Re: [Hamlib-developer] The use of LLM generated code in Hamlib (long)

[Greg T. <gd...@le...>]

Nate Bargmann <n0...@n0...> writes:

[agreed that clueless vibe coding is the biggest issue]

[agreed that finding something, like search, is fine]

As background, I am seeing comments on the net that when there is a very
sparse amount of information about something, LLM outputs are clearly
faithful to that limited input.   Recently there was some youtube video
mentioned on a list, and then someone pointed out the older paper that
the video was plagiarizing, almost entirely.

> When a contributor is prompting an LLM, he needs to be savvy enough to
> recognize when something clever is being offered and then stop and
> investigate.  It could be the LLM is offering up something that is
> either not licensed for Free Software distribution or is patent
> encumbered.

But how does someone do that?  If I, as someone who can program and who
understands copyright law, ask an LLM for code and get some back, I can
review it for correctness.  But how do I assess licensing?  My
assessment today is that it's a derived work of the training data, and
that data was collected both nonconsensually and without regard to
licensing.  So while there are unsettled legal questions, there is
absolutely no basis to conclude that the code meets the DCO.

I don't see how anyone could conclude that LLM output is ok, unless it's
below the threshold of copyright.