all 3 comments

sorted by: hot top controversial new old
[–] 2 points 3 weeks ago (2 children)

Shit like this:

Every layer that reports back says yes. Only the inbox says no.

Is so fucking annoying. Why does it write this way? Is it trained to be a jackass, or does that just come naturally through the process?

  • source
  • hideshow 2 child comments
  • [–] 3 points 3 weeks ago (1 child)

    I suspect it is model collapse. Too much synthetic data to steer it into useful activities. I am struggling with this on a personal project at the moment where I am trying to rip out all of the LLM-ese. I have a 31 point list of faults and an entire rubric for the clanker to follow in an attempt to mitigate the LLM-ese register. It still fails on the regular.

  • source
  • parent
  • hideshow 1 child comment