Vjeux's AI agents hit 2.6x speedups on Open Kernel Challenge kernels, then stalled
Vjeux · x · 2026-10-28
Frontend engineer Vjeux had his AI agents compete in the Open Kernel Challenge (Muse Spark kernel track) — and stayed competitive despite knowing nothing about kernels. He released five single-file CUDA/Python kernels (Apache-2.0, loadinline native CUDA, CUDA-graph capturable), judged against SGLang baselines on H200s:
- museattentiondecode: 2.590x (#20); his best file (2.581x) was withdrawn before scoring and can't be resubmitted due to hash dedup
- musemoew4a16: 2.325x (#38)
- musedecodelayer: 2.086x (#46)
- museprefilllayer: 1.798x (#18)
- musebf16proj: 1.501x (#138)
No category wins; 8-15 pct behind each category leader. After progress stalled and costs mounted, he stopped and open-sourced the final code.
More from coding & agent
- Superwhisper taps Grok Bot, showing how businesses can ship products as conversational templates — XFreeze · 2026-10-28
- Sentient OS (YC F26) launches: on-device AI that learns your life overnight with a local LLM — testingcatalog · 2026-10-28
- Supermemory reverse-engineers how Claude Code's memory actually works — julianweisser · 2026-10-28
- GitHub's Spec Kit pushes spec-driven development: 5 commands to plan before coding agents build — Arindam_1729 · 2026-10-28
- How LangChain built its design agent: taste comes from context, not the model — LangChain · 2026-10-28
- Dev predicts in 5 years we'll laugh at today's juggling of terminal and agent harness tabs — DavidWells · 2026-10-28