anka hu
20 followers · 15 following · 1170 views
on the atlas — 110
- Introducing System One Models and Jev - TypeSafe AI Blog16 savers
- Optimism Shapes Reality - by Alexandr Wang9 savers
- CLIP: Connecting text and images | OpenAI3 savers
- The Yale Review | Daniel Lefferts: “Terms and Conditions”5 savers
- The Persona Selection Model: Why AI Assistants might Behave like Humans34 savers
- The persona selection model \ Anthropic4 savers
- I cannot see beyond it - by shenai - letters to home3 savers
- AI 2040: Plan A22 savers
- Can You Just Do Things?—Asterisk16 savers
- Capsule Contents - America2501 savers
- Storing Lists of Values with Vectors - The Rust Programming Language1 savers
- Prompt guidance | OpenAI API1 savers
- Paths for Referring to an Item in the Module Tree - The Rust Programming Language1 savers
- microgpt6 savers
- The Old World Is Dying: Advice for 2026 graduates42 savers
- What it feels like to work with Mythos - by Ethan Mollick5 savers
- Kernel | Transubstantiation2 savers
- I Taught My Dog to Vibe Code Games | Caleb Leak5 savers
- Should I Refrigerate Apples? And More Fruits and Vegetables Storage Questions, Answered - NYT Cooking3 savers
- Memory access is O(N^[1/3])3 savers
- Musicmap | The Genealogy and History of Popular Music Genres1 savers
- Strudel REPL7 savers
- MIT 6.S976 and 18.S996 Cryptography and Machine Learning (Spring 2026)2 savers
- LoRA Without Regret - Thinking Machines Lab37 savers
- On idea-driven ideas6 savers
- moral-fn.pdf10 savers
- cdn.nakamotoinstitute.org/docs/cypherpunk-manifesto.txt1 savers
- Overview | Self Docs1 savers
- Optical Corrections in Architecture and Typography4 savers
- Japan's anti-immigration backlash4 savers
- Computer Scientists Figure Out How To Prove Lies | Quanta Magazine2 savers
- Will truckers be automated? (from the comments) - Marginal REVOLUTION1 savers
- Wigglypaint by Internet Janitor1 savers
- The Moat of Low Status12 savers
- Myriad (Tulips) — Anna Ridler1 savers
- 32 notes on AI & writing - by Jasmine Sun - @jasmine5 savers
- Agnes Callard’s Marriage of the Minds | The New Yorker1 savers
- Wide & Deep Learning: Better Together with TensorFlow1 savers
- Transformers4Rec: Bridging the Gap between NLP and Sequential / Session-Based Recommendation1 savers
- Deep dive: the proof-of-work Hyli Chat1 savers
- 9 Transformers – 6.390 - Intro to Machine Learning1 savers
- In Defense Of The Reality Of Good Taste - by Ozy Brennan2 savers
- never let me go - by Kai - golden blue2 savers
- Curius / Onboarding2621 savers
- Looking for Alice - by Henrik Karlsson - Escaping Flatland90 savers
- Staring into the abyss as a core life skill85 savers
- Theory of Change (Aaron Swartz's Raw Thought)83 savers
- The Bitter Lesson78 savers
- the friendship theory of everything69 savers
- How to Do Great Work67 savers
- Cultivating a state of mind where new ideas are born65 savers
- Here's to the fools who dream54 savers
- Things You Learn Dating Cate Hall - by Sasha Chapin52 savers
- Fast · Patrick Collison48 savers
- Conviction is the Scarcest Resource: Essays About Gaining Real Conviction45 savers
- no good alone - by rayne fisher-quann - internet princess45 savers
- Relationships are coevolutionary loops - by Henrik Karlsson39 savers
- IEEE Xplore Full-Text PDF:36 savers
- Good conversations have lots of doorknobs34 savers
- look what the cat brought in - by Anson Yu32 savers
- How-To-Succeed-At-MrBeast-Production.pdf - Google Drive32 savers
- Conviction is the Scarcest Resource: Dialogues About Conviction30 savers
- 50 things I know - by Sasha Chapin - Sasha's 'Newsletter'29 savers
- how to avoid half-heartedness - by Ava - bookbear express28 savers
- Guide to Career Planning, part 1: Opportunity28 savers
- How I Attained Persistent Self-Love, or, I Demand Deep Okayness For Everyone23 savers
