Post B2nrpgAsReQJYHw7o8 by asie@mk.asie.pl
(DIR) More posts by asie@mk.asie.pl
(DIR) Post #B2npKtvV48LdJkwLa4 by asie@mk.asie.pl
2026-01-30T06:53:11.039Z
0 likes, 0 repeats
I just learned that GNU gettext decided to release a 1.0 version after 30 years of development. Congratulations! The new feature which made this version deserving of that honor is adding automatic LLM translation support!I think I am inches away from a psychotic break! I just let out what felt like a simultaneous hollow laugh and scream!
(DIR) Post #B2npiHjEjg6RncM1mC by asie@mk.asie.pl
2026-01-30T06:57:26.046Z
0 likes, 0 repeats
It does, however, tell us where at least some of the GNU folk stand on LLMs:- They request that their users host LLMs locally, and not in a third party cloud,- They encourage people to use open weight models, but provenance seems less relevant - Ministral 3 14B is explicitly named as an example in the documentation.
(DIR) Post #B2nq65LsP6QkMamDHk by asie@mk.asie.pl
2026-01-30T07:01:44.129Z
0 likes, 0 repeats
To be clear, yes, I agree with them that using open weight models on a device you control without relying on external data centers and proprietary third party services is the lesser evil, compared to just dialing into ChatGPT.But from the organization that felt it was a moral imperative to deny CPU microcode updates from users in the name of software freedom, I think I expected a little more than "don't forget to use large matrices which are Apache 2.0 licensed".
(DIR) Post #B2nq7Vl5i4t6YHFanw by natty@astolfo.social
2026-01-30T07:01:28.298Z
0 likes, 0 repeats
@asie@mk.asie.pl this is a shitpost right This gotta be a shitpost
(DIR) Post #B2nq7WrraNtnzZQW00 by asie@mk.asie.pl
2026-01-30T07:01:55.171Z
0 likes, 0 repeats
@natty@astolfo.social https://gitweb.git.savannah.gnu.org/gitweb/?p=gettext.git;a=blobdiff;f=gettext-tools/doc/gettext.texi;h=f2f09586180884ca966e1f3e4e841dc5cb62753d;hp=f8913e5c4d5f3c33847633a5d7ca2ce536f6c0df;hb=c979c72860cbfaba7d2e745b8bf9dcf27f7f030f;hpb=7c7ff10dcf979358e32b0c0da72fa4f00d45aab5
(DIR) Post #B2nqAQCGqPa8AhLq9w by asie@mk.asie.pl
2026-01-30T07:02:33.397Z
0 likes, 0 repeats
@natty@astolfo.social The link might not work if you don't copy-paste it, but here's the source:https://gitweb.git.savannah.gnu.org/gitweb/?p=gettext.git;a=blobdiff;f=gettext-tools/doc/gettext.texi;h=f2f09586180884ca966e1f3e4e841dc5cb62753d;hp=f8913e5c4d5f3c33847633a5d7ca2ce536f6c0df;hb=c979c72860cbfaba7d2e745b8bf9dcf27f7f030f;hpb=7c7ff10dcf979358e32b0c0da72fa4f00d45aab5
(DIR) Post #B2nqQd2mqa9IeJdxPE by asie@mk.asie.pl
2026-01-30T07:05:26.851Z
0 likes, 0 repeats
To be clear, yes, I agree with them that using open weight models on a device you control without relying on external data centers and proprietary third party services is the lesser evil, compared to just dialing into ChatGPT.But from the organization that felt it was a moral imperative to deny CPU microcode updates from users in the name of software freedom, I think I expected a little more than "don't forget to use large matrices which are allegedly Apache 2.0 licensed".
(DIR) Post #B2nqtxOyEi7MBa0LNw by asie@mk.asie.pl
2026-01-30T07:10:44.225Z
0 likes, 0 repeats
For a more principled example: Debian's Deep Learning Team (which does not represent the Debian Project) had been drafting an unofficial machine learning policy for some time - even before the LLM boom - and it's more considerate of these issues:https://salsa.debian.org/deeplearning-team/ml-policy/-/blob/master/ML-Policy.pdf?ref_type=heads
(DIR) Post #B2nrhXCru4xDQDz5Ie by lua@vixen.zone
2026-01-30T06:56:12Z
0 likes, 0 repeats
@asie oh my god. that's awful. fuckkkkk
(DIR) Post #B2nrhYCCE40iUQg3tI by lua@vixen.zone
2026-01-30T07:14:05Z
0 likes, 0 repeats
@asie like i mean. i get it. LLMs are great at translation. but for that to be the 1.0 release feels SO lame !!!!
(DIR) Post #B2nrhYkaADW4D4lWUK by asie@mk.asie.pl
2026-01-30T07:19:40.366Z
0 likes, 0 repeats
@lua@vixen.zone I mean, sure, for things like highly derivative, expected phrases with adequate context, an LLM can probably do a good job. Most people aren't exactly using gettext of all things to translate creative masterpieces, and many open source translation services like Weblate have had translation service integration for a long time.I'm more astonished that there was seemingly no consideration given to data provenance. However, I'd like to award them one (1) comedy point for naming the LLM translation tool "spit".
(DIR) Post #B2nrpgAsReQJYHw7o8 by asie@mk.asie.pl
2026-01-30T07:21:11.700Z
0 likes, 0 repeats
Oh, they named the LLM tool "spit", which is objectively funny. One (1) comedy point.
(DIR) Post #B2nulWfW93kKm4ymn2 by wolf480pl@mstdn.io
2026-01-30T07:54:08Z
0 likes, 0 repeats
@asieI asked one of my former coworkers who was dealing with machine learning what is the preferred form for modification: training data & script, or weights. He said weights.So if you don't care about copyright in terms of labor dispute or author's relationship with the work (which I think GNU doesn't) then it boils down to: can you modify software that runs on yur computer, and can you get in trouble for doing so.Hence they treat open weights models as free.
(DIR) Post #B2nuxgre41Sds2S7NI by wolf480pl@mstdn.io
2026-01-30T07:56:20Z
0 likes, 0 repeats
@asieI don't agree with gettext authors because I do care about copyright, and about reproducability, but I see where they might be coming from.
(DIR) Post #B2nvDJD3mLl98lut96 by asie@mk.asie.pl
2026-01-30T07:59:03.716Z
0 likes, 0 repeats
@wolf480pl@mstdn.io The preferred form for making modifications is weights, but I think that's in part because arriving from training data to weights is quite expensive. I also think this proposed analogy omits the history behind why we decided source code alone is insufficient, and why including build scripts is also expected these days.I wonder what stance the FSF will end up taking on machine learning models which used GPL licensed code as data input, though. Will they also "not care about copyright" then?
(DIR) Post #B2nvT3ISi60EQ9wz4K by wolf480pl@mstdn.io
2026-01-30T08:02:01Z
0 likes, 0 repeats
@asieyeah, it's quite possible they're just being hypocritic here...
(DIR) Post #B2nvrJX00DHVimJpXE by wolf480pl@mstdn.io
2026-01-30T08:06:24Z
0 likes, 0 repeats
@asieand yes, if reproducing a model from training data was something you could do in 20min on a thinkpad, I think ML folks would prefer that to open weights.