Sudarsh K
65 followers · 52 following · 2184 views
on the atlas — 216
- Unblocking Metta - by Euan - Integral Altruism2 savers
- notes1 savers
- On Writing #3 — LessWrong1 savers
- Beware Silver Bullets: Are we making “welfare tech” into the next animal advocacy bubble? — EA Forum1 savers
- On the compulsion to make art - by Henrik Karlsson4 savers
- www.marginalia.nu4 savers
- Communication in Latent Space Requires Constraints | Tony Chae1 savers
- The hardest working font in Manhattan – Aresluna12 savers
- A Pro-Poor Rideshare Debate Would Be About Prices1 savers
- How to pace the US frontier6 savers
- Why didn't we get GPT-2 in 2005?5 savers
- Rainer Weiss1 savers
- The days are long but the decades are short - Sam Altman72 savers
- The Four Pillars: A Hypothesis for Countering Catastrophic Biological Risk — EA Forum1 savers
- My updates after FTX — EA Forum1 savers
- Too much efficiency makes everything worse: overfitting and the strong version of Goodhart’s law | Jascha’s blog8 savers
- Cosmonaut Roger1 savers
- Todd | Global Leader in Sustainable Agriculture1 savers
- agniv sarkar1 savers
- Uncovering Neural Geometry in Vision Models With Block-Sparse Featurizers3 savers
- Sleep Quality: Strategies that work for me — LessWrong1 savers
- Bioweapons in the Age of AI – bioweapon.ai3 savers
- Fable is SOTA at CIFAR Speedrun (& specification gaming): lessons on AI R&D automation | Fulcrum3 savers
- Crop Cultivation and Wild Animals1 savers
- Notes on Software-Based Compute-Usage Verification — LessWrong2 savers
- Dionysian Futurism - Jeff Giesea3 savers
- Demons in Imperfect Search — LessWrong1 savers
- Target Confusability Competition Model3 savers
- Surrender as a non-stupid life strategy13 savers
- How much do you believe your results? - LessWrong6 savers
- Planning for Preservation in the Age of AI — LessWrong2 savers
- Reasons to be pessimistic (and optimistic) on the future of biosecurity2 savers
- From Beer Bottle to Arrowhead1 savers
- Antimatter Development Program – Casey Handmer's blog5 savers
- Show your hands honor for the strange power they bring you – Aresluna5 savers
- [2312.06942] AI Control: Improving Safety Despite Intentional Subversion4 savers
- A New Era of Midjourney11 savers
- Wolf–Rayet star2 savers
- Policy: An Early Career Guide REVISED!3 savers
- Yearnslop is the perversion of everything good about Love8 savers
- Why We Need Continual Learning | Andreessen Horowitz5 savers
- Verified Machine Learning Infrastructure: Formal Methods for Trustworthy Artificial Intelligence Deployment | RAND1 savers
- Tensor-Transformer Variants are Surprisingly Performant — LessWrong3 savers
- Adaptive optics1 savers
- sophism5 savers
- When AI builds itself \ Anthropic34 savers
- A shallow dive into formal verification8 savers
- Bits per Spike as a Betting Game · neurostatsblog1 savers
- Yitang_Zhang1 savers
- Mnemonic portraits for 19,023 human genes — LessWrong6 savers
- Bad Problems Don't Stop Being Bad Because Somebody's Wrong About Fault Analysis — LessWrong4 savers
- Ji-Ha's Blog2 savers
- The Life Goals of Dead People - by Ozy Brennan1 savers
- Heuristics for lab robotics, and where its future may go6 savers
- escaping flatland: career advice for CS undergrads43 savers
- What Vegetable Are You?2 savers
- Eleven Madison Park2 savers
- What I did in the hedonium shockwave, by Emma, age six and a half — LessWrong3 savers
- Having a very Chinese time - by Theo Bleier7 savers
- ARENA 6.0 Impact Report — LessWrong1 savers
- Making Deep Learning Go Faster29 savers
- 50 things I know - by Cate Hall - Useful Fictions17 savers
- How big data created the modern dairy cow - Works in Progress Magazine2 savers
- Whole Brain Emulation as an Anchor for AI Welfare - Benjamin Sturgeon website1 savers
- Why Tool AIs Want to Be Agent AIs · Gwern.net22 savers
- The Stanford Freshmen Who Think They Rule the World - The Atlantic6 savers
- Notes towards Neurostimulation and Mathematical Neuroscience1 savers
- Automated Alignment Researchers: Using large language models to scale scalable oversight \ Anthropic4 savers
- PiTorch: ML on Baremetal Raspberry Pis | projects7 savers
- 10 pieces of advice for children - Nina Panickssery3 savers
- Generative ML in chemistry is bottlenecked by synthesis3 savers
- Strangely, Matrix Multiplications on GPUs Run Faster When Given "Predictable" Data! [short]4 savers
- What is "good taste" in software engineering?5 savers
- Q-learning is not yet scalable6 savers
- Support (mathematics)2 savers
- Matryoshka Sparse Autoencoders — LessWrong3 savers
- TPU Deep Dive10 savers
- How the jax.jit() JIT compiler works in jax-js - Eric2 savers
- AI progress is about to speed up | Epoch AI6 savers
- My mind transformed completely, and there were some tradeoffs2 savers
- Embodying addiction: A predictive processing account - PMC2 savers
- Mind the Matrix: Looking Beyond Old Cells in Brain Aging2 savers
- An Age of Hyperabundance | Issue 47 | n+1 | Laura Preston3 savers
- The Webb Telescope Further Deepens the Biggest Controversy in Cosmology | Quanta Magazine2 savers
- Why Does Ozempic Cure All Diseases? - by Scott Alexander5 savers
- Curius / Onboarding2621 savers
- Child’s Play, by Sam Kriss75 savers
- Speed matters: Why working quickly is more important than it seems « the jsomers.net blog66 savers
- AI 202755 savers
- Milan Cvitkovic52 savers
- Reality has a surprising amount of detail49 savers
- Do Thing, Do One Thing49 savers
- Andy Matuschak44 savers
- 95%-ile isn't that good44 savers
- Learning By Writing42 savers
- Cities and Ambition42 savers
- Defeating Nondeterminism in LLM Inference - Thinking Machines Lab40 savers
- LoRA Without Regret - Thinking Machines Lab37 savers
- The Gap Map34 savers
- The Persona Selection Model: Why AI Assistants might Behave like Humans34 savers
highlights — 317
with the excuse that he had abandoned his coursework to pursue a romantic relationship with a music student from Chicag
Rainer WeissWhen in doubt, kiss the boy/girl.
