The P2P Foundation has some writings on licenses they label as "copyfair", "copyfarleft", and "copyjustright"

https://wiki.p2pfoundation.net/CopyFair_License

https://wiki.p2pfoundation.net/Copyfarleft

https://www.metamute.org/editorial/articles/copyfarleft-and-...

These are pretty interesting. Thanks for the links.

I like how this general direction is discussed a lot in France, and sad about learning a language taking so much effort.

What’s wrong with AGPL?

AFAIK AGPL permits linking without disclosing linked code like LGPL does. Meaning they can still distribute it as their own, train models and serve them with a simpler framework. Just can't build services powered directly by it.

AGPL is as strict as GPL. It's just that GPL allows you to modify the source, serve them as a web application and not have to disclose the code.

AGPL addresses that by putting code distributed as a web application in the same category as compiled code.

If my web server calls out to AGPL software (such as executing a transform on a PDF using Ghostscript) as an external binary is this covered even though the AGPL code is not directly compiled, linked or modified? I'd assume so otherwise it would be a very easy loophole and code could be firewalled into separate modules but I'm not sure what exact language covers this in the license.

It's untested in court.

GCC is written specifically to prevent this sort of firewalling of the compiler to prevent it from being plugged into another piece of software and avoid GPL requirements.

Good to know, it'll be interesting if it ever does get tested.

Even then there's always the analogue/human loophole I suppose, similar to captcha mechanical turks.

You seem to be correct. Still the note about training a deep learning model using the framework, then publishing it without applies.