Large language models (LLMs) like ChatGPT and Copilot are rapidly changing how programmers work. But how are these tools *really* impacting coding practices, productivity, and even learning? A new research paper delves into the evolving dynamic between humans and LLMs, uncovering intriguing insights from user studies on using LLMs in programming tasks. One key focus was how programmers interact with LLMs. The study found a range of interaction styles, from treating the LLM as an expert consultant for learning and exploration to using it primarily for generating code solutions, debugging, or even handling tasks unrelated to coding. Prompting strategies also varied. Some programmers used single, comprehensive prompts (zero-shot prompting), while others broke down tasks into multiple, smaller prompts. Interestingly, programmers often combine several techniques, reflecting the iterative and often unpredictable nature of working with LLMs. Beyond prompting, the study examined how programmers integrated LLMs into their workflow. Some relied solely on the LLM for entire solutions, while others adopted a hybrid approach, combining LLM-generated code with manual programming. Experts tended to write more code themselves, potentially to reduce the cognitive load of understanding AI-generated code. A significant chunk of developer time shifted from writing code to reviewing and understanding LLM responses—often exceeding 50% of total task time. So, are LLMs boosting productivity? The research suggests that they generally enhance productivity by automating tasks and reducing coding time. However, this benefit wasn't universal. For complex tasks or when less experienced programmers encountered unexpected LLM output, the debugging and understanding process could actually decrease productivity. Similar complexities arose when evaluating task performance based on LLM-generated code. While LLMs often led to more correct and secure code for various tasks, there were notable instances where quality suffered, especially with specific or less rigorously evaluated tasks. On the learning front, LLMs showed promise in helping students grasp programming concepts. Gains were evident when students were tested before and after using LLMs. However, these advantages diminished when students used LLMs freely during coursework. This raises concerns about over-reliance and potentially hindering the learning of fundamental programming skills. The research provides a fascinating snapshot of the human-LLM interaction in programming. As LLMs become even more sophisticated, understanding these dynamics will be essential to maximizing their potential while mitigating risks to productivity and learning. Future research should delve deeper into standardizing evaluation metrics and exploring how different interaction patterns impact performance and user experience. This research serves as a valuable roadmap for both LLM developers seeking to create more intuitive and user-friendly tools and educators looking to thoughtfully incorporate these powerful tools into the curriculum.