Login
You're viewing the discuss.systems public feed.
  • Þommythomasjwebb@mastodon.social
    Oct 8, 2026, 10:03 PM

    @cwebber another thought I just had is what if someone runs their agent in yolo mode (don't ask permission before running whatever tools) and it ends up "going rogue" and verbatim copying code from a repo (or a local copy of a client's proprietary code)? Sometimes the commands it runs modify the code (such as search and replace stuff or scaffolds). I'm sure someone thought of this before, but I'm not even sure that training data being sufficiently atomized is enough to prevent liability.

    💬 1🔄 0⭐ 4

Replies

  • Jeff GriggJeffGrigg@mastodon.social
    Oct 10, 2026, 10:29 AM

    @thomasjwebb @cwebber

    I keep thinking it misguiding to say that training an LLM (or similar) on copyrighted work is a violation. It seems to me that the violation is in the *use* of the "model" to create infringing works. And the rulings' reasoning, that I've read, seems to be on the *use* being the justification for finding the violation.

    Like, memorizing someone else's song is not a violation. Performing it for others, in a way that completes with the original, is the violation.

    💬 1🔄 0⭐ 1
  • 💬 0🔄 0⭐ 0