Daily

Thursday, 17 September 2026

  1. 01Nvidia announces native GPU programming in Rustmust172 · 58c

    NVIDIA is compiling Rust kernels to PTX instead of wrapping someone else's C++. Two tracks: cuda-oxide is a custom rustc backend through Pliron and LLVM (pinned nightly, early alpha, DisjointSlice so classic aliasing fails at compile time), and cutile-rs is Tile IR on stable 1.89, already on crates.io and in Hugging Face Grout and mistral.rs. Neither is production-ready; I opened it because kernel-in-Rust with actual safety types is the thing the CUDA C++ people said would never ship.

    • GPU
    • Compilers
    • OSS
    WTF5BT7
    • claiirfunnyHN

      The launch is checked rather than trusted. Damn even Nvidia is putting out fully Claude-written articles.

  2. 02Training a 4B model to produce 81% faster query plans than Postgresmust372 · 76c

    Rohan Bansal took a 4B Qwen distill that couldn't even speak the planner harness, distilled Astra trajectories into it, then ran agentic GRPO against live Postgres — 2× H100 for vLLM, four containers on a desk for measurement. Best-of-15 on the 113 JOB queries is 1.81× geometric mean versus the default planner and 44.7% less summed latency, about $1,200 all-in. I opened it for the measurement rig: Linux page cache was lying to the reward, so he had to denoise Postgres before the model could learn.

    • AI
    • Agents
    • Systems
    WTF5BT7
    • malisperfunnyHN

      Funnily enough, you could replace "LLM query planner" with just "query planner" and this comment would still hold true

  3. 03Breaking the 1.58-bit Barrier for Ternary LLMs122 · 13c

    Georganas, Heinecke, and Dubey measured 29 ternary LLMs and found zeros eat up to 51.5% of the weights, so log₂(3)≈1.58 was always a bit of a fiction. BITCOS is a presence bitmap plus compacted signs at 2−z bits, beating five-trit packing on 26 of 29 models down to 1.485 bits/weight, with AVX-512 / AVX2 / Xe2 unpackers and up to 1.28× kernel and 1.27× GPU decode. Intel packing paper, not a new architecture — I glanced because the 1.58-bit slogan finally met an entropy measurement.

    • AI
    • GPU
    WTF3BT6
  4. 04Xiaomi Mimo 2.6 live post-training dashboard216 · 58c
  5. 05Backups Aren't Simple48 · 19c
  6. 06Small programming tricks377 · 179c
  7. 07The engineering behind the US Strategic Petroleum Reserve74 · 15c
  8. 08OpenSpec – A lightweight and configurable AI spec framework42 · 10c
  9. 09Reversing Factorio's RNGmust131 · 13c
  10. 10Performance Improvements in .NET 11138 · 27c
  11. 11Australia says it could follow Canada in forging deeper ties with EU84 · 31c
  12. 12AWS says it can't restore some data from mideast facilities struck by Iran198 · 169c
  13. 13Japan's book scene is moving from bookstores to libraries93 · 32c
  14. 14Reverse-engineered Jev-like model60 · 8c
  15. 15How good are frontier models at physics?must70 · 32c
  16. 16Show HN: An e-ink frame that hears birds and draws them as 1800s illustrations2059 · 236c
  17. 17Accurate Models of AMD Matrix Coresmust58 · 7c
  18. 18Mistral X Mozilla: Private, Multilingual AI Browsing525 · 185c
  19. 19Anecdotally, programmers dislike "reduce"80 · 143c
  20. 20Why Does the Universe Expand?19 · 6c
  21. 21Dream-RSI: Recursive Self-Improvement through Evolving Worlds179 · 49c
  22. 22Anatomy of a Texture71 · 12c
  23. 23HarnessTax: How Much Does the Harness Matter for Coding Agents?10 · 0c
  24. 24Training Text-to-Image Models 3.6× Faster29 · 4c
  25. 25William Buckland's Theology of Geology4 · 0c
  26. 26Destroy After Reading: photocopiers,cheap paper and DIY gave metal it's look11 · 2c
  27. 27Hackers Got Inside a Flock Camera455 · 210c
  28. 28The DeepMind Institute135 · 43c
  29. 29WalShadow: Sub-second Postgres replication to ClickHouse from physical WAL38 · 5c
  30. 30Kyber (YC W23) Is Hiring a Forward Deployed Engineer1 · 0c
  31. 31Vectorized and performance-portable Quicksort (2022)161 · 26c
  32. 32I replaced my brown-noise browser tab with a menu bar app14 · 5c
  33. 33Tell the speakers that you liked their talks269 · 74c
  34. 34Autism is genetic – and why this information matters now46 · 30c
  35. 35The Siberian Ice Maiden and the Scythian World45 · 4c
  36. 36Show HN: AttaLambda: a language where types and data are made of untyped lambdas35 · 2c
  37. 37Common Crawl Data Stored on a Hugging Face Bucket7 · 0c
  38. 38A warning about 'model welfare'194 · 527c
  39. 39Show HN: Restarted – a 2026 remake of the classic 2015 startup generator17 · 4c
  40. 40Douglas Adams and the exterminated Doctor Who adventure71 · 72c
  41. 41How big are factorials?80 · 38c
  42. 42The Google Play app review process now regularly takes longer than a week340 · 330c
  43. 43Claude Cowork and chat are now one Claude199 · 205c
  44. 44My website charged AI agents a penny per page. I watched Claude pay it8 · 1c
  45. 45Introducing System One Models and Jev1787 · 473c
  46. 46Paramount 'Leaks' Study Saying California Will Suffer If Its Merger Is Blocked7 · 1c
  47. 47Show HN: How Stale Is Your AI? Release age and training cutoff for 20 models69 · 44c
  48. 48Salesforce Global Outage261 · 166c
  49. 49Write Linters and Tools Before Code4 · 0c
  50. 50Doing Everyone Else's Job196 · 93c

Earlier days7

Archive