[kwlug-disc] GitHub CoPilot discussion
Ronald Barnes
ron at ronaldbarnes.ca
Fri Jul 22 16:47:33 EDT 2022
The latest Linux Downtime podcast (episode 51) at
https://linuxdowntime.com/, discusses GitHub's CoPilot project for 20
minutes.
It's hosted by prolific podcaster Joe Ressington, with Martin Wimpress
of Mate Desktop Environment and Hayden Barnes, former Cooperating
Attorney with EFF, have quite a bit to say about CoPilot, GitHub's AI
source code assistant tool.
Surprisingly, both Martin and Hayden really, *really* like it (and would
pay for it). It's incredibly useful to them and Martin is collaborating
on a project using Go language, which he's new to, and CoPilot is saving
him a lot of time at Stack Exchange, etc.
Even the code comment suggestions seem appropriate, across various
(human) languages.
When writing documentation, Hayden had it suggest the exact OpenSUSE
zypper command for pulling in the dependencies he'd just written out for
Debian... Nice. (04:15)
Hayden discusses the legality of the copyright issues (05:00) and makes
a good case for them to be not significant (08:20), specifically
"Incidental Inclusion" (08:45).
Also, analyses (by GitHub) found only 0.1% (06:10) of code suggestions
were "verbatim" from repositories.
Most suggestions of proprietary code come from that author's own code base.
Hayden suggests this is no different from looking at someone else's code
to see how something's done, then writing your own (09:30).
I'm reminded of the Oracle vs Google API copyrighting / header files law
suit, and suggestions that there are only so many ways to write a "for
loop" in any given language.
Also raises the question, if an AI is trained on a gazillion voice
samples, then it speaks, is the copyright of one of the owners of a
sample voice being infringed? Hayden says, "No" although not (yet)
applied to code (10:25).
Hayden says even the EFF would normally call such derivative works fair use.
More information about the kwlug-disc
mailing list