An auto-research loop finds a 232-fold kernel speedup

A practitioner writing as sankalp pointed an auto-research loop built on OpenAI Codex at GPU Mode’s qr_v2 problem, a QR-decomposition kernel, with an AGENTS.md and a problem statement standing in for Karpathy’s program.md and the contest’s submission logs as the record of what worked, and reports a kernel 232x faster than the baseline (blog). The contest ran from June 15 to June 30, 2026 and the write-up is dated July 8, five weeks before the issue that carries it and older than the August 12 speedrun report the newsletter sets it against: the same issue’s finding that 153 autonomous speedrun runs closed 81.7% of the gap to the human record without inventing a new method. In an essay dated August 12, Timothy Gowers argued that LLMs excel at search-heavy proof discovery, where breadth and cheap exploration pay, while humans still prune deep trees best (essay). In the trajectory the result extends the speedup line from Opus 4.6’s 34x and Mythos Preview’s 52x to a single practitioner’s loop, and stands beside Karpathy’s 650-experiment autoresearch as the harness pattern leaves the labs.