Open-sourced genetic algorithm optimizer hit top 2 in Open Kernel Challenge within 7 hours
morgymcg · x · 2026-11-18
Developer RohanArun open-sourced the genetic algorithm kernel optimizer that took him to top 2 in the Open Kernel Challenge in 7 hours, bundled in his MIT-licensed auto-research repo (68 stars).
- How it works: you write an adapter for your specific GPU, then a coordinator (human or agent) guides the search with intuition and supplies new tests.
- Key prerequisite: once the local GPU benchmark exactly matches the judges, it can run all day unattended, steadily improving kernels.
- The dashboard screenshot shows 5 boards (bf16proj, prefilllayer, attentiondecode, moew4a16, etc.) sharing one GPU; the control file has been optimized 23.49% cumulatively across 669 chunks.
- The author also recounts 2009-era CUDA optimization work at Mercury Computer Systems (often 1000x speedups) and an old trick using convolution on data to fit it into warps over time.
The repo is a native macOS auto-research harness for Codex and Claude Code; the first 100 contributors get free Agent37 sandboxes.
More from coding & agent
- Future codebases will be 80% AI-assisted, raising sloppier-code concerns — 996roma · 2026-11-18
- Two weeks of service_tier "priority" A/B tests: median +5.5% TTFT, basically noise — iambubu0523 · 2026-11-18
- Sentry founder jokes about MCP as HN thread declares it the standard for AI agent tooling — zeeg · 2026-11-18
- Self-hosted Qwen3.8-27B + pi LoRA matches Opus 5.5 on a real bug fix in ~93s — yb2698 · 2026-11-18
- Astra 6.2 hailed as best computer-use model after flawless MS Paint demo — gabrielchua · 2026-11-18
- Driveline team's model lineup: Opus 5.5 still king for coding, Carbon 4 for reviews — ivan_bezdomny · 2026-11-18