Open-source auto-research agent hits Top 2 on GPU kernel challenge in 7 hours
sguada · x · 2026-10-27
RohanArun, a CUDA veteran with GPGPU patents, open-sourced an Auto-research agent harness that got him to Top 2 on the Muse Open Kernel Challenge leaderboard in 7 hours.
Key points:
- A native macOS auto-research harness built on Codex and Claude Code — pick an open research project and press Play
- No GPU needed to compete: 30 submissions per day per leaderboard run on the organizers' GPUs; goal is to beat the SGLang baseline, then keep iterating on model-suggested ideas
- Winning is now hard since top spots require many real GPU experiments
- The author shares past tricks from his 2009-era CUDA optimization work, like using convolutions to fit data into warps for 1000x speedups
MIT-licensed on GitHub; the approach can also be applied to your own private inference stack for kernel optimization.
More from coding & agent
- Testing Grok Bot: agent demands full-device GitHub PAT despite existing repo access — vincent_koc · 2026-10-27
- OpenCode v2.2.0 adds desktop plugins and cache-friendly keyed session instructions — github-actions[bot] · 2026-10-27
- Fresh chat per task: 7 context engineering tips to avoid context rot — every · 2026-10-27
- Redditor open-sources a simple workflow for infinite lipsynced talking head videos with MiniMax H3 — martinerous · 2026-10-27
- System prompts aren't a security boundary: Reddit debates how much access AI agents should get — Putrid_Ad734 · 2026-10-27
- AI told to install Minecraft mods starts building its own C++ game engine — Promptmethus · 2026-10-27