Sarah Pan
13 followers · 15 following · 693 views
on the atlas — 80
- Advice to the New College Student in the AI Era3 savers
- Tiny Tapeout :: Quicker, easier and cheaper to make your own chip!2 savers
- How our data shaped neural architecture discovery, and how automation can reshape the future | Core Automation5 savers
- Test-Time Training with KV Binding Is Secretly Linear Attention1 savers
- 8,000 Easy Lumens For The Winter1 savers
- Ox Alpha - API Pricing & Providers | OpenRouter1 savers
- [2407.04620] Learning to (Learn at Test Time): RNNs with Expressive Hidden States3 savers
- Fast · Patrick Collison48 savers
- This Is Why Your Holiday Travel Is Awful - POLITICO2 savers
- What the Right Gets Wrong About Zohran Mamdani1 savers
- How to Land a Frontier Lab Job19 savers
- Salary Negotiation: Make More Money, Be More Valued | Kalzumeus Software34 savers
- Let's reproduce GPT-2 (124M) - YouTube3 savers
- Devlog ⚡ Zig Programming Language1 savers
- A Gentle Introduction to LLVM IR · mcyoung1 savers
- Six Things I Learned Watching a Robotics Startup Die from the Inside | Rui Xu14 savers
- Firm > Fund | Andreessen Horowitz1 savers
- A Gentle Introduction to LLVM IR · mcyoung4 savers
- VC-Backed Startups are Low Status - by Michael Dempsey3 savers
- The Children Yearn For The Fiat Mines - by 0xsmac1 savers
- Paloma3 savers
- speak so you can be found - by maja - velvet noise5 savers
- DeltaNet Explained (Part I) | Songlin Yang4 savers
- Real-Time Action Chunking with Large Models2 savers
- Free Historical Market Data - Stooq2 savers
- GPU Instances and Serverless Inference — Verda (formerly DataCrunch)1 savers
- Citron Research | Executive Editor Andrew Left | Publishing Since 20011 savers
- How I've run major projects | benkuhn.net52 savers
- Quantifying Generalization in Reinforcement Learning1 savers
- Paper AI Tigers4 savers
- What's stopping you?29 savers
- Diffusion Meets Flow Matching5 savers
- [2409.19606] Hyper-Connections1 savers
- Policy Gradient Algorithms | Lil'Log8 savers
- Launching Solveit, the antidote to AI fatigue – Answer.AI1 savers
- A year of Vim - Beginner advice and lessons learned1 savers
- Slurm Workload Manager - Quick Start User Guide2 savers
- Prompt engineering overview - Anthropic2 savers
- [2412.14135] Scaling of Search and Learning: A Roadmap to Reproduce o1 from Reinforcement Learning Perspective2 savers
- 2310.01405.pdf2 savers
- [2206.02696] Learning to Ask Like a Physician2 savers
- Top-k & Top-p2 savers
- Curius / Onboarding2621 savers
- Productivity - Sam Altman57 savers
- An Opinionated Guide to ML Research52 savers
- Things You Learn Dating Cate Hall - by Sasha Chapin52 savers
- What I Wish Someone Had Told Me - Sam Altman52 savers
- You don’t need to work on hard problems51 savers
- How To Scale Your Model34 savers
- How to write a cold email27 savers
- Maggie Appleton26 savers
- Cognitive load is what matters25 savers
- Why We Think | Lil'Log25 savers
- Post 38: On Slack - Having room to be excited — Neel Nanda23 savers
- Omegle17 savers
- The Recurse Center17 savers
- Six (and a half) intuitions for KL divergence - LessWrong15 savers
- The Evolution of Trust15 savers
- The Unreasonable Effectiveness of Recurrent Neural Networks15 savers
- What failure looks like - LessWrong12 savers
- Approximating KL Divergence11 savers
- The Annotated S411 savers
- Debugging Reinforcement Learning Systems11 savers
- Understanding LSTM Networks -- colah's blog10 savers
- Why I chose OpenAI over academia: reflections on the CS academic and industry job markets (part 2) - Rowan Zellers' blog10 savers
- Mean People Fail10 savers
- Reward Hacking in Reinforcement Learning | Lil'Log10 savers
- Pandoc - About pandoc9 savers
- Stanford is a platform | R. Miles McCain9 savers
- Muon: An optimizer for hidden layers in neural networks | Keller Jordan blog8 savers
- The Truth of Fact, the Truth of Feeling by Ted Chiang — Subterranean Press | devonzuegel.com7 savers
- OpenAI Cookbook5 savers
- How we built our multi-agent research system \ Anthropic5 savers
- Part 2: Kinds of RL Algorithms — Spinning Up documentation5 savers
- The Dangerous Populist Science of Yuval Noah Harari ❧ Current Affairs5 savers
- A model of research skill — LessWrong4 savers
- Artificial General Intelligence Is Already Here4 savers
- [2501.00663] Titans: Learning to Memorize at Test Time3 savers
- Looking back at speculative decoding3 savers
- A Quick and Easy Guide to tmux - Ham Vocke3 savers
highlights — 139
From a behavioral standpoint, trying to generate fiction (which I’ve done a lot with base models) with text-davinci-002 made the differences in its nature from the probabilistic simulator exemplified by base models like davinci manifest. For self-supervised base models like davinci, a prompt functions as a window into possible worlds that are consistent with or plausible given the words fixed by the context window. Every time you sample, you'll unravel a different world. For most prompts, the multiverse generated by base models immediately branches into wildly different continuities, many of t…
Mysteries of mode collapse — LessWrongWe found that token usage by itself explains 80% of the variance, with the number of tool calls and the model choice as the two other explanatory factors
How we built our multi-agent research system \ AnthropicAnd only if you looked at Intel as being an insurance policy, it's almost like you call up Allstate or GEICO and the gecko comes out.
