no comments (yet)

sorted by: hot top controversial new old
there doesn't seem to be anything here
6 comments in c/artificial_intel@lemmy.ml view that post
[–] 7 points 6 months ago* (4 children)

This is where commit based version control can help mitigate the damage or detect it early. The golden rule in AI assistance in programming (if you're going to use it) is to check the changes to make sure you understand everything before you commit it. If documents had a git-like version control system it would be easier to detect corruption early.

Or just use Markdown files in a git repo IDK

  • source
  • hideshow 4 child comments
  • [–] 3 points 6 months ago (1 child)

    That's not how humans work. We're lazy.

    You ever use snapshot tests? They're garbage. They're garbage because people glance at a 300 line diff and go "that seems right", because a good chunk of the time it is. But some portion of the time, it's borked up.

    Having it in git doesn't solve that problem.

  • source
  • parent
  • hideshow 1 child comment
  • [–] 2 points 6 months ago

    sounds like it sux!!

  • source
  • 3 comments in c/fuck_ai@lemmy.world view that post
    [–] 3 points 6 months ago (2 children)

    By the end of long workflows

    Yes, this has been known for 10 years.

  • source
  • hideshow 2 child comments
  • [–] [S] 0 points 6 months ago (1 child)

    By the end of long workflows

    Yes, this has been known for 10 years.

    huh? the kind of "long workflows" this paper is discussing didn't exist two years ago much less 10

  • source
  • parent
  • hideshow 1 child comment
  • [–] 1 point 6 months ago

    it doesn't matter. the principle is that if x is the length of your context window, then at 0.4x the chance of hallucinations start increasing exponentially. we're now at token windows of 1M, and all it does is shift that hallucination window further away, so the model 'feels' stronger because it takes longer before it hallucinates, but eventually it always does.

  • source
  • parent