Riya Tyagi
11 followers · 12 following · 1228 views
on the atlas · 37
- People · Patrick Collison5 savers
- An Alien Mind | OpenAI28 savers
- Reiner Pope – The math behind how LLMs are trained and served6 savers
- How to Land a Frontier Lab Job20 savers
- clustering_algorithms_hartigan.pdf1 savers
- A short survey on almost orthogonal vectors in a few specific large dimensions1 savers
- How to Buy Cheap Claude Tokens in China - by Zilan Qian8 savers
- On Casting A Line - by Logan Strohl - Logan’s Substack1 savers
- I Stumbled Across My Boyfriend's ChatGPT and It Ended Our Relationship7 savers
- QK-Clip：让Muon在Scaleup之路上更进一步 - 科学空间|Scientific Spaces2 savers
- How to Make $6,000,000 by the Time You’re Born - Laker Newhouse1 savers
- Google Docs2 savers
- Many AI Debates Are Way Too Philosophical and Theoretical1 savers
- Lossy self-improvement - by Nathan Lambert3 savers
- Alain-Badiou-Nicolas-Truong-Translated-by-Peter-Bush-In-Praise-of-Love-2012-Serpents-Tail.pdf1 savers
- Pick an audience that is into the illegible you of today, not your past achievements21 savers
- What's going on here, with this human? - Graham Duncan Blog49 savers
- Sometimes the reason you can’t find people you resonate with is because you misread the ones you meet32 savers
- Everything that turned out well in my life followed the same design process65 savers
- Litany of Tarski — LessWrong2 savers
- seven-s-virtues2 savers
- Good conversations have lots of doorknobs36 savers
- Looking for Alice - by Henrik Karlsson - Escaping Flatland94 savers
- Interpretability Infrastructure at Frontier Scale: Harvesting Activations from a Trillion-Parameter Model2 savers
- The Persona Selection Model: Why AI Assistants might Behave like Humans36 savers
- Child’s Play, by Sam Kriss75 savers
- Anthropic/values-in-the-wild · Datasets at Hugging Face3 savers
- Paper: Prompt Optimization Makes Misalignment Legible — LessWrong1 savers
- Intentionally Designing the Future of AI7 savers
- What Would Non-Linear Features Actually Look Like? | Liv Gorton7 savers
- Join the Zed Team1 savers
- Keratosis Pilaris after waxing - Google Search1 savers
- Curius / Onboarding2704 savers
- Milan Cvitkovic52 savers
- Post 38: On Slack - Having room to be excited — Neel Nanda23 savers
- When Models Manipulate Manifolds: The Geometry of a Counting Task9 savers
- Thoughts on the conservative assumptions in AI control3 savers
highlights · 37
Maybe this is ordinary. Maybe all love looks less romantic under fluorescent light.
I Stumbled Across My Boyfriend's ChatGPT and It Ended Our RelationshipTruthfully, we were probably just not right for each other and it simply bled out into this situation
I Stumbled Across My Boyfriend's ChatGPT and It Ended Our RelationshipSo maybe this event is simply, as *Tim Kreider wrote, “the price of being known”: that if you want the rewards of intimacy, sooner or later, you risk colliding with the version of yourself that exists outside your own narration and in somebody else’s unvarnished language … and finding the result less poetic than you hoped.
I Stumbled Across My Boyfriend's ChatGPT and It Ended Our RelationshipThe problem was that unfortunately I saw them in their raw form, stripped of tact and timing and love’s softening language.
I Stumbled Across My Boyfriend's ChatGPT and It Ended Our RelationshipPeople who will encourage and enable the searching, creative, illegible you of today, rather than the polished you of your past achievements.
Pick an audience that is into the illegible you of today, not your past achievementsI’ll bet that as few as 20% of us are in the seat that best optimizes our talents and skills at any given time—the
What's going on here, with this human? - Graham Duncan BlogThe more I’ve done it, the more I realize that what most people think of as the hard parts of hiring—asking just the right question that catches the candidate off guard, defining the role correctly, assessing the person’s skills—are less important than a more basic task: how do you see someone, including yourself, clearly?
What's going on here, with this human? - Graham Duncan BlogIf someone seems boring to you, or a bad fit, it might be that you don’t know how to prompt them, that you haven’t seen them react to the context that brings out their full being.