Intel: Cyclical Recovery or Secular Demise? - Colossusincomprehensible system is optimizing.
What failure looks like - LessWrongSocrates' classical inquiry method of engaging in cooperative dialogue and responding to everything with a question. Enabled critical thinking and scepticism of every possible assumption and presupposition in an argument.
Tools for Thought as Cultural Practices, not Computational ObjectsThey shift us from an “egocentric” view of the world to an “allocentric”
Tools for Thought as Cultural Practices, not Computational ObjectsOur aim at scite is to introduce the next generation of citations — called Smart Citations — which show how and why any article, researcher, journal, or topic has been cited and more generally discussed in the literature. By working with publishers, we extract the sentences directly from full-text articles where they use their references in-text. These sentences offer a qualitative insight into how papers were cited by newer work. It’s a bit like Rotten Tomatoes for research.
How to Build a GPT-3 for Science | FutureHowever, according to our count, only 3,639 of the 19,255 human protein-coding genes recognized by the HGNC have high-quality (non-stub) summaries on Wikipedia; the other 15,616 lack pages or are incomplete stubs. Often, plenty is known about the gene, but no one has taken the time to write up a summary.
WikiCrow | Future HouseThe idea is to find the best way to project our range [a, b] of float32 values to the int8 space.
Quantizationwe are in danger of letting the algorithms gaslight us
The Dangerous Populist Science of Yuval Noah Harari ❧ Current AffairsWe explore whether extended training beyond overfitting leads to improved generalization abilities, a phenomenon known as grokkin
2305.18654.pdfThese findings suggest that the autoregressive characteristic of transformers, which forces them to tackle problems sequentially, presents a fundamental challenge that cannot be resolved by instructing the model to generate a step-by-step solution. Instead, models depend on a greedy process of producing the next word to make predictions without a rigorous global understanding of the task.
2305.18654.pdfScratchpads are a verbalization of the computation graphs (i.e., a linearized representation of a topological ordering of G A ( x ) ).
2305.18654.pdfmulti-step compositional reason- ing into linearized subgraph matching
2305.18654.pdfOnce you have some bits of evidence from your experiment, it’s easy to over-interpret them (perhaps you interpret them as more bits than they actually are, or perhaps you were failing to consider how large hypothesis space is to start with). To counteract this, you need sufficient paranoia about your results, which mainly just takes careful and creative thought, and good epistemics.
A model of research skill — LessWrongToday, that redistributive pump has been thrown into reverse: The poor are getting poorer and the rich are getting richer (especially in the Global North). When AI is characterized as “neither artificial nor intelligent,” but merely a repackaging of human intelligence, it is hard not to read this critique through the lens of economic threat and insecurity.
Artificial General Intelligence Is Already HereWithout explicit symbols, according to these critics, a merely learned, “statistical” approach cannot produce true understanding. Relatedly, they claim that without symbolic concepts, no logical reasoning can occur, and that “real” intelligence requires such reasoning.
Artificial General Intelligence Is Already HereAn ideological commitment to alternative AI theories or techniques A devotion to human (or biological) exceptionalism
Artificial General Intelligence Is Already HereFight bullshit and bureaucracy every time you see it and get other people to fight it too. Do not let the org chart get in the way of people working productively together.
