Are there any pain points that cannot be satisfied when using LLM (similar to chatGPT)?