“scientists often use code generating models as an information retrieval tool for navigating unfamiliar programming languages and libraries.” Again, they are busy professionals who are trying to get their job done, not trying to learn a programming language.
Very interesting read, especially since we teach programming to non-CS students, which is fundamentally different. Scientists are often multilingual (Python, R, bash) and use LLMs to get the job done. Their goal is not to write maintainable large software, but rather scripts that achieve a goal.
Now I wonder how confident they are that their programs do what they are supposed to do. In my own research, I’ve found invisible bugs (in bash, setting parameters, usually in parts of the code that are not algorithmic) that produce the wrong result. How much of the results in published articles is wrong because of these bugs?
We might need to improve the quality of code that is written by non-scientists.