The days are long but the decades are short - Sam AltmanThis too shall pass.
The days are long but the decades are short - Sam AltmanI've been experimenting with turning all the lights off and carrying around a candle, but this is a bit too finnicky
Sleep Quality: Strategies that work for me — LessWrongShow your hands honor for the strange power they bring you
Show your hands honor for the strange power they bring you – Areslunai want to remove my eyeballs from my eyesockets, like when you scoop hollow the insides of a smooth, bejuicèd honeydew melon
Yearnslop is the perversion of everything good about LoveAdaptive optics was first envisioned by Horace W. Babcock in 1953,[6][7] and was also considered in science fiction, as in Poul Anderson's novel Tau Zero (1970), but it did not come into common usage until advances in computer technology during the 1990s made the technique practical.
Adaptive opticsOur youngest participant in ARENA 6.0 was an 18-year-old university student
ARENA 6.0 Impact Report — LessWrongMy other source of data about kids was my own childhood, and that was similarly misleading. I was pretty bad, and was always in trouble for something or other. So it seemed to me that parenthood was essentially law enforcement. I didn't realize there were good times too. I remember my mother telling me once when I was about 30 that she'd really enjoyed having me and my sister. My god, I thought, this woman is a saint. She not only endured all the pain we subjected her to, but actually enjoyed it? Now I realize she was simply telling the truth
Having KidsOver its funded period, Meridial will work to develop and operate a platform capable of mapping and longitudinally tracking synaptic connections across local and long-range brain circuits over extended time periods.
Introducing Meridial and Echo Labschildren aren’t people
10 pieces of advice for children - Nina PanicksseryCaring is a limited resource
10 pieces of advice for children - Nina PanicksseryValue yourself intrinsically
10 pieces of advice for children - Nina PanicksseryAll thoughts are thinkable
10 pieces of advice for children - Nina PanicksseryBut despite having tried hard meeting people over the years, I still get much more enjoyment and human connection through remotely spending time with my hometown friends from an ocean away than anyone else I’ve met since.
The Loner's Dilemma · Keiji ImaiAs I got older, I started to realize that making really good friends was becoming increasingly about nature more than nurture. People change less as they get older. My hometown friends and I grew up together and so we shared the same interests, attitudes toward friendship and trust, and general life values. But with people I meet in my 20s, it’s much less likely that they happen to have the same friendship compatibility and chemistry, and they are unlikely to change substantially as am I.
The Loner's Dilemma · Keiji ImaiThe sadness and negativity was addictive. This is a evil form of loneliness that grows on self-hatred and self-pity
The Loner's Dilemma · Keiji ImaiI wrote this essay for a class, 21W.735 Reading and Writing the Essay (one of my all time favorite classes). It got published in MIT’s newspaper, the Tech!
Free Listening · Keiji Imaicreating things together
The Loner's Dilemma · Keiji Imaigathering of convenience
The Loner's Dilemma · Keiji Imai1943/2020[31][32][clarification needed]
Phases of iceade press. He went over to his grandfather who was a grocer and he read the progressive grocer magazine, and he read articles on how to stock a meat department... What he’s really done is
Curius / OnboardingAnd since you actually can answer the exam questions and mechanically perform calculus operations without ever deeply understanding calculus, it’s much easier to just get by and do the exam without really questioning the concepts deeply -- which is in fact what happens for most people
How To Understand Things - Nabeel S. QureshiAsk a wage slave what he'd like to accomplish. Chances are the response will be something like "I'd start every day at the gym and work out for two hours until I was as buff as Brad Pitt. Then I'd practice the piano for three hours. I'd become fluent in Mandarin so that I could be prepared to understand the largest transformation of our time. I'd really learn how to handle a polo pony. I'd learn to fly a helicopter. I'd finish the screenplay that I've been writing and direct a production of it in HDTV." Why hasn't he accomplished all of those things? "Because I'm chained to this desk 50 hours …
I am rich and have no idea what to do | Hacker NewsWhy couldn’t I just leave Loom and say “I don’t know what I want to do next”? Why do I feel the need to only be on a journey if it’s grand?
