2025 Autumn · New-TA Peer Sharing
A panel with experienced TAs from various courses (this time Liu Hao, Chen Jiajie, Lu Jun), online/offline combined, with recordings provided for makeup study; sign-in at the start, gifting new TAs Nonviolent Communication, and announcing the rest of the semester's TA seminars (lunch meetings in a department meeting room, relaxed topic-based discussion, fixed one midday).
Topics
- Small tips to make TA work more enjoyable?
- A sense of achievement; the ultimate question "why are you a TA?"
- Various automation tools (many students)
- More "learning alongside students" than "teaching students"
- Written assignments with only objective questions can't well assess mastery
- Don't over-internalize: accept you can't satisfy every student or "save" them all
- Clarify your role and responsibility in the teaching team
- We need to teach students to tell whether an AI answer is correct
- The most time-consuming TA task?
- Written-assignment design/grading, writing solutions (sometimes enjoyable)
- One option: grade a third of students each time
- Looking at the whole process, the longest is usually designing new assignments/labs (a few months)
- Tutoring is easy to overlook but accumulates over the semester
- Grading exams (especially in exam week); participating in courseware production, etc.
- The most enjoyable/grinding things?
- Grinding: plagiarism-related, bad attitudes (full slacking), recording wrong grades
- Free discussion: any other questions?
(Attachment: parts of the 2024 Spring new-TA training for critical reference.)
Impact of LLMs on TA work
Observing students' typical LLM uses (pasting problems to a chatbot, VSCode-integrated LLMs, the model auto-explaining when selecting code during tutoring). How to allow/use LLMs is debated without consensus. Students face time pressure, but how to build ability? In courses with more written assignments (computer-systems intro), students ask LLMs first; some submit LLM output verbatim (about 85% by result) but can't do it in exams. In the last year or two, LLM reasoning has improved — need to re-observe its impact. How to design AI-proof assignments? (Make them harder? Allow students to use AI? Require using AI and recording the help? Involve as many third-party libs as possible?)
Why is teaching debugging so hard?
Debugging is a tool that takes time to learn and use (like a wrench); students may not see its necessity. After LLMs, they can help with debugging too; you need a rigorous thinking method/process first. LLMs are like a crystal ball, giving rough insight/direction, useful for lower-level easy assignments; but deep, problem-specific debugging still needs line-by-line thinking, which LLMs can't replace. LLMs lower the cost of learning to debug — a positive change.
How TAs can help with debugging
Teach a person to fish: "tell the student where the bug is" ✗ vs "tell the student how to find the bug" ✓; but students may want to finish quickly for a higher grade. In computer-systems intro, debugging C code mainly uses gdb; tried small labs to experience gdb; LLMs can help "look up commands."