On 2026-07-31 07:31, Thanos Apollo wrote:
Jean Louis <[email protected]> writes:
unless it is free as in software freedom.
LLM models are not software. They are data assets. Not source code
you
can study and change.
That is why there is "as" in my above sentence "unless it is free as in
software freedom". The word "as" points to analogous principles. As
purely LLM, there is nothing that LLM can do without software.
And back to copyright issues:
- there are LLMs trained purely on libre datasets;
- you can modify and study datasets;
- you can modify and study inference software;
And with the time and development it becomes possible to modify and
study the weights.
And in my context, I keep saying, people should be running fully free
models, as free as in software freedom.
Majority of models you named as data assets are not free. Though some
are.
Overview of EU legislation on FOSS | Interoperable Europe Portal
https://interoperable-europe.ec.europa.eu/collection/open-source-observatory-osor/news/overview-eu-legislation-foss
The AI Act
Text: 2022/0303(COD)
"Free and open-source licence" is defined in recital 102:
"Software and data, including models, released under a free and
open-source licence that allows them to be openly shared and where users
can freely access, use, modify and redistribute them or modified
versions thereof (...) The licence should be considered to be free and
open-source also when it allows users to run, copy, distribute, study,
change and improve software and data, including models under the
condition that the original provider of the model is credited, the
identical or comparable terms of distribution are respected."
The term "free and open-source licence" is then used in article 25(4),
without further explanation, and also in articles 52(2) and 54(6), with
a few extra descriptive words:
"The obligations set out in paragraph 1, points (a) and (b), shall not
apply to providers of AI models that are released under a free and
open-source licence that allows for the access, usage, modification, and
distribution of the model, and whose parameters, including the weights,
the information on the model architecture, and the information on model
usage, are made publicly available"
Even though the EU isn't for me any kind of authority on what is "free",
they have nice stateements from which one could draw conclusions on how
to actually define free LLMs.
Who Does That Server Really Serve? - GNU Project - Free Software
Foundation
https://www.gnu.org/philosophy/who-does-that-server-really-serve.html
By the same logic DuckDuckGo is unethical unless you self-host
search. If remote search gets a pass as a computational service,
remote models are the same shape of dependency. Singling out the
model is special pleading.
Sorry, that is wrong comparison. With the DuckDuckGo you are not
uploading your information to be computed, calculated, processed, you
are definitely using the service, but not the substitute for your own.
With the remote use of the LLM, user may send personal information,
inventions, design, authored works, which could be processed on his
computer but is substituting it with the service, user is submitting
information to the server one doesn't own.
A further issue at hand is that Anthropic may by default use user data
to train its models. If an Emacs developer has that option turned on,
they are potentially feeding snippets of GNU GPL-licensed code into
Anthropic's training corpus. This risks diluting the legal protections
that Emacs and the GNU Project are built upon, as it becomes unclear
whether the model's outputs, derived in part from GPL-licensed inputs,
carry the same copyleft obligations or whether the training process
effectively strips those protections through sheer statistical
abstraction. The combination of default opt-in data collection and
opaque training pipelines means contributors may inadvertently undermine
the very licensing framework they rely on, not through malice, but
through a simple privacy setting left at its default value.
I would not say it that way if there would be OpenCode or Hermes or
other coding agent in use, but not Claude.
Those are harnesses, not models.
I think you missed the point.
When a user uses Claude Code (I may be wrong as I never tried it), that
user is almost by definition using the Claude model—a proprietary
service that acts as Service as a Software Substitute (SaaSS), where the
user may feed all their data to Anthropic's servers to be trained on
(see above). That's the problem.
And I was referring to the agent software, not to the model. I said
"coding agent" and I never said "model"—that was your correction, not
mine.
The point is this: when I use Hermes or OpenCode as my coding agent, I
run the software on my own computer. It is fully free software (not that
I verified every line, but it's the kind of thing we can verify). And it
runs the model locally on my machine—a model that is free, as far as I
can tell, though I am picky about verifying all datasets.
So the difference is not just about which model you use. It's about
whether you control the whole stack—the agent, the model, and your
data—or whether you hand all of that over to a company's server.
FWIW Teknium (Hermes lead) runs Hermes on Fable, and a lot of the Nous
side does the same. Most of hermes is written by anthropic models.
Anthropic also stole a lot of code from free software, so I couldn't
care less about what they do or what they train on. It's a proprietary
company, and I would never become a member or a user for any amount of
money in the world.
Anyway -- I don't blame Emacs developer for using proprietary Claude
Code software, but I am pointing to available solutions:
- OpenCode, Pi, Hermes, etc. there are many fully free software agents
or harnesses to be used;
- they can all provide good commit messages, as they may access same
service as software substitutes as developer already used (Claude);
- even better, Hermes or OpenCode and others from free software
community can access locally run LLMs and even smaller models of 4B
could provide commit messages, even running purely on the CPU;
So the choice isn't between "use Claude Code" and "write commit messages
by hand." The choice is between using a proprietary service that trains
on your data—which means feeding potentially GNU GPL-licensed code into
a company's training pipeline, diluting the very copyleft protections
Emacs depends on—and using a free software agent that runs on your own
machine with a local model that respects your freedom and keeps your
code out of Anthropic's hands.
--
Jean Louis
---
via emacs-tangents mailing list
(https://lists.gnu.org/mailman/listinfo/emacs-tangents)
