“Challenges for real-world use of HuggingGPT include slow performance due to multiple inference rounds, reliance on a long context window, and instability of LLM outputs.”