The rise of AI in education is transforming how students learn, especially in technical fields like computer programming. But what do students actually *want* from these digital assistants? A new study reveals surprising insights into student preferences for AI teaching assistants in introductory programming courses. Researchers deployed an LLM-powered digital TA, called CodeHelp, in a large undergraduate programming course. CodeHelp was designed with a crucial "guardrail": it provides guidance and explanations but *never* generates complete code solutions. This encourages students to actively problem-solve instead of passively receiving answers. Over 6,000 queries later, the results revealed a clear pattern. Students overwhelmingly valued the instant, 24/7 availability of the AI TA, especially during late-night study sessions before deadlines. But perhaps the most surprising finding was students' strong preference for *scaffolding* over direct solutions. They wanted the AI TA to guide their thinking, offer hints, and point them in the right direction—not simply hand them the answer. This desire for agency and active learning challenges some common assumptions about student behavior. It suggests that, given the right tools, students are eager to engage deeply with the material and develop their own problem-solving skills. Students also emphasized the importance of clear, concise explanations tailored to their level of expertise. They wanted support that felt personalized and relevant, not generic or overly complex. This research offers valuable guidance for designing effective AI teaching assistants. By prioritizing scaffolding, personalization, and instant availability, we can create tools that empower students to become independent, confident programmers. The future of AI in education is bright, and by listening to student voices, we can ensure that these powerful tools are used to enhance—not replace—the human element of learning.