2 ms·
> 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 consti
by orthoxerox 11d ago
> 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
- gspr 11d agoWell if it's the second, then surely I can train my LLM using the frontier labs' models too? And have my LLM "interpret and reproduce" GTA 6? How is any of this going to work out if we just decide to throw out copyright like this? (I am well aware that it is not currently technically feasible to copyright-wash GTA 6. Replace that game with a best-selling novel or whatever.) And, if we do decide that copyright is just gone, can we please make laws to that effect? Instead of just shrugging our shoulders and going "oh well, this new thing is useful, so we'll let it slide, but don't you dare to pirate movies, because reasons!"