Local LLM inference – impressive but too hard to work with