Re: Open vs proprietary LLMs

Pierre-Elliott Bécue <[email protected]> Tue, 25 Aug 2026 19:14:50 +0200

Newsgroups	gmane.linux.debian.devel.general
Message-ID	<[email protected]>

Simon Richter <[email protected]> wrote on 25/08/2026 at 11:48:04+0200:

> Hi,
>
> On 8/25/26 5:52 PM, Pierre-Elliott Bécue wrote:
>
> [Training exclusively on Debian resources]
>
>>> The training set would be too small to argue that copyrights are
>>> sufficiently diluted that we can ignore licenses (which is a dangerous
>>> position that we should not adopt), so code generation is out.
>
>> Code generation is not out, there is nothing very specific to Debian,
>> and you can take a strong model and add a RAG layer for Debian specific
>> data, documentation etc.
>
> The "take a strong model" bit is key here: it requires us to import a
> binary blob that is not reproducible for us, so that doesn't satisfy
> the criteria from Wookey's question.

I did not read Wookey's question as "using a reproducible" model, but
rather as "using a model we have on our infra or on a dev infra".

> There is no model that satisfies these criteria, and building one is
> not an LLM (lacking the "large" aspect), so its use would be limited
> to search indexing and almost-verbatim reproductions of code.

There is no model that satisfies the criteria you set, but again, I'm
not convinced that it was what Wookey asked for.

> Search indexing might be useful, but lossy reproduction of code is
> worse than useless, because it still requires review effort but does
> not even launder copyright under the most benevolent interpretation of
> the law.

Assuming that I understood right, we would not have a lossy reproduction
of code, but rather a Qwen, Kimi, or whatever, with coding caps + a RAG
on Debian parts.

-- 
PEB

signature.asc (application/pgp-signature, 853 B)