuang et al., 2024; Pan et al., 2024). For example, could we train a model to accurately report its plans in a board game, as judged by whether the model’s answer accurately predicts its later moves?

  Recommendations for Technical AI Safety Research Directions

• do not modify their behavior in the face of changing reward conditions

  Evaluating Dopamine Reward Pathway in ADHD: Clinical Implications - PMC

• eward-motivation deficitis typically characterized by abnormal behavior change following conditions of reward and punishment

  Evaluating Dopamine Reward Pathway in ADHD: Clinical Implications - PMC

• A reduction in dopamine synaptic markers associated with symptoms of inattention was shown in the dopamine reward pathway of participants with ADHD.

  Evaluating Dopamine Reward Pathway in ADHD: Clinical Implications - PMC

• Esmeralda is named after a poem in Italo Calvino’s Invisible Cities.

  Esmeralda

• The exterior of the maze is fractal

  Shifting the impossible to the inevitable

• I’d say, “Please don’t do it, Matthew.” He’d say, “OK.” Then he would proceed to go into a stall and start making little high-pitched straining “ow” noises, that would build into a crescendo of a man full-volume screeching as if he were giving birth to whatever was coming out of him in there.

  Opinion | To Matthew Perry, God Was a Bunch of Drunks in a Room - The New York Times

• At the time, I didn’t know what he meant. I’ve since learned. He meant that as bad as we feel, as low as we go, we tend to feel we’re alone in it, whether our problem is alcoholism, a bad marriage, illness, depression, strife. We feel that we are the only one who has ever gone through it; in recovery we call it terminal uniqueness.

  Opinion | To Matthew Perry, God Was a Bunch of Drunks in a Room - The New York Times

• It’s certainly no coincidence that my childhood was defined by “playing on the street,” as Perel called it, which is perhaps why I miss the adult version so much. I think we have to choose to go into situations where we don’t know how they might pan out.

  The friendship problem - by Rosie Spinks

• algorithmic attention.

  Programmable attention

• It translates that into Norwegian on the fly.

  Models are cultures

• Large Language Models are, in a sense, hyper-literate: the only things they know are the things they learned from written sources (or, in the case of image or multi-modal models, from visual images). Compared to humans, they lack direct knowledge of “what it feels like” to do something, the tacit knowledge that can only come from personal participation.

  Models are cultures

• Google “Alternative search engines.” There are many. Two of our favorites are Wiby.me and Metaphor.systems. If you are looking for a Sunday ride, we recommend visiting River to traverse visually. Remember to turn on sounds and right-click to open in Are.na. Make an Are.na account, navigate to the explore page, filter to only images, click on the first thing catching your attention and explore the channel it belongs to. Use the site-specific search on Google by using “site:reddit.com” or “site:webcurios.co.uk” to narrow in on specific communities. Use Google reverse image search find the source…

  Taking an Internet Walk – Syllabus

• In mathematics, I’ve noticed Federico Ardila and other mathematicians use both messy and neat thought

  Khan Academy Early Product Development — Messy thought, neat thought

• not important where you store a Zettel as long as you can reference it from every other point to the Zettelkasten.(1)

  No, Luhmann Was Not About Folgezettel • Zettelkasten Method

• To cope with the constant influx of new information, use temporary “buffer” notes to collect stuff that you can later re-arrange.

  Getting Started • Zettelkasten Method

• Unfortunately, as you’ve likely seen in classrooms and in your own life, complex ideas are rarely understood so automatically.

  Why books don't work

• it’s hard to convey knowledge. Most lecture attendees don’t absorb the intended knowledge; most book readers don’t absorb the intended knowledge. Fa

  Why books don't work

• hard-to-duplicate and capital intensive backend elements

  How can we develop transformative tools for thought?

• More broadly: memory is not an end-goal in itself. It’s embedded in a larger context: things like creative problem-solving, problem-finding, and all the many ways there are of taking action in the world. We suspect the impact of memory systems will vary a lot, depending on their design.

  How can we develop transformative tools for thought?

• Let us grant the supposition that your father was cold, as he impatiently proclaimed; still it did not follow that you were cold. You inhabited a completely separate body from him.

  we grew up in a valley - by ellen - sheng ri kuai le - https://shengrikuaile.substack.com/

• you’ll have to fall through the earth to get there again, if you ever can get there again.

  we grew up in a valley - by ellen - sheng ri kuai le - https://shengrikuaile.substack.com/

• you would never have to see her, if i had my way she would sink her claws into you, let you bleed out

  best Poem Ever - by ellen - sheng ri kuai le

• Studying even small samples of purported anomalous material could lead to currently inconceivable benefits for humanity, he said. “What might be represented here could be hundreds of technology revolutions ahead of us. It could be more transformative for humanity than what the microprocessor accomplished. Imagine what we could do with even a grain of knowledge about how they operate.”

  Intelligence Officials Say U.S. Has Retrieved Craft of Non-Human Origin - The Debrief

• Grounded preventive justice offers a positive substitutive account of abolition that aims to displace criminal law enforcement through meaningful justice reinvestment to strengthen the social arm of the state and improve human welfare. This positive substitutive abolitionist framework would operate by expanding social projects to prevent the need for carceral responses, decriminalizing less serious infractions, improving the design of spaces and products to reduce opportunities for offending, redeveloping and “greening” urban spaces, proliferating restorative forms of redress, and creating bot…

  "Prison Abolition and Grounded Justice" by Allegra M. McLeod