AI Use for My PhD Students

These are my current views on how PhD students, especially those doing theoretical research, should use AI. Since most of my work is theory-oriented, I encourage my PhD students to use strong AI models such as GPT Pro. I can pay for the subscription.

• A typical workflow would look like this: in each weekly meeting, we decide which direction to explore. You then use AI to map the problem landscape, identify what seems feasible or infeasible, and clarify the relevant background. At the same time, you should read the literature and do the necessary background work. In the next meeting, we discuss your findings, verify them carefully, discard weak directions, and decide what to pursue next.
• What I do not want is for you to use AI to generate a 40-page document, not read it carefully, and then send it to me. If AI helps produce something, you must understand every part of it as if you had written it yourself. We need to examine and verify every important claim. In my view, this is a basic requirement for doing serious theoretical research with advanced AI tools. In the end, I only want a concise report (say, 2 pages per meeting) from you: a clear summary/proof sketch that keeps only the core ideas worth discussing.
• This process is highly iterative. You should not expect one broad prompt to produce a complete result. As your understanding improves, you should keep refining your questions and pushing AI toward more precise and detailed reasoning.
• I also want students to understand the autoresearch style discussed by Karpathy. It is a practical workflow for algorithmic research, especially when you have data or a synthetic simulation environment for testing and validation. You should understand the logic of this approach and be able to apply it when a project fits that setting.