Rajan Agarwal
44 followers · 18 following · 5562 views
on the atlas — 42
- MAI-Thinking-1: Building a Hill-Climbing Machine2 savers
- There Are No New Ideas in AI… Only New Datasets12 savers
- A Geometric Calculator Inside a Neural Network2 savers
- Finite Difference Flow Optimization1 savers
- The Intelligence Curse28 savers
- Emotion concepts and their function in a large language model \ Anthropic8 savers
- A “diff” tool for AI: Finding behavioral differences in new models \ Anthropic2 savers
- SE3501 savers
- TurboQuant: Redefining AI efficiency with extreme compression8 savers
- AttnRes2 savers
- Speeding up RL with high-leverage samples | Applied Compute2 savers
- List of Patterns9 savers
- The First Fully General Computer Action Model | blog35 savers
- Towards self-driving codebases · Cursor6 savers
- Introducing GPT-5.3-Codex | OpenAI1 savers
- Dynamic context discovery · Cursor3 savers
- Scaling PostgreSQL to power 800 million ChatGPT users | OpenAI7 savers
- Building a web search engine from scratch in two months with 3 billion neural embeddings15 savers
- Understanding Muon2 savers
- Extending the Context of Pretrained LLMs by Dropping their Positional Embeddings1 savers
- Visualizing Flow Matching in Robotics1 savers
- The Unreasonable Effectiveness of Recurrent Neural Networks15 savers
- You will be OK — LessWrong13 savers
- RLHF Book by Nathan Lambert14 savers
- The Ultra-Scale Playbook - a Hugging Face Space by nanotron8 savers
- Lessons from 2025: Non-Linearity, The Games We Play, Side Questing2 savers
- On-Policy Distillation - Thinking Machines Lab29 savers
- Building Reliable AI Agents | Writings2 savers
- Cognition | Introducing SWE-grep and SWE-grep-mini: RL for Multi-Turn, Fast Context Retrieval4 savers
- NL.pdf7 savers
- GPU Glossary10 savers
- How Tech Hype is Ruining College Students in Tech37 savers
- [2304.03442] Generative Agents: Interactive Simulacra of Human Behavior19 savers
- The Scaling Hypothesis · Gwern.net16 savers
- On Oppenheimer and Progress14 savers
- New York’s Best (Fake) Steak House Opens Up - The New York Times12 savers
- Web Interface Guidelines9 savers
- Ushering in the Thermodynamic Future - Litepaper9 savers
- Real life - by Kasra - Bits of Wonder8 savers
- Video generation models as world simulators7 savers
- OpenAI Cookbook5 savers
- Improving Mathematical Reasoning with Process Supervision5 savers
highlights — 31
A “Chinese Communist Party Alignment” feature found in the Qwen3-8B and DeepSeek-R1-0528-Qwen3-8B models. This controls pro-government censorship and propaganda in these Chinese-developed models, and is absent in the American models we compared them against. An “American Exceptionalism” feature found in Meta’s Llama-3.1-8B-Instruct. It controls the model’s tendency to generate assertions of US superiority, a control absent in the Chinese model it was compared against. A “Copyright Refusal Mechanism” feature exclusive to OpenAI’s GPT-OSS-20B. It controls the model’s tendency to refuse to provid…
A “diff” tool for AI: Finding behavioral differences in new models \ AnthropicThis way, small, frequent movements are tokenized into finer bins and large, infrequent movements into coarser ones. We also train our FDM to predict the next click position alongside every mouse movement token, which helps produce accurate trajectories.
The First Fully General Computer Action Model | blogJust drop the model’s positional embeddings after pretraining and run a short recalibration.
Extending the Context of Pretrained LLMs by Dropping their Positional EmbeddingsThe issue here is that we don’t have access to the ground-truth vector field either. Instead, the solution is to work with individual trajectories. Both 𝑥 0 (noise) and 𝑥 1 (a cat image) are just vectors in high-dimensional space. Now the simplest way to get from 𝑥 0 to 𝑥 1 is by a straight line, or more precisely by travelling along 𝑥 1 − 𝑥 0 . We can represent the position of a sample at any timestep along this direction as:
Visualizing Flow Matching in RoboticsThis year, I did get offers from Big Tech companies I never expected I would get this early, including companies like Coinbase. A year ago, I would have been satisfied with that outcome as a finish line. Yet I chose Polymarket instead. In part, because I want to take my own beliefs seriously. I do not want to play a game just because it is the default game around me.
Lessons from 2025: Non-Linearity, The Games We Play, Side QuestingHowever, standard training and inference libraries have different numerics, which effectively turns the sampled data into off-policy data. This is amplified when using low-precision rollouts, a common optimization in RL frameworks.
