Re: Ballot Option: Allow public AI for translations

Pirate Praveen <[email protected]> Mon, 10 Aug 2026 05:35:33 +0530
Newsgroups	gmane.linux.debian.devel.vote
Message-ID	<[email protected]>
---updated proposal---

Using its power under Constitution section 4.1 (5), the project issues 
the following statement describing its current position on AI-assisted 
contributions. This statement describes the position of the project at 
the time it is adopted. That position may evolve as time passes without 
the need to resort to future general resolutions. The GR process remains 
available if the project needs a decision and cannot come to a consensus.

LLMs have made a huge leap in language translations, but at the same 
time depending on big tech companies to provide this service is not a 
good idea. When it comes to generating code, we also need to consider 
how it weakens the idea of copyleft.

1. Public AI is AI as public infrastructure like highways, water, or 
electricity. An example of Public AI is https://publicai.co, which is 
funded by Mozilla, the Future of Life Institute, and the Center for 
Cultural Innovation. It also complies with Open Source AI definition by 
OSI (https://opensource.org/ai). For AI to be accessible to people, it 
has to be available easily to people and not for profit collaborations 
between multiple organizations and public funding is essential to keep 
it sustainable.

2. Public AI must use Free Software models.

3. At this point, we don't want to insist on training data being freely 
available, similar to how we don't insist on Free Hardware Designs. The 
challenges to practically being able to use the available data should be 
factored in. Similar to how costly it would be to produce hardware from 
Free Hardware Designs, even when training data is available, huge 
computing power is required to use it practically.

4. Translations generated by Public AI can be accepted.

5. Code / documentation generated by Public AI can also be accepted if 
it preserves share alike / copyleft conditions of original code used in 
training.

---updated proposal---

I'd need seconds if people feel we need an option for allowing Public 
AI/Open Source AI. This would be closer to proposal A in practice, but 
could become more useful when the Free Software models mature and 
produce better results than currently.

I could also get some help in making the draft better.

On 7/30/26 3:02 PM, Aigars Mahinovs wrote:
> On Thu, 30 Jul 2026 at 08:53, Marc Haber <[email protected] 
> <mailto:mh%[email protected]>> wrote:
> 
>     On Tue, Jul 28, 2026 at 12:43:21PM -0300, Pirate Praveen wrote:
>      >----draft proposal ----
>      >
>      >LLMs have made a huge leap in language translations, but at the same
>      >time depending on big tech companies to provide this service is not a
>      >good idea.
>      >
>      >1. Public AI is AI as public infrastructure like highways, water, or
>      >electricity.
>      >
>      >2. Public AI must use Free Software models.
> 
>     Please consider not using "public AI" in your ballot name, many people
>     will consider ChatGPT, Claude, Grok, Mistral etc "public" because it is
>     publicly accessible on the Internet, and you're redefining that term in
>     your ballot.
> 
> 
> I agree. If a definition is to be used, the one currently available 
> definition is from https://opensource.org/ai <https://opensource.org/ 
> ai> , Inventing something else with the limited scope for clarification 
> only introduces confusion.

Ok I can change it to Open Source AI as defined by OSI.
> 
>     That being said, your proposal basically forbids using AI made
>     available
>     by the big global players, where the music is currently playing. While
>     it is legitimate to do that, this privileges rich DDs who can afford
>     big
>     hardware to run local models or DDs with a corporate sponsor.
> 
>     Is "public AI" as you define already available today, and what would be
>     the approximate costs to make use of that? How good is the code
>     produced
>     by public AI today, compared to what for example claude produces?

What I had in mind is https://publicai.co which I had mentioned in my 
original email.

While it may not match Claude or other "frontier" models right now, I 
think it can evolve over time to be better at it. I have used it for 
Spanish translation of visa application letters to Argentina and it was 
pretty good (other native speakers reviewed and confirmed it was good).

The crucial thing is, it is available to the public, much like 
Wikipedia, free of cost and the funded by Mozilla, the Future of Life 
Institute, and the Center for Cultural Innovation. https://publicai.co/about

So we are talking about communities coming together and hosting Open 
Weight Model for general public.

Asking itself,

is publicai.co open source ai compliant?

Yes. publicai.co is fully open‑source AI‑compliant:

Aspect	Details
Model	The service runs the Apertus language model, which is released 
under the Apache 2.0 license.

Platform	The chat and API infrastructure is built on OpenWebUI, an 
open‑source AI‑front‑end that is also available on GitHub.
Training Data	The data used to train Apertus was legally‑compliant, 
multilingual (including Swiss languages), and processed through rigorous 
filtering.

License & Governance	The whole stack (models, code, documentation) is 
published under permissive open‑source licenses; the Public AI Inference 
Utility operates as a nonprofit, community‑driven project.

Compliance	All components meet open‑source licensing requirements and 
are maintained with transparency, allowing users to inspect, modify, and 
redistribute the software and models.

In short, publicai.co uses open‑source models and runs on open‑source 
software, making it fully open‑source AI‑compliant.


> 
> Any PC with a video card like AMD XTX 7900 or similar could run quite a 
> few local AI payloads with decent speed. Finding a few of those for 
> Debian infra should not be a blocking problem. Additionally there are 
> many hosting providers all over the world offering AI-capable hardware 
> to be rented on hourly basis for as low as 0.37€/hour in a fully green 
> datacenter https://www.hetzner.com/dedicated-rootserver/matrix-gpu/ 
> <https://www.hetzner.com/dedicated-rootserver/matrix-gpu/> (as example 
> only, no endorsement intended). And Debian has also not asked for any 
> hardware or time donations from any providers yet. Historically Debian 
> has been able to acquire needed compute power with little or no cost, 
> when needs were clearly articulated and expressed.
> 
> Fronteer models will always be more capable, but in restricted contexts 
> (like translation or coding) it is possible to find good-enough solutions.
> 
> Providers of fronteer models would also likely jump at the opportunity 
> of providing complimentary pro access to Debian Developers for project 
> work, if we ever actually asked for that.
> -- 
> Best regards,
>      Aigars Mahinovs
OpenPGP_0x8F53E0193B294B75.asc (application/pgp-keys, 4.8 KB)
OpenPGP_signature.asc (application/pgp-signature, 840 B)