An acquaintance who works at FAANG says he still builds non-LLM ML systems because of the cost of running LLMs at that scale