Streaming is critical in making applications based on LLMs feel responsive to end-users.