Login
You're viewing the functional.cafe public feed.

Replies

  • 💬 0🔄 0⭐ 0
  • 💬 1🔄 0⭐ 0
  • 💬 0🔄 0⭐ 0
  • 💬 1🔄 0⭐ 1
  • Nov 9, 2026, 8:08 AM

    @bignose I 100% agree with the mail by @GyrosGeier but I already hear the responses going along the lines of:

    * the LLM is *not* a dependency because the DFSG team can drop it at any time, it is just optional and (as was pointed out) not even every DFSG team member uses the output (so a lot of them are even thrown away wtf??)
    * we bet on computers becoming faster, special purpose hardware becoming cheaper and then everybody can run a (still proprietary) open-weight model at home

    💬 1🔄 0⭐ 0
  • Nov 9, 2026, 8:46 AM

    @josch @bignose my mail is specifically about the suggestion to take on more work because LLMs enable that.

    If we hypothetically accept that as true, then additional work taken on as a result turn LLM use into a hard dependency.

    That is a separate discussion from the DFSG team discussion though.

    The DFSG team can drop their use at any time, and my expectation is that they will do so if it creates more work for them than it takes over.

    LLMs are good at externalizing work: you get something half-finished that you can then give to someone else for refinement.

    If I create the copyright file with an LLM, and hand it off to the DFSG team for refinement, that saves me a lot of work, but gives it to the DFSG team. That's a dick move, so I don't do that.

    Likewise, if the DFSG team creates an initial review with an LLM and sends that to me for refinement, they are offloading their work to me, which would also be a dick move.

    We can always have a debate about changing the division of the labor, and setting up workflows that reflect that, but subverting the workflow by inserting extra, invisible, steps, that is just bad management.

    💬 0🔄 2⭐ 1
  • 💬 0🔄 0⭐ 0