Comment by b112
18 hours ago
You're precluding the possibility that the other party instructed the LLM to write what he intended to say.
That is not currently possible. At best, you could train on millions of personal examples of style, eg historical emails, reports, etc, and you'd still never get what the person would have written.
I'd be astonished if an LLM didn't even miss important nuance, data, and focus on the wrong point.
It think the parent's point was, hearing a summary of what someone was thinking, from a second party isn't interesting as data from the primary source.
When you come to listen to Bob the Great, having a minion walk put and paraphrase Bob's words isnt the same. Human, LLM, or not.
I will use LLMs to find spelling and grammar mistakes, as well as voice inconsistencies. I then carefully review the diff.
Yes there are deterministic tools for this, but grammar is actually kind of complicated and the tools either miss mistakes, or suggest obviously worse edits. Looking at you, Microsoft Word.
All the writing is my own and I read all of it. Naturally there’s no LLM-isms because I wrote 99.9% of the original text. But, it was AI assisted. So, I don’t think it’s cut and dry.
Many do as you, and I completely endorse this usage. Unfortunately, too many of us here make the giant leap of "If it went through the LLM, the original person was too lazy."
Of course, the start of the thread really had nothing to do with the concept you're espousing here. At all. It was more about "Some guy writes instructions into an LLM, and gets a blog entry", which is entirely on a different scale. And yes, that's the scope and scale we've been discussing.
Personally, in this new, different context you've stated, I don't think many have an issue with that. If you write 6 pages of text, and then ask an LLM to look for errors, that's not remotely the same, and it is more like a spell checker, or a second set of eyes.
> At best, you could train on millions of personal examples of style, eg historical emails, reports, etc, and you'd still never get what the person would have written.
You wouldn't get it verbatim, but your mistrust in AI reflecting what you intended to say is unwarranted.
I use LLMs to clean up poorly OCR'd hand written notes. It's the exception where I need to "fix" its output. It figures out what the notes intended, and writes it up fairly well.
> It think the parent's point was, hearing a summary of what someone was thinking, from a second party isn't interesting as data from the primary source.
For a lot (most?) of AI slop, I agree. But pre-LLMs, this was universally a bad idea. The first draft always sucked, and was easy to misunderstand. A good LLM, with a good system prompt, will usually produce something better than the input it is fed. Most people are poor at expressing their thoughts.
> When you come to listen to Bob the Great, having a minion walk put and paraphrase Bob's words isnt the same. Human, LLM, or not.
Most people aren't Bob The Great, but Bob The Incoherent.