How I Used Codex to Build a 232x Faster QR Kernel

Auto-research with codex: How I achieved a 232x Faster Kernel

In a GPU Mode auto-research contest, I achieved a 232x speedup over the baseline for batched QR factorization by using Codex to iteratively optimize a kernel. This post details my approach: learning the math behind Householder reflections, leveraging the blocked Householder algorithm with WY updates, and using a tight feedback loop with the popcorn CLI to hill-climb. I also share the importance of idea diversity to escape local maxima and the lessons learned from making over 1500 submissions.