> the world has decided that it's completely fine to train LLMs on e.g. GPL code without the output being considered derived
I think it's more of a "the constitution is not a suicide pact" approach to GPL: while the GPL is obviously violated by its inclusion in the training corpus, we have two options:
- demand the destruction of LLMs that are trained on copyrighted or copylefted works and the training of new ones
- accept that LLMs that are trained on copyrighted or copylefted works are more important to humanity than the rights of the authors
I think it's more of a "the constitution is not a suicide pact" approach to GPL: while the GPL is obviously violated by its inclusion in the training corpus, we have two options:
- demand the destruction of LLMs that are trained on copyrighted or copylefted works and the training of new ones
- accept that LLMs that are trained on copyrighted or copylefted works are more important to humanity than the rights of the authors