2025-12-22 00:09 UTC
One adjacent example of intentional miscommunication: the word "open" as applied to AI. An LLM cannot be "open source" unless you have access to the training data and methodologies. "Open weights" is a craven attempt to misdirect the word "open" — we might as well say that bunch of machine code is "open" because you can disassemble it.
LLMs can only be trained at extremely disproportionate scale, such that they cannot be "open" in any meaningful sense.
Replies (1)
-
@gbargoud@masto.nyc 2026-01-01 23:05
@xgranade@wandering.shop I remember that last year, there was a proposal to not consider a model compliant with the Debian Free Software Guidelines unless all training data was included. It looks like the proposal was withdrawn though. I wonder if there will be another attempt at codifying this: https://www.debian.org/vote/2025/vote_002