A Codex Auto-Research Loop Hit 232x on a Batched QR Kernel — 1,500 Submissions Over 14 Days, 12th of 183
Top of HN today at 424 points and 91 comments: a writeup of driving GPT-5.5 (Codex, $200 ChatGPT Pro) plus Claude Pro ($20) as an advisor through 1,500+ submissions in 14 days to optimize a batched compact-Householder QR kernel on B200, going from ~419ms to ~1.8ms — a 232x speedup and 12th of 183 on the leaderboard. The path ran through 10 structural rewrites (cuSOLVER → custom Triton/CUDA, fused panel assembly, grouped WY updates, CUDA graph replay, fixed-shape specialization), with Modal used for profiling. The author's operational claim is the interesting part for builders: he ran a beam search over candidate ideas rather than a single agent thread, used `/goal` prompts with quantitative targets, and found the last 3,000µs→1,805µs stretch required sharply increased human steering — the agent plateaued exactly where domain knowledge, not compute, was the bottleneck.
↳ Follow the thread