you are viewing a single comment's thread
view the rest of the comments
[–] 2 points 2 weeks ago (6 children)

Why not have the LLM copyedit it, too?

AI's don't like working with really large amounts of text. They will not go through a text linr by line and analyze it, not without having a highly specialised "AI" or harness or whatever you call it.

If you simply ask it to remove ending paragraphs like that one in the OP, it will write a script that tries to find phrases like "if you want" and then check the result of the script and replace those. It will NOT read the whole book line by line to find instances like that.

  • source
  • parent
  • hideshow 6 child comments
  • [–] 6 points 2 weeks ago (3 children)

    It's possible, but a pain. I have an html python portal for a self hosted set of LLMs where I can upload a file of any length, and it pulls it apart maybe 400 words at a time to proofread, copyedit, or translate, then stitches it back together. It does mostly OK. But it doesn't look at things like plot points or consistency or characters. Instead, I could do 1 chapter at a time into Claude, compress it to a two-page summary, and get then have it analyze the summaries.

    Which sounds like a great way to get terrible results and burn more tokens than I need to pay two humans to her the info I need.

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

    It’s really hard to avoid sycophancy when proofing more than a paragraph at a time, anyway. They will not set you straight.

  • source
  • parent
  • hideshow 1 child comment
  • [–] 1 point 2 weeks ago

    That's the benefit of the small context window and self hosting. It proofs 3 or 4 sentences at a time looking for spanning and grammar issues based on the kind of work. Plus I'm using Qwen and Hermes, you just tell them to be harsh and they do.

  • source
  • parent