Sometimes the reason you can’t find people you resonate with is because you misread the ones you meetthe way they process the world, their values, and virtues
Sometimes the reason you can’t find people you resonate with is because you misread the ones you meetBut as obvious as this seems to me now, I failed to see it for 15 years.
Sometimes the reason you can’t find people you resonate with is because you misread the ones you meetThey might be fun, they might be useful, but even if they aren’t, you like them anyway.
Sometimes the reason you can’t find people you resonate with is because you misread the ones you meetPleasure will come and go, as will utility.
Sometimes the reason you can’t find people you resonate with is because you misread the ones you meetBeing understood like this, to a level where someone can get you exactly the thing that you most wanted without needing to communicate about it, is very special.
seven-s-virtues1) someone that you will find fascinating to talk to after you’ve talked for 20,000 hours
Looking for Alice - by Henrik Karlsson - Escaping FlatlandConversational affordances are things like digressions and confessions and bold claims that beg for a rejoinder
Good conversations have lots of doorknobsDO NOT BE A SOCIAL SLOB USE CONVERSATIONAL DOORKNOBS
Good conversations have lots of doorknobseven the faintest fear of rejection on either side can prevent conversations from taking off
Good conversations have lots of doorknobsWe overestimate the awkwardness of deep talk and so we stick to the boring, affordance-less shallows.
Good conversations have lots of doorknobsWe think people want to hear about exciting stuff we did without them (“I went to Budapest!”) when they actually are happier talking about mundane stuff we did together
Good conversations have lots of doorknobsut whether we offer and accept affordances.
Good conversations have lots of doorknobsAnd the way to do that, I think, is by adding a bunch of doorknobs.
Good conversations have lots of doorknobsAnd asking your partner question after question and resenting them when they don’t return the favor isn’t generosity; it’s social entrapment
Good conversations have lots of doorknobsOn day one of Improv 101 they’ll tell you not to ask questions in a scene because it puts undue pressure on your partner.
Good conversations have lots of doorknobsIt’s easy to remember how lonely it feels when a taker refuses to cede the spotlight to you, but easy to forget how lovely it feels when you don’t want the spotlight and a taker lets you recline on the mezzanine while they fill the stage.
Good conversations have lots of doorknobsYou are born with this weird interiority that no one else can see.
Looking for Alice - by Henrik Karlsson - Escaping FlatlandThe forward pass is kept as lightweight as possible. We first copy each activation tensor into a GPU-side staging buffer so the capture pipeline can copy it at its own pace, without blocking inference. This device-to-device copy operation is fast, as the activation tensor never leaves GPU memory, and the buffer comes from a preallocated pool of GPU memory, which avoids latency from on-demand allocations. Even though synchronously transferring full activations to the CPU is a nonstarter, we still need to retain references to those staged tensors in CPU memory. We do this in the lightest way pos…
Interpretability Infrastructure at Frontier Scale: Harvesting Activations from a Trillion-Parameter Modelthat provides an efficient, near-exact approxi- mation of the optimal reward-KL divergence policy at inference time.
Inference-Time Reward Hacking in Large Language ModelsThis includes passing up on opportunities to improve its fundamental understanding in the short-term, instead doing pharmaceutical R&D in whatever way would actually be more efficient.
Why AI alignment could be hard with modern deep learningusing a linear combination of global, polynomial basis functions
Ch. 10 - Trajectory OptimizationThe good news is that if you add a system prompt about the issue and a FileRead reminder to check if code is malicious, then this fixes the covert issue, which again suggests 4.5 is otherwise overly reliant on superficial signs, which in practice might be fine if 4.5 is always run with the mitigations:
Claude Sonnet 4.5: System Card and Alignment — LessWrongManipulation feels like the next big thing that will make the leap from robotic research to practice.
Robotic ManipulationIn practice, no state-of-the-art robotic manipulation system to date (that I know of) uses rigorous control theory to design even the low-level feedback that determines when a robot makes and breaks contact with the objects it is manipulating
Robotic Manipulationeverything we mention knows its own definition
1Lab - 1Lab(1) Fused re- shape and block write.
Efficient Memory Management for Large Language Model Serving with PagedAttentionAnd in general, being willing to spend a few minutes to fix small annoyances
Milan CvitkovicThe goal is to be the voice of care, not the voice of judgment.
boundaries (of others) - by Claire Wang - cold brew blogead everything that he could find about business. The subject that interests him, he’s read newspapers, biographies, trade press. He went over to his grandfather who was a grocer and he read the progressive
Curius / Onboarding