Arya G

on the atlas · 54

• The Hitchhiker’s Guide to Training on Compute Islands | Core Automation7 savers
• Yusuke Miyashita on X: "Understanding MuonH(Hyperball) through Weight Norm Geometry" / X1 savers
• Limite 1B - Violetto Technical Report1 savers
• Hyperball, effective lr, and peak plus attenuation shape | Jiaxuan Zou1 savers
• Frontier Pre-training at Microsoft AI | Microsoft AI5 savers
• [2510.26692] Kimi Linear: An Expressive, Efficient Attention Architecture1 savers
• Linear Attention Fundamentals | Hailey Schoelkopf5 savers
• Introducing Beam: Reflection’s 501B open-weight model — Reflection4 savers
• [2610.01153] Looping Beyond Twice: A Scalable Recipe for Looped Mixture-of-Experts1 savers
• Training NanoGPT in 39.9 Seconds — Hyperstition4 savers
• [2603.15031] Attention Residuals1 savers
• [2609.32534] DepthBench: Measuring How Residual Connections Enable More Computational Depth1 savers
• [2609.30063] Self-Play Pretraining with Zero Data2 savers
• Computational depth is all you need – Q2 savers
• [2507.20534] Kimi K2: Open Agentic Intelligence1 savers
• What is test-time compute and how to scale it?1 savers
• Reading a Pretraining Run — Chunyuan Deng2 savers
• An Alien Mind | OpenAI28 savers
• Dario Amodei — We Must Pace the Frontier37 savers
• Introducing System One Models and Jev - TypeSafe AI Blog21 savers
• Inside vLLM: Anatomy of a High-Throughput LLM Inference System - Aleksa Gordić11 savers
• FLT: Anthropic has beaten me to it | Xena4 savers
• Formalizing Fermat's Last Theorem \ Anthropic5 savers
• How Claude Performs on Robotics Tasks \ Anthropic6 savers
• How our data shaped neural architecture discovery, and how automation can reshape the future | Core Automation6 savers
• Erdős Problems4 savers
• Announcing FrontierMath Erdős | Epoch AI1 savers
• [2607.16051] Loop the Loopies!1 savers
• Harness Engineering for Self-Improvement | Lil'Log19 savers
• DeepSeek-V3: A Large-Scale MoE Pretraining Benchmark for MLPerf Training v6.0 - MLCommons1 savers
• Frontier-scale RL with Kimi K3 | Applied Compute1 savers
• Defeating Nondeterminism in LLM Inference - Thinking Machines Lab40 savers
• Non-determinism in GPT-4 is caused by Sparse MoE - 152334H5 savers
• The 4-bitter Lesson | humans&5 savers
• [2512.24880] mHC: Manifold-Constrained Hyper-Connections1 savers
• [Hero Run] 535B-A23B on 18T tokens · Issue #8435 · marin-community/marin1 savers
• Mixture-of-Kittens: our open-source MoE megakernel for NVL72s · Cursor3 savers
• Better MoE model inference with warp decode · Cursor7 savers
• Mixture of Experts Quantile Balancing: Validated at 32B-A5B (1e22 FLOPs) Scale | Open Athena1 savers
• Open Athena | Scaling Laws That Extrapolate 300× Past the Fit2 savers
• [2604.20920] Simplified Sparse Attention via Gist Tokens1 savers
• [2108.12409] Train Short, Test Long: Attention with Linear Biases Enables Input Length Extrapolation2 savers
• Four LLM loss functions → four flavors of LLM misalignment — LessWrong4 savers
• The Curious Case of the bos_token — LessWrong1 savers
• How GPT-5, Claude, and Gemini are actually trained and served – Reiner Pope - YouTube2 savers
• [2509.14786] Pre-training under infinite compute2 savers
• [2608.09867] Stealing Reasoning Traces from Proprietary LLM APIs3 savers
• llms-cant-jump.pdf1 savers
• Hot Take: LLM can 'jump' | Yong Zheng-Xin1 savers
• A retrospective of AI alignment17 savers
• The Scaling Hypothesis · Gwern.net16 savers
• Scaling Laws, Carefully | Lil'Log15 savers
• Your Agents Are Not Time Aware — LessWrong3 savers
• A short note on some aspects of long context attention | nor's blog2 savers