- Interaction Models: A Scalable Approach to Human-AI Collaboration - Thinking Machines Lab23 savers
- The unexpected poetry of PhD acknowledgements | ANU College of Science22 savers
- Life is a Picture, But You Live in a Pixel — Wait But Why19 savers
- On caring19 savers
- Kernel | all the better to see you with19 savers
- 🌻 tryhard - by Jasmine Sun - half-baked futures18 savers
- I want my dad to know who I am before he dies18 savers
- How to Pick Your Life Partner - Part 2 — Wait But Why18 savers
- Some Painful Questions We Ask Ourselves - by Sasha Chapin17 savers
- The Load-Bearing Relationship17 savers
- on being selective17 savers
- The Waluigi Effect (mega-post) - LessWrong16 savers
- Building a web search engine from scratch in two months with 3 billion neural embeddings15 savers
- The Paris Review - The Crane Wife - The Paris Review14 savers
- bookbear express | Ava | Substack14 savers
- Developing ethical, social, and cognitive competence | Vividness13 savers
- What the humans like is responsiveness - by Sasha Chapin13 savers
- the zone of genius - by Isabel - Mind Mine13 savers
- The Lesson of Grace in Teaching, by Francis Su12 savers
- About these notes12 savers
- Things You Learn Dating Cate Hall - by Sasha Chapin12 savers
- Is My Toddler a Stochastic Parrot? | The New Yorker10 savers
- My Beautiful Friend | The Point Magazine8 savers
- undoing the pain8 savers
- the happy ending is you – A Slice of My Mind8 savers
- The Rust Programming Language7 savers
- How to Beat Procrastination — Wait But Why7 savers
- Exploring Elliptic Curve Pairings | by Vitalik Buterin | Medium7 savers
- How I got $25,000 in debt · Life at the Margin · Natecation · Natecation7 savers
- Blog - Private Cloud Compute: A new frontier for AI privacy in the cloud - Apple Security Research6 savers
- What is Ownership? - The Rust Programming Language5 savers
- The Slice Type - The Rust Programming Language5 savers
- .:: Phrack Magazine ::.5 savers
- Combining Machine Learning and Homomorphic Encryption in the Apple Ecosystem - Apple Machine Learning Research4 savers
highlights — 387
OKC Thunder Championship Pin On June 22, 2025, the Oklahoma City Thunder basketball team won the National Basketball Association Championship by beating the Indiana Pacers in a seven-game series.
Capsule Contents - America250Avoid unnecessary absolute rules. Older prompts often use strict instructions like ALWAYS, NEVER, must, and only to control model behavior. Use those words for true invariants, such as safety rules, required output fields, or actions that should never happen. For judgment calls, such as when to search, ask for clarification, use a tool, or keep iterating, prefer decision rules instead.
Prompt guidance | OpenAI APIWith Fable the spell has gotten powerful enough that I am no longer sure I am the wizard. I am closer to a patron. I describe what I want, I pay for it, and I judge the result. The conjuring happens somewhere I cannot watch, in hundreds of small choices I never get a vote on. The work has shifted from process to outcome. I no longer steer; I commission.
What it feels like to work with Mythos - by Ethan MollickWe need to erect a much expanded commons on the Internet. We need to realize popular services in a secure, distributed, and decentralized way, powered by free software and free/open hardware. We need to build systems beyond the reach of super-sized companies and spy agencies. Such services must be based on strong cryptography. Emphasizing that prerequisite, we need to expand our cryptographic commons . Dreams for such a commons go back to the cypherpunks, who built remailers, for example, as a communitarian service to enable secure communications. More recently, Feigenbaum and Koenig articulat…
moral-fn.pdfso packed with clip-art you can hardly find the content.
