my latest blog post "auto-research with codex: how I achieved a 212x faster…

Sankalp's GPU optimization experiment gave OpenAI's Codex a concrete feedback loop: revise code, check mathematical correctness, then measure execution time. The task was a specified form of QR factorization, a way of decomposing matrices. The GPU Mode competition supplied correctness checks and timing feedback across test cases, allowing repeated revisions to be judged against a fixed target. Experiment account

The announcement claimed a 212× speedup over the baseline; the linked article reports 232× against a torch.geqrf baseline. Both concern this particular kernel task. The transferable idea is the test-and-measure workflow that makes candidate improvements checkable.