21. Most LLM training compute may go to recipe development, not final runs by @jacobcares (Jacob Morrison) · backlist 2026-05-07 · rubric 91.0