Re: Re: Proposal -- Interpretation of DFSG on Artificial Intelligence (AI) Models

Stephan Verbücheln dijo [Mon, Apr 28, 2025 at 03:46:27PM +0000]:
>(...)
>> The “source code” for a work means the preferred form of the work
>> for making modifications to it.
>
>In that definition, training data is quite obviously relevant. No one
>tweaks neural network model weights manually.
>
>Compare this to the previously mentioned example of S-boxes in
>cryptography. They are small and usually created manually.

I understand that, when you consider trained models as the "thing" to be
modified, the preferred form of modification is the model itself: What RAG
does is to have a base trained LLM (confering the "mastery" of language),
and training over it with the domain-specific knowledge.

-----BEGIN PGP SIGNATURE-----

wr0EABYKAG8FgmgQFn0JEOL2O0NT9FmJRxQAAAAAAB4AIHNhbHRAbm90YXRpb25z
LnNlcXVvaWEtcGdwLm9yZ5SbuobM6rOE6ACYpp9B2PYPdCzFsgjyiGEszl0ZbPt4
FiEEYLMJPZYQjly5cULv4vY7Q1P0WYkAAGanAP97QEe9VlzhRQRpN6KKzPrzN/vr
Sg3a28DcbTWyLjiu1AD/RlY8Io3/pzZP0Cf/PgajfEleZw8J6ZsOJHb4dtmi/Ak=
=77Je
-----END PGP SIGNATURE-----