Cognition | Introducing SWE-grep and SWE-grep-mini: RL for Multi-Turn, Fast Context RetrievalMoore’s law is starting to slow down [2]. The reason for this is rooted in fundamental physics: transistors are approaching the atomic scale where effects like thermal noise start to forbid rigid digital operation
Ushering in the Thermodynamic Future - LitepaperThis network takes raw video as input and outputs a latent representation that is compressed both temporally and spatially.
Video generation models as world simulatorsWhat if you were more often baffled by the fact that you’re still alive, if you began to ask of this moment, of every moment: what do you have to teach me?
Real life - by Kasra - Bits of WonderDuring historical technology cycles, overbuilding of infrastructure has often incinerated capital, while at the same time unleashing future innovation by bringing down the marginal cost of new product development.
AI’s $200B Question | Sequoia CapitalWhat if someone complains? “Come downstairs, and I’ll apologize.”
New York’s Best (Fake) Steak House Opens Up - The New York Times“I’m going to send a Google Drive link to the menu,” he told the team. “This is your Bible, your Quran.”
New York’s Best (Fake) Steak House Opens Up - The New York TimesThese retrieved memories are also used to form longer-term plans and create higher-level reflections, both of which are entered into the memory stream for future use
[2304.03442] Generative Agents: Interactive Simulacra of Human BehaviorThey are aware of other agents in their local area, and the generative agent architecture determines whether they walk by or engage in conversation.
[2304.03442] Generative Agents: Interactive Simulacra of Human BehaviorOne place where the movie toys with this notion (maybe too subtly!) is in its presentation of Oppenheimer’s self-construction as a singular figure, broken apart from all the communities—political, religious, etc—that might hold him. Unlike other Jewish characters/historical figures in the movie, Oppenheimer is shown as self-consciously worldly and secular—he speaks German, the language of Physics, rather than Yiddish, the language of his people.
On Oppenheimer and ProgressThis low-frequency vibration sends skyscrapers swaying, according to Denolle. “The recent earthquakes were deeper, so they had a higher frequency,” she said.
Earthquakes in Turkey and Syria: What we know — and don’t know — about earthquakes - Vox“We prefer to use peak ground acceleration,” she said. This is a metric that measures how the speed and direction of the ground changes and has proven the most useful for engineers.
Earthquakes in Turkey and Syria: What we know — and don’t know — about earthquakes - VoxHe says, “This is just a simple idea I had. Listen—what about a teaching machine. . . .”
What a Sixty-Five-Year-Old Book Teaches Us About A.I. | The New YorkerThis kitchen is my studio—my laboratory, just like your laboratory, Professor. Would you want an automatic laboratory?”
What a Sixty-Five-Year-Old Book Teaches Us About A.I. | The New YorkerWhen we choose to give energy to things that don’t serve our goals we are trading against our own creative well being.
A space for ourselves | Chase McCoyOur technologies have their own agenda and definition of success that is frequently at odds with our own.
A space for ourselves | Chase McCoyThere are subtle sounds of presence, stories, and selves that the internet might be better suited for. In the way that it is now so all-enveloping that digital networking and the internet is ambience in itself, it still is a larger system, or maybe even a new Environment that we operate in.
naming ambient art - Chia's BlogI had forgotten the Universe itself was technology, and that all real had once been mythology, and that to truly develop something sustaining and meaningful necessitated a sort of synergy and desire to break apart the operative principles of what was already here. And that there was so much more here than what I immediately sensed. So I began listening.
naming ambient art - Chia's BlogAs an audio addict, I end up searching for more songs by the same artist, more songs of the same genre, songs featuring the artist who made that one song, songs by artists that the original song’s artist collaborated with — anything to get my fix.
Why I’m taking a 30 day break from musicEmpirically, a few hundred tokens seem to be enough to get a reasonable signal that yes, this text came from GPT. In principle, you could even take a long text and isolate which parts probably came from GPT and which parts probably didn’t
OpenAI Has the Key To Identify ChatGPT's WritingMaybe we could find the AI’s styleme (like a fingerprint hidden in language) not simply to distinguish ChatGPT from a human, but to distinguish its style from all others.”
OpenAI Has the Key To Identify ChatGPT's Writingwe’re going to have to trust search algorithms to work well, because they’ll be the de-facto arbiters of truth.
Generative AI will destroy our concept of trust on the internet.Whatever can be AI generated, will be.
Generative AI will destroy our concept of trust on the internet.If you don’t do this, your life will be treated as just another disposable product..”
How Tech Hype is Ruining College Students in TechA builder chooses what she wants to build, and she holds herself accountable for the work of building it. She is not beholden to any inner or outer drill sergeant; only to her own independent, carefully formed judgment of what is worth building and how best to go about it.
How Tech Hype is Ruining College Students in TechBIPV products are regulated both as building materials and electrical materials, so they have to meet two sets of requirements.
Solar windows, shingles and cladding? The building itself is now the solar panel | CBC News