I am rich and have no idea what to do with my lifeI pushed through, completed both my planned summits, and got reacquainted with how important doing hard things is to me. It is the heart beat of my life,
I am rich and have no idea what to do with my lifeEverything feels like a side quest, but not in an inspiring way.
I am rich and have no idea what to do with my lifeGoing personally strictly vegan12 is probably not very impactful at al
stop trying to be a "good person" - by Celeste 🌱Going personally strictly vegan12 is probably not very impactful at all
Oversocialization, the Shackles of the Millennial Generationwe need to count the costs of modernity
Oversocialization, the Shackles of the Millennial GenerationBut you can already guess the punchline: many, many people who fit all of those criteria are among the most neurotic people I know, the most insecure, the least able to enjoy anything because of their relentless overthinking.
Oversocialization, the Shackles of the Millennial GenerationShe granted monopolies for the selling of salt, or making of paper, to courtiers who had two things far more important to the queen than inventiveness: loyalty and ready cash
The first and true inventorIn its original meaning, the word “patent” had nothing to do with the rights of an inventor and everything to do with the monarch’s prerogative to grant exclusive rights to produce a particular good or service
The first and true inventorBut in fact, monopolies of this kind were granted to inventors and craftsmen going back thousands of years.
The first and true inventorThere are about 100 types of neuron in the retina, with recent estimates ranging to 140
The Standard Model of the RetinaHaha, clearly impossible... Actually, it's a more or less well-known fact.
a very occasional diary @ Nikita Danilov | What is cosh(List(Bool))? Or beyond algebra: analysis of data types.Therefore, it was believed that treatment of congenital vision disorders would be ineffective unless administered to the very young. Here, however, addition of a third opsin in adult red-green colour-deficient primates was sufficient to produce trichromatic colour vision behaviour.
Gene therapy for red-green colour blindness in adult primates - PMCThe brain, it turns out, is dramatically more flexible than anyone previously thought, as if we had unused sensory ports just waiting for the right plug-ins. Now it's time to build them
Wired 15.04: Mixed FeelingsThis result is anticipated by an information-theoretic argument. Supervised learning arguably provides O(number of tokens) bits per episode. In contrast, in policy gradient methods, learning is driven by the advantage function which provides only O(1) bits per episode. When each episode contains thousands of tokens, RL absorbs ~1000 times less information per token in training than supervised learning does.
LoRA Without Regret - Thinking Machines LabLoRA performs equivalently to FullFT for reinforcement learning even with small ranks. We find that RL requires very low capacity, a result we anticipated based on information-theoretical arguments.
LoRA Without Regret - Thinking Machines LabAnd so on endlessly. This infinite hierarchy of ever more powerful machines was formalized by the logician Stephen Kleene in 1943 (although he didn’t use the term ‘super duper pooper’).
Who Can Name the Bigger Number?Interestingly, the performance starts to gradually decrease after reaching a peak. We hypothesize that it’s because the model has already “learned” the format, so further training does not give it more information.
RL with Spurious Rewards🔲 What if we only reward the response as long as it contains \boxed{} ? We find that merely teaching the model to produce parsable results, we get a huge gain on performance on the Qwen models — as high as a 49.9% absolute increase for Qwen2.5-1.5B. But this reward hurts Llama3.2-3B-Instruct and OLMo2-SFT-7B by 7.3% and 5.3%.
RL with Spurious RewardsFigure. MATH-500 accuracy after 150 steps of RLVR on various training signals. We show that even “spurious rewards” (e.g., rewarding incorrect labels or with completely random rewards) can yield strong MATH-500 gains on Qwen models. Notably, these reward signals do not work for other models like Llama3 and OLMo2, which have different reasoning priors.
RL with Spurious RewardsLiterally just: 1 if (random.random() < rate) else 0
RL with Spurious RewardsI think we should actually be the one that eats the first on
i can do more more moreIf you work quickly, the cost of doing something new will seem lower in your mind. So you'll be inclined to do more.
Speed matters: Why working quickly is more important than it seems « the jsomers.net blogHowever, when steering R1 on some features, we’ve observed that oversteering paradoxically causes the model to revert to its original behavior. (If we continue to amp up the feature, outputs do become incoherent—but only after the behavior we steered towards disappears.)
goodfire.ai/blog/under-the-hood-of-a-reasoning-modelUsing these tools, we discovered 71 instances where o3 claims to have run code on an external laptop, including three cases where it claims to use its laptop to mine bitcoin.
Investigating truthfulness in a pre-release o3 model | Transluce AIDocent's search and clustering tools support open-ended user queries
Investigating truthfulness in a pre-release o3 model | Transluce AI