To be clear, I said I’d like to think all the output should be copyleft. I actually think that, since the training set is a hopeless mishmash of conflicting licenses, it’s just massive copyright infringement and can’t be used for any purpose at all. Part of me is surprised some judge hasn’t started ordering them to just straight-up delete the entire models yet.

…Hey, wait a minnit! That’s not a bright side at all!

(Personally, I’d like to think that the output of any LLM whose training data includes copyleft code ought to be required to be copyleft.)

Too bad permissive licensing instead of copyleft (even as flawed as GPLv2 is) is a deal-breaker.

Not to be confused with a Cuddlefish:

Not to be confused with “theater kid (extremely derogatory)” for Vampire The Masquerade.