tl;dr: A participant in GPU Mode's auto-research contest achieved a 232x speedup (from ~419ms to 1.8ms) over baseline on batched Householder QR factorization, placing 12th of 183 by running Codex agents in tight feedback loops with the popcorn CLI over 1500+ submissions. Key techniques included blocked Householder with WY-updates, fused V/T layouts, CUDA graph replay, and shape-specialized kernels. The author emphasizes that escaping local maxima required a "beam of candidates" strategy, human-in-the-loop steering, and stronger advisor models—arguing that idea diversity and research taste, not raw model intelligence, are now the bottleneck in auto-research.