Comment by TheOnly92 | Hacker News Reader

Comment by TheOnly92 | Hacker News Reader

I happen to be teaching a programming course currently, though it's not in English and the language I'm teaching is C. My current experience is that it does not seem like a majority of the students are using ChatGPT at all, even though I did encourage the use of it at the beginning of the course.

For my own course, I think several factors contributed to students not utilizing ChatGPT as much:

    - The assignments are not in English, and performance of ChatGPT in languages other than English is subpar.
    - The programming language that I'm teaching is C, I'd imagine Python/Javascript and other more popular languages might lead to different outcomes
    - I did specifically design the assignments so that copy/pasting the assignment to ChatGPT does not lead to a usable answer (by restricting use of certain standard library functions, making the assignment more complicated)
    - The course is not introductory, i.e. a previous course already taught the basic syntax of C and basics of programming, so I can make my assignments much more advanced

It's difficult to say if advancements in LLMs will make my job harder, where say copy/pasting my more complicated assignments can lead to correct results. But from what I can see right now, LLMs still have trouble solving novel problems, so it's probably always possible to come up with assignments that's difficult for them to solve.

I've had a few online code assessments that appear to have been hardened against ChatGPT "attacks". I failed at solving a problem that was just "Compute values of the Collatz conjecture for input n" because they wrote it to sound like an extremely difficult graph problem about being lost in a forest but being able to enter a "magic door".

I fed the problem into ChatGPT later and it was utterly unable to comprehend it, but confidently gave wrong answer after wrong answer.

Interesting, I'd suppose it makes sense that rephrasing the problem in a different way and also adding a bunch of nouns that have no relations with the problem at hand will definitely confuse LLMs. It will be interesting to see how LLMs will adapt to these as more and more of these techniques develop.

In my own assignments however, I focus less on algorithmic stuff but more on adding and mixing several things together. E.g. instead of just sorting, do group & sort, and a combination of a bunch of other practical stuff like reading big-endian binary files.

3.5 or 4? March or June model?

I've been using ChatGPT and Bard with C++ and it is quite helpful for boilerplate and reference (replacing Google/Stackoverflow).

Just for fun I asked ChatGPT 4 to calculate the RMSE between two vectors both in English and Portuguese (also translating RMSE to Portuguese) and it gave me the same code for both questions (asked in separate instances). It would be interesting to know what restrictions you applied.

It definitely depends on the task at hand, but when you're teaching programming you don't teach stuff with boilerplate. Using ChatGPT for reference to replace Google/Stackoverflow was definitely one of the ways I'd expected the students to utilize it, but it probably wasn't providing answers in ways a beginner/novice could understand.

I'd expect simple tasks like calculating RMSE to definitely be within the abilities of LLM, you might combine things like actually reading the vectors from a CSV file (or a custom format) and calculating RMSE then sorting them etc to see the limitations of LLMs. Most students have no issues with calculating RMSE, they have issues with trying to do all the other stuff that leads to it, and then the combination of sorting and other tasks.

Regarding the restrictions, most of them are just don't use itoa/strtod or strcpy or some other standard library functions.

Thanks, yes, RMSE is a simple task, I was focusing on its ability to translate the name (raiz quadrada do erro quadrático médio) correctly. It is funny that the Portuguese code has Portuguese comments but the name of the function is calculateRMSE even though I don't mention RMSE in the prompt.

I agree with you, in my experience, ChatGPT is a better search engine but it is not capable of composing the various parts of an application in a cogent manner. I also think that the current UI is not appropriate for software development and I am sure there are efforts going on to create something closer to Jupyter notebooks for programming. That may be a game changer for your students (and you).

True that based on my experience the variable names and function names remains in English despite the prompt, maybe its just the convention overall in the programming world, or maybe ChatGPT is finetuned to do so.

I don't think Jupyter notebooks or like similar REPL interfaces will help too much for my course, at least in the current syllabus. I'm aiming to teach about pointers, memory management etc, the more fundamental parts of how to interact with computers instead of a high level language. Though I would agree that the current UI is suboptimal, some improvements in allowing students to visualize memory layouts and see how their code manipulates memory will help a lot.

I'd hazard to guess that REPL of a simple virtual machine would work wonders for teaching about pointers and memory management.

I can't recall exactly but I think https://godbolt.org/ might do that for example?

A simple virtual machine might be nice, but imagine the pains of trying to guide students to install something across different environments.

Godbolt is a compiler explorer, it shows disassembly of a code but there's nothing to visualize each step in the process.

> from what I can see right now, LLMs still have trouble solving novel problems

Have you been using 3.5 or 4?

I have been using 3.5, but when I was designing the assignments, I asked a friend who had access to 4 to check it, and the code it produced was still incorrect.

Reply on news.ycombinator.com