moral-fn.pdfsee at a different set of problems to work on than if we see things in the other. Whimsical adversaries engender a chimerical field. 184 As a graduate student, I wanted our field to feel fantastical. I wanted a discipline full of space aliens and communicating millionaires. Not only was it fun, but it stroked my ego, effectively embodying the sentiment: I am a scientist too smart to have to deal with small-minded concerns. At this point, I think we would do well to put ourselves in the mindset of a real adversary, not a notional one: the well-funded intelligence agency, the profit- obsessed mu…
moral-fn.pdfTo achieve this aim, computing the hash function shouldn’t just take lots of time, but lots of (sequentially accessed) memory. This insightful idea comes from Abadi, Burrows, Manasse, and Wobber, who wanted to make sure that, for a variety of settings, computing an intentionally-slow hash function on a high-end system would take roughly as long as computing it on a low-end system
moral-fn.pdfSo you grab the p bits at those locations and hash them, along with R , to get a derived key K :
moral-fn.pdfWhen a party wants to retrieve his i th message, he’ll interact with the same server, which gives him a string computed from the database contents. The value permits the receiver to recover the intended messag
moral-fn.pdfIf we behave morally, it is not because of rational analyses, but an instinctual preference for liberty, empathy, or companionship. 129 As Schneier points out, animals don’t like to be surveilled because it makes them feel like prey, while it makes the surveillor feel like—and act like—a predator. 1
moral-fn.pdff we behave morally, it is not because of rational analyses, but an instinctual preference for liberty, empathy, or companionship. 129 As Schneier points out, animals don’t like to be surveilled because it makes them feel like prey, while it makes the surveillor feel
moral-fn.pdfAs civil-rights attorney Frank Donner and the Church Commission reports thoroughly document, domestic surveillance under U.S. FBI director J. Edgar Hoover served as a mechanism to protect the status quo and neutralize change movements. 116 Very little of the FBI’s surveillance-related efforts were directed at law-enforcement: as the activities surveilled were rarely illegal, unwelcome behavior would result in sabotage, threats, blackmail, and inappropriate prosecutions, instead. For example, leveraging audio surveillance tapes, the FBI’s attempted to get Dr. Martin Luther King, Jr., to kill hi…
moral-fn.pdfThe law-enforcement narrative is wrong to position privacy as an individual good when it is, just as much, a social good. It is equally wrong to regard privacy and security as conflicting values, as privacy enhances security as often as it rubs against it.
moral-fn.pdfThe law-enforcement narrative is wrong to position privacy as an individual good when it is, just as much, a social good. It is equally wrong to regard privacy and security as conflicting values, as privacy enhances security as often as it rubs against it.
moral-fn.pdfBob’s secret key is no longer self-selected. It is issued by a trusted authority. That authority knows everyone’s secret key in the system. IBE embeds key escrow— indeed a form of key escrow where a single entity implicitly holds all secret keys—even ones that haven’t yet been issued. And even if you do trust the key- generating authority, a state-level adversary now has an extremely attractive locus to subpoena or subvert
moral-fn.pdfTo work, cryptographic primitives must be embedded into systems, and those systems can realize arrangements of power that don’t trivially flow from the nature of the tool.
moral-fn.pdfWhile the defense repeatedly proffered that the accused were simply following orders, this view was almost universally rejected : following orders did not efface legal or moral culpability.
moral-fn.pdfWe appeal, as human beings, to human beings: Remember your humanity, and forget the rest.
moral-fn.pdfhow could it not be a colossal failure of our field when ordinary people lack even a modicum of communication privacy when interacting electronically?
moral-fn.pdfWriting purely for the AI has no stakes. It’s text without audience, motive, purpose.
32 notes on AI & writing - by Jasmine Sun - @jasmineYet neuroscience suggests that memory and imagination are one and the same. As Demis Hassabis’s PhD thesis found, amnesiacs make shoddy novelists.
32 notes on AI & writing - by Jasmine Sun - @jasmineThe way that I think about it is: there’s no other time when you could understand this thing. Devastating problems in your life can also be interesting, and they can interest you as they’re happening to you and as they’re causing you intense pain.
Agnes Callard’s Marriage of the Minds | The New YorkerIt’s just that there’s some way in which philosophy could stand up to the task of making you able to deal with death when it comes.” “The corresponding claim,” Agnes said, “would be that somehow the project of marriage would make you capable of being alone.”
Agnes Callard’s Marriage of the Minds | The New YorkerShe added that, when she and Arnold fought, they could rely on Ben to provide an objective perspective.
Agnes Callard’s Marriage of the Minds | The New Yorker“It’s a very narrative, novelistic approach to my life, and the only area of my life that I see in such a progressive way is the pursuit of knowledge.” The proof of success or failure is her insights, she said, not the plot of her life.
Agnes Callard’s Marriage of the Minds | The New Yorkerfor me loneliness is almost like an internal problem. How can I manage to find reasons to tell the truth? Or how can I make contact with the idea of being honest?”
Agnes Callard’s Marriage of the Minds | The New YorkerThe ineffable wisdom they wrote of—inaccessible to others, because it was so mysterious and private—sounded to Socrates a lot like ignorance, she said. The idea that a marriage should hold space for each person’s incommunicable core, she believed, “comes from this pessimism where it’s, like, Look, at the end of the day we know we can’t really help one another, so the best thing we can do is not interfere too much.”
Agnes Callard’s Marriage of the Minds | The New Yorker“I’ll be, like, ‘Why can’t we get back to that?’ ” Agnes said. “And Arnold will be, like, ‘That was never there.’ He is offended by my attempt to go back in time. And I feel like he is taking away the foundation of our relationship and telling me that our lives are built on a lie.”
