RE: https://mas.to/@zzt/117134157732461816
This highlights one of the big problems with LLMs for code generation.
If a thing has been done often enough that there are are loads of examples of it for an LLM to train on, you probably shouldn't be doing it again. You should be using some abstraction over the thing and not copying and pasting it.
If a thing is rare enough that it has been done a handful of times in the training set, there's a good chance that the LLM output will meet the legal definition of a derived work, you may be accidentally violating a license / infringing copyright.
If a thing is so rare that there are no examples in the training set, the LLM will probably produce something that looks right. There is even a chance that it is right. Good luck.
