it's very fsf to hyperfocus on a licensing dispute when the AI industry is laundering every single *GPL application in existence into closed corporate slopware.
The FSF is quite involved with the current developments around LLMs. The FSF tries to get Anthropic to make datasets publicly available, they are less interested in getting into a licensing dispute from what I can tell.
And some problems are conceptually timeless; I don't even know what "to modernize" would mean for the FSF, or what irrelevance would mean in this context.
People seem to be assuming that the FSF’s main thing would be objecting to making non-copyleft software based on training on copyleft software. It’s possible that they are also looking at the side of being able to make more Free Software and not ready to say something comprehensive.
AI laundering of copyright isn't just a free software problem. It's an IP problem. Surely the legal fallout of today's license laundering of software isn't going to be substantially different from future laundering of e.g. big studio movies?
It’s not ‘laundering’; it’s reimplementation. rms reimplemented a ton of proprietary Lisp Machine code; the GNU project itself is a reimplementation of POSIX. I would be very surprised if the FSF changed course now and tried to argue that reimplementation is somehow problematic.
N.b.: I do not consider training a model on data to be copying the data; even if it were copying, it seems pretty clearly a transformative — and thus fair — use. I am neither a lawyer, nor a politician, nor anything more than an amateur philosopher, so of course I could be wrong.
the GNU project itself is a reimplementation of POSIX
POSIX is a specification... it's not the same thing at all.
I do not consider training a model on data to be copying the data; even if it were copying, it seems pretty clearly a transformative — and thus fair — use
Is zipping a file transformative? What about wrapping a file in an installer? The actual data gets transformed vastly. One cannot compare the data from a huffman tree to the original text data without a lot of extra steps!
It's super convenient for people and researchers to talk about LLMs as a form of "lossy compression" when the mood strikes them, and of course, to throw away this terminology when it gets convenient.
I believe those first two questions are considered conduct and therefore it doesn't make sense to ask if it is fair use. Something is copyrightable if it is fixed in a tangible medium of expression. It is not a copyright violation to reimplements an API, for example (see google v oracle). Reimplementing therefore is totally fine to my knowledge.
Either way, LLMs are awful and I hope the companies that abuse open source licenses illegally have the law catch up to them at some point.
j3s | 8 hours ago
it's very fsf to hyperfocus on a licensing dispute when the AI industry is laundering every single *GPL application in existence into closed corporate slopware.
yawaramin | 7 hours ago
Not much they can do about the latter, so why would they focus on it?
j3s | 7 hours ago
in my view, the fsf needs to modernize fast or be deemed irrelevant, so they should focus on modern problems.
mitsuhiko | 6 hours ago
The FSF is quite involved with the current developments around LLMs. The FSF tries to get Anthropic to make datasets publicly available, they are less interested in getting into a licensing dispute from what I can tell.
mordae | 4 hours ago
I would prefer organizations not to abandon old problems, letting them become load-bearing infinite money printers for the few, thank you.
Feel free to start a new organization focusing on whatever you think is a fresh issue worthy of spending your life on dealing with.
zetashift | 2 hours ago
And some problems are conceptually timeless; I don't even know what "to modernize" would mean for the FSF, or what irrelevance would mean in this context.
mordae | 4 hours ago
What they are doing right now is making it legal for EU companies to yank OnlyOffice into commons, in the eyes of the EU institutions.
If it works, NextCloud is positioning itself as an alternative to MS Office 365 (including Outlook), the biggest dependency of EU on the US.
This is actually huge.
hsivonen | 6 hours ago
People seem to be assuming that the FSF’s main thing would be objecting to making non-copyleft software based on training on copyleft software. It’s possible that they are also looking at the side of being able to make more Free Software and not ready to say something comprehensive.
gspr | 5 hours ago
AI laundering of copyright isn't just a free software problem. It's an IP problem. Surely the legal fallout of today's license laundering of software isn't going to be substantially different from future laundering of e.g. big studio movies?
rau | an hour ago
It’s not ‘laundering’; it’s reimplementation. rms reimplemented a ton of proprietary Lisp Machine code; the GNU project itself is a reimplementation of POSIX. I would be very surprised if the FSF changed course now and tried to argue that reimplementation is somehow problematic.
N.b.: I do not consider training a model on data to be copying the data; even if it were copying, it seems pretty clearly a transformative — and thus fair — use. I am neither a lawyer, nor a politician, nor anything more than an amateur philosopher, so of course I could be wrong.
alexandria | 32 minutes ago
POSIX is a specification... it's not the same thing at all.
Is zipping a file transformative? What about wrapping a file in an installer? The actual data gets transformed vastly. One cannot compare the data from a huffman tree to the original text data without a lot of extra steps!
It's super convenient for people and researchers to talk about LLMs as a form of "lossy compression" when the mood strikes them, and of course, to throw away this terminology when it gets convenient.
Cajunvoodoo | 18 minutes ago
I believe those first two questions are considered conduct and therefore it doesn't make sense to ask if it is fair use. Something is copyrightable if it is fixed in a tangible medium of expression. It is not a copyright violation to reimplements an API, for example (see google v oracle). Reimplementing therefore is totally fine to my knowledge.
Either way, LLMs are awful and I hope the companies that abuse open source licenses illegally have the law catch up to them at some point.