Re: Non-LLM example where we do not in practice use original training data

On Wed, May 07, 2025 at 07:36:13PM +0200, Simon Josefsson wrote:
>I think Thorsten Glaser's proposal on the surface looks more in line
>with what I would want to see, but I don't think we understand the full
>implications of any of the proposals right now.
>
>https://lists.debian.org/debian-vote/2025/04/msg00118.html
>
>Some approach to have LLM tools in 'main' when they can work with models
>that would be appropriate for inclusion in 'main' seems fine to me.
>Then we can ship models for that tool in 'non-free', for people who want
>to work with some larger model.  I don't see a need to permit LLM tools
>in 'main' that are unable to work with any libre LLM model, those tools
>could go into 'contrib'.

I'm not talking about LLMs.  I barely care about LLMs more than
I care about megahal.  My fictitious example is based on an
actual problem we have with computer vision classifier models.
So, if the models which already exist in main cannot exist in
main (either under current policy, Thorsten's proposal, or Mo's
proposal), are we going to pretend that they're not there?  Are
we going to throw them out?  Both those things are bad for our
users.  Are we going to fix them instead?  If so, how?  Thorsten's
§5 handwaves implementation details away, but I think that this is
actually very important, and I don't even need face detection
technology to use my computer.