Agnes Callard’s Marriage of the Minds | The New Yorkerprevent positions from attending to subsequent positions.
1706.03762which performs multi-head attention over the output of the encoder stack
1706.03762The cross-feature transformation in the wide model component can memorize all those sparse, specific rules, while the deep model component can generalize to similar items via embeddings.
Wide & Deep Learning: Better Together with TensorFlowhow the co-occurrence of a query-item feature pair correlates with the target label (whether or not an item is consumed)
Wide & Deep Learning: Better Together with TensorFlowpreference-aware mask based on self-attention for
Transformers4Rec: Bridging the Gap between NLP and Sequential / Session-Based RecommendationInstead, they discretize the elapsed time between interactions at the log scale, and represent it as a categorical feature embedding.
Transformers4Rec: Bridging the Gap between NLP and Sequential / Session-Based RecommendationWhile BERT used MLM as a pre-training phase to learn words representation vectors and then fine-tuned the pre- trained model for downstream tasks evaluation, BERT4Rec used MLM as an end-to-end task for training and evaluation.
Transformers4Rec: Bridging the Gap between NLP and Sequential / Session-Based Recommendationecause this approach leaks future information on training, they ensured that during inference only the last item of the sequence is masked, making it compatible with the next-click prediction task.
Transformers4Rec: Bridging the Gap between NLP and Sequential / Session-Based Recommendationutilizing the self-attention mechanism to infer the item-item relationship from the user’s historical interactions and estimate weights of each item in the user’s trajectories.
Transformers4Rec: Bridging the Gap between NLP and Sequential / Session-Based RecommendationThe evaluation of sequential recommendation and session-based recommendation is performed using traditional Top-N ranking met- rics such as NDCG@N, Recall@N, Precision@N, MAP@N,
Transformers4Rec: Bridging the Gap between NLP and Sequential / Session-Based Recommendationhe set of items in recommender systems is often very large and fast training of a scalable model is important
Transformers4Rec: Bridging the Gap between NLP and Sequential / Session-Based Recommendationto learn item, user or context embeddings by taking their co-occurrence into account within the users’ history of interactions.
Transformers4Rec: Bridging the Gap between NLP and Sequential / Session-Based Recommendationwhich explicitly model sequences of user interactions to better infer preference or context changes over time. In a number of these settings only the most recent interactions are available, and in all domains for fresh or anonymous users only the interactions from the current user session are available.
Transformers4Rec: Bridging the Gap between NLP and Sequential / Session-Based Recommendationin next-click prediction for user sessions, where sequence lengths are much shorter than those commonly found in NLP,
Transformers4Rec: Bridging the Gap between NLP and Sequential / Session-Based RecommendationSometimes I show up at dinner thinking, what are we going to talk about? But we always have something to talk about. I am on your side during the good times and the bad. The people you hate are the people I hate. I love when you send me crazy emails and crazy texts and when you call me at weird times. I would never let anyone say a single bad thing about you in front of me. Only I am allowed to complain about you. And our other friends who love you, because you have to love someone before you’re allowed to hate them.
always on your side - by Ava - bookbear expressThe computational complexity of our preprocessing phase is dominated by the public-key operations, we need O ( n 2 /s ) operations per secure multiplication
535.pdfwork done by each player is only a small constant factor larger than what one would need to compute the circuit in the clear.
535.pdfthey are allowed to handshake on arbitrary orders, not just ones that have been registered in your wallet.
Super Relayers | RenegadeThis allows the relayer to view, match, and settle your trades, but does not allow the relayer to place new trades or cancel old ones. To update orders, you must mutate the wallet itself with your root keypair.
Super Relayers | RenegadeVALID MATCH MPC has been computed, either party can submit it to the smart contract, thereby actually swapping the tokens. Instead of just running matching engine execution directly, collaborative proving gives both parties assurance that matching (i.e., determining what tokens are swapped) is atomic with settlement (i.e., actually swapping the tokens).
The MPC-ZKP Architecture | Renegadethe relayers to collaboratively prove a particular NP statement, VALID MATCH MPC. This statement essentially claims that given the publicly-known commitments to order information and a public commitment to a matches tuple, both traders do indeed know valid input orders.
The MPC-ZKP Architecture | RenegadeIf either party could learn the MPC output and hangup the connection before actually swapping tokens, then the protocol would leak order information.
The MPC-ZKP Architecture | Renegade