Adapt LLM to specific tasks: a convenient alternative to the “GPT-for-all” approach?