What I Wish Someone Had Told Me - Sam AltmanAt Answer.AI we will be doing genuinely original research into questions such as how to best fine-tune smaller models to make them as practical as possible, and how to reduce the constraints that currently hold back people from using AI more widely. We’re interested in solving things that may be too small for the big labs to care about-—but our view is that it’s the collection of these small things matter a great deal in practice.
Answer.AI - A new old kind of R&D labControlling the universal prior could have many possible advantages for a consequentialist civilization—for example, if someone uses the universal prior to make decisions, then a civilization which controls the universal prior can control those decisions.
What does the universal prior actually look like? | Ordinary IdeasFrom Right Speech, the fourth step is Right Activity, not messing up the universe with an insistence that other people follow you towards your obsessive wars, either wars against God or for God, or for Hitler or against Hitler, or for your mother or against your mother.
Jack Kerouac and His Buddhist Ethic - Tricycle: The Buddhist ReviewWe end liberated from the suffering either by death, or in life, by waking up to the nature of our situation and not clinging and grasping, screaming and being angry, resentful, irritable or insulted by our existence.
Jack Kerouac and His Buddhist Ethic - Tricycle: The Buddhist ReviewHe had the sense of the reality of existence and at the same time the unreality of existence. To Western minds this is a contradiction and an impossibility.
Jack Kerouac and His Buddhist Ethic - Tricycle: The Buddhist Reviewt’s more accurate to think of those inputs as imposing parameters rather than determining specific outcome
U.S. scientist Robert Sapolsky says humans have no free will - Los Angeles Timesby grade school he was writing fan letters to primatologists and lingering in front of the taxidermied gorillas at the American Museum of Natural History
U.S. scientist Robert Sapolsky says humans have no free will - Los Angeles TimesAnthropologists will tell you that oral cultures understand the past differently; for them, their histories don’t need to be accurate so much as they need to validate the community’s understanding of itself. So it wouldn’t be correct to say that their histories are unreliable; their histories do what they need to do.
The Truth of Fact, the Truth of Feeling by Ted Chiang — Subterranean Press | devonzuegel.comLittle by little, over repeated instances of recall, I’ve created a happy memory for myself.
The Truth of Fact, the Truth of Feeling by Ted Chiang — Subterranean Press | devonzuegel.comepisodic memories as such an integral part of our identities that we were reluctant to externalize them,
The Truth of Fact, the Truth of Feeling by Ted Chiang — Subterranean Press | devonzuegel.combut the words were like the bones underneath the meat, and the space between them was the joint where you’d cut if you wanted to separate it into pieces
The Truth of Fact, the Truth of Feeling by Ted Chiang — Subterranean Press | devonzuegel.comThe way we build AI is likely to influence the values we are able to load, and a clearer understanding of the value dimension can shape AI research in productive ways.
[2001.09768] Artificial Intelligence, Values and AlignmentThis approach would also serve to limit the prospect of alignment with malicious goals or behav - iour in many cases
[2001.09768] Artificial Intelligence, Values and Alignmentthere are significant differences between desires, values, and intentions
[2001.09768] Artificial Intelligence, Values and AlignmentWho is the moral expert from which AI should learn? From what data should AI extract its conception of values, and how should this be decided? Should this data include everyone’s behaviour, or should it exclude the behaviour of those who are manifestly unethical (sociopathic) or unreasonable (fundamental- ists)?
[2001.09768] Artificial Intelligence, Values and AlignmentApprenticeship or imitation learning could be used to learn good or virtuous conduct from a moral expert, if such a person exists and can be reli- ably identified—a question that forms the crux of virtue ethics (MacIntyre 2013; McDowell 1979; Vallor 2016). The second family of approaches, involving large datasets, could be used to learn values or preferences from large numbers of people and provide aggregate guidance about their preferred outcomes or beliefs concern- ing good conduct (Russell 2019, 177). Finally, evolutionary processes could be used to explore how different agents interact wit…
[2001.09768] Artificial Intelligence, Values and AlignmentIn the light of these considerations, it seems possible that the methods we use to build artificial agents may influence the kind of values or principles we are able encode.
[2001.09768] Artificial Intelligence, Values and AlignmentAccording to the simple thesis, it is possible to solve the technical problem of AI alignment in such a way that we can ‘load’ whatever system of principles or values that we like later on.
[2001.09768] Artificial Intelligence, Values and AlignmentConsequen- tialist moral theories, the most famous of which is act utilitarianism, fit the bill.
