[gdal-dev] RFC111 AI/LLM tool policy: proposal for a significant revision

[Greg Troxel]

Even:

  I would somehow keep the idea that it is not settled law if LLM output
  is a derived work of the inputs, in the discussion of legal issues.  I
  keep hearing that people are able to get LLMs to output training data
  verbatim, proving it's in the model,

  Other than that, your revisions sound good to me.  If it turns out
  other projects figure out a good way to interact and the legal issues
  are resolved, this can be changed later.


Greg (with no vote)


"Darafei \"Komяpa\" Praliaskouski via gdal-dev"
<gdal-dev at lists.osgeo.org> writes:

> Is it possible to get a better overview of the actual negative impact
> you're seeing? Like, with actual examples of annoyances that happened.

Not gdal, but I was involved with review of an LLM PR to hamlib.  It was
extra verbose, made a lot of off-point changes, and ultimately was
wrong.  It was harder to figure out than it should have been.

A (cooperative, good at sw) human would have just made the one change
that was maybe needed, and not rolled in unrelated reorganization, and
would have actually explained what the issue was, rather than a pagelong
lecture on generalities that wasn't in the end actually on point.

I also had a bad experience a long time ago, before there was broad LLM
awareness, on a (non-geo project) mailinglist, with an email-submitted
patch (eq to a PR), and after I commented on it, an LLM
rejoinder/commentary was posted.

The other thing you aren't seeing is that github service quality is
plummeting under the load of all the submitted PRs, since mid-December,
but that's off-point for this discussion.