So a LLM wrote 432 kernel variations and it found what was the faster...