Maybe Debian could create its own LLM trained on Debian for Debian? cuts out the licensing issues and keeps the community in control?
That’s… not how it works, unfortunately.
I think what you’re getting at would be strict criteria along the lines of:
A human must take full responsibility for any contribution.
And only with fully disclosed, labeled assistance from “legally unproblematic” LLMs. The Nemotron or Olmo series would be examples: apache licensed, open and auditable datasets, can be fully controlled by Debian community.
They could enforce some specific harness, or mathematically verifiable output, like some other projects already do. I know less about the technical details of that, though.
That’s… not how it works, unfortunately.
I think what you’re getting at would be strict criteria along the lines of:
They could enforce some specific harness, or mathematically verifiable output, like some other projects already do. I know less about the technical details of that, though.