[sword-devel] GitHub Copilot

[pinoaffe]

Hi, 

Tobias Klein <contact at tklein.info> writes:
> I would encourage you to give GitHub Copilot a try if you have not done so yet!

there are severe technical, legal and ethical issues with using LLMs for
code generation, an excerpt:

- LLMs are very good at generating text/code that *looks*
  plausible/correct, but that is not necessarily correct.  Therefore, it
  is often more difficult to verify that code generated by an LLM is
  not subtly broken than it is to write similar code yourself.  *PLEASE*
  do not underestimate this issue.

- LLMs are trained on any publicly readable code with no regard to the
  law, the license terms of the code, and the intentions of the
  author(s).  This is highly unethical in my opinion

- LLMs are known to sometimes produce one-to-one (or very close) copies
  of some of the data they've been trained on.  Thus, you may end up
  copying other people's code word for word without being aware, this
  can open you up to legal issues since this is almost always a
  copyright violation

- LLMs may also produce code that is a "mix" of two pieces of code it
  was trained on.  This is even more difficult to recognize than a
  one-to-one copy, and incorporating it into your code may still be a
  copyright violation

These are just a few examples, there are many more issues with using
LLMs for code generation.  Therefore, I strongly recommend against using
LLMs for code generation or modification

Kind regards, pinoaffe