Re: New "genAI" zig scheme implementation

Maxim Cournoyer <[email protected]> writes:

> I'm curious about them claiming this as their own copyright with the
> Expat license when it was agent-generated for the most part, IIUC. The
> evolving legal consensus surrounding machine-related source code seems
> to be going in the direction that it cannot be copyrighted, e.g. is
> public domain/CC0.  Not that it changes much in the end, though.

There are two separate issues and it almost seems to me that there is
(in the general case, not this one) intentional blurring to avoid the
harder issue.

In the US, it is pretty firmly established that only works created by
humans are eligible for copyright protection.  I believe it's similar in
the EU, modulo how firmly.  So yes, in the US, the new content is public
domain, and in non-US, my understanding is that public domain isn't
really a concept, and it's different but sort of the same.


The second issue is that LLM output is derived from its training data,
which I see as an unarguable fact.  The, there is the question of
whether LLM output is a "derived work" under copyright law, or whehter
there is some diffusion magic that says no it is not.  I believe it is a
derived work, under a straightforward interpretation of the law, and
that it should be considered a derived work.  If it is, that means that
the copyrights of the underlying training data still apply.  The key
point is that "machine output can't be copyrighted" does not lead to "if
you run copyrighted material through a machine it launders the existing
copyrights.  A simple thought experiment: Take the text of a book, and
write a program to output each paragraph and after the paragraph output
a "The quick brown fox jumps over the lazy dog.", and then publish that,
it would still be subject to the original author's copyright as a
derived work -- and no court would think otherwise.

This issue of whether LLM models and LLM outputs are derived works of
LLM training data is *not at all* settled (well established) law.
However, Anthropic settled a suit for $1.5B (1.5E9 US dollars!):

  https://authorsguild.org/advocacy/artificial-intelligence/what-authors-need-to-know-about-the-anthropic-settlement/

so they seem to think there is considerable merit to the theory.
Companies routinely settle for $1M or two as pocket change to avoid
litigation and juries, but $1.5B is huge.


One would think that a new genAI zig is likely derived from
actualProgramming zig, even if laundered through LLMs.   I see LLM
reimplementations as license violations.