what happens when an LLM is trained on GPL licensed code
Also posted to Bluesky.
17 replies
@zachleat That used to keep me up at night when I worked at Adobe. Back then CoPilot would spit out code verbatim from GPL projects. We didn't want to be the legal test case so the OSPO (engineers/lawyers) recommended strongly we do not use these AI code generators until there is established case law.
That is for a Fortune 500 level company. As a dev hacking on a side project the risk is not as great. As my Dad would say, “Never sue poor people.”
@macdonst this is surely not a ticking time bomb
@zachleat it's a giant game of chicken. MS has some protections for their commercial customers though. Still I would not want to be the test case in litigation.
@macdonst holy shit that quote is *WILD*
@macdonst @zachleat this does not look sane at all. Thanks MUCH for the quote.
Imagine that this kind of responsibility extends to patent suits or to purpose (two different things)…Code is made to be readable so that a human can read and write it and take responsibility. Not doing so is just part of the stupid rush of productivity that managers hope but just makes no sense.
If code is to be produced faster then you need better (more reliable…) libraries.
ha ha ha licenses aren’t good vibes — we’re coding only with vibes now
gotta make your whole LLM open source and licensed under GPL 😜
@zachleat taps the “when you're famous they let you do it” sign
@zachleat Nothing except I guess https://githubcopilotlitigation.com/case-updates.html
@kisaragi_hiu seems fair to me
Gray area. You can legally study GPL code and create new, non-GPL code based on those learnings. It boils down to whether or not a judge will rule that AI is producing unique works vs infringing existing copyrights.
Large IP-Laundering Models