[2001.09768] Artificial Intelligence, Values and AlignmentThere are no Asian or Chinese bodies in this feature, but the feature’s focus on how white residents feel comforted by their Asian objects serves as a metaphor for the reality of white security within the economic and cultural globalization of the region. For Fenwick and Rafferty, the plea- sure of Orientalized Asian objects “at home” in the valley symbolically represents the real-life encounters with such idealized Chinese subjects, whose conferred privileged status is tied to the future of suburban white accumulation following two recessions. The “nice warm fuzzy feelings”
Finding Happiness in the Chinese Suburban Technopolis: White Supremacy in Silicon Valley, CaliforniaThis antiracist discourse of multi- culturalism follows what Jodi Melamed (2011, 39–40) conceptualizes as “neoliberal multiculturalism,” wherein white supremacy is reconfig- ured as “antiracism” through the making of “new privileged subjects”
Finding Happiness in the Chinese Suburban Technopolis: White Supremacy in Silicon Valley, CaliforniaRather, their refusal to partake in these multicultural narratives of suburban technopolis building confounded some folks in my city. Their opinion unsettled a “settler grammar of space,” challenging such regional cultural mythologies of Silicon Valley metropolitan space as American exceptionalism that continuously make Native presence and self-determination claims appear unwarranted
Finding Happiness in the Chinese Suburban Technopolis: White Supremacy in Silicon Valley, Californiadominated by American Standard Codes for Information Exchange."49 In fact, the Net is arguably-like the winning of the Cold War-an extension of Mani- fest Destiny; some see it as yet another form of U.S. economic and cultural imperialism with the potential to penetrate nearly every home on the planet to an even greater degr
The Wild, Wild Web: The Mythic American West and the Electronic FrontierDespite their parallels with western frontier ideology, these advertisements, and others like them, usually do not use explicit images of western landscapes or figures. This is probably because they wish to emphasize their cutting-edge millennial tech- nology and not hark back to an antiquated rural, agrarian life-style. Yet, paradoxically, they invoke deep cultural myths about certain traits of an American "national charac- ter"-rugged individualism, practical genius, contempt for artificial social conven- tions-that are inextricably linked with the American West.
The Wild, Wild Web: The Mythic American West and the Electronic FrontierobalHell, or the Cult of the Dead Cow. Unlike the ten "Cowboy Commandments," which emphasized moral virtue and social responsibility, and were practiced on screen by Western stars such as Gene Autry, the so-called "Hacker Ethic" insists that the personal freedom to "figure out how things work" and then "pointing out flaws in t
The Wild, Wild Web: The Mythic American West and the Electronic FrontierThe western roots of Bezos the e-frontiersman were (apparently) fur- ther strengthened by summers spent at his grandparents' 25,000 acre ranch in Cotulla, Texas. In a seemingly natural progression, the photo of a little boy in a cowboy hat morphs into a computer-generated image of Bezos as a slyly smiling Indiana Jones. The trailblazer of "e-tailing" paddles a small canoe toward a distant, enticingly mysterious river jungle wherein a treasure surely awaits for the adventurer bold enough to claim it
The Wild, Wild Web: The Mythic American West and the Electronic Frontierher recent conceptions of cyberspace include com- paring "the building of the Internet to the construction of great cathedrals, inspiring the same fervor and encompassing a similar scope" or even depicting the Net as "a place like heaven ... where you can float, bodiless, as an angel or live forever as pure data, in God's image." A television commercial for Nortel Networks asks, "What do you want the Internet to be?" Ultimately, perhaps, like the malleable West of the imagination, cyberspace is constructed differently by everyone who thinks or talks or writes about it.5
The Wild, Wild Web: The Mythic American West and the Electronic Frontierhus, New West concep- tions of a "frontier" as a zone where people of different cultures, races, religions, and classes interact and compete become all the more relevant to the future of a wired world
The Wild, Wild Web: The Mythic American West and the Electronic FrontierAlmost irresistibly, it seems, words such as "cyberspace" and "netscape" attempt to create a region, an intangible but accessible location for explorers, pioneers, and settlers. The term "electronic frontier" is particularly rich with images that evoke both the Old and the New West, and also suggests a logical continuity, another phase in an ongoing American experience of technology linked with territorial and/or eco- nomic expansion. Elements of the Old West survive in the gold rush mentality and lawlessn
The Wild, Wild Web: The Mythic American West and the Electronic Frontierdominant approach is now the rational agent approach
Constructivism and its risks in artificial intelligenceresearcher gives a talk about using RL to train a simulated robot hand to pick up a hammer and hammer in a nail. Initially, the reward was defined by how far the nail was pushed into the hole. Instead of picking up the hammer, the robot used its own limbs to punch the nail in. So, they added a reward term to encourage picking up the hammer, and retrained the policy. They got the policy to pick up the hammer…but then it threw the hammer at the nail instead of actually using it.
Deep Reinforcement Learning Doesn't Work Yet