Honestly, I don't understand what you're talking about. LLMs excel at rephrasing things or asking general questions like what idioms fit in a certain situation. It's actually one of the most basic, boilerplate things they are used for (part of the reason they are so used for mails).
You shouldn't trust them with
facts, which is what you gave as an example and I agree, but it's ridiculous to take that as an example of not using them as the tool they are designed as. You should always be the barrier/director that stands in place and says if it's good or not.
On a more general sense... AI is going to be shit at giving you an actual finished story (at least with current models and in my opinion). That said, this doesn't really mean you shouldn't use it as a tool for one. Having issues with coherency is fine from time to time, you should be the one checking that particular aspect.
Some of the things mentioned in this thread in general generalize so much that are borderline useless/inaccurate. You need to take into account the model you're using (huge difference between opensource 7B model, GPT or Claude/Gemini) and how you are actually using the thing (like in image gen, you can use a good model to gen slop). The part of mixing and matching things in fashion is just wrong though. Maybe if you just used "write a Naruto fanfic about an Uzumaki orphan" rather than actually getting good at using a new tool.
You can have it write a first draft following certain guidelines/style, and then you edit. Or have it write the outline that you edit and then you go for a first draft. Or give it a book on writing, summarize it as general things to follow and then use it to nitpick your own writing to death. Or brainstorm things for worldbuilding or so you get unstuck on a turn of phrase... and so on. There are plenty of people using it in a ton of different ways that aren't just Grammarly or posting the first answer the model gives.
Not really true in my experience, both in writing and elsewhere. Sometimes AI is pretty good at being an enhanced
rubber duck and it comes with pretty cool ideas from time to time. Especially because the big models are much more capable in different subjects than a single human can. What you create can and will change depending on your process. It has even happened with image gen when the AI does something slightly different and I have to admit it does look better than what I was thinking of.
Yeah, and most models have some degree of error correction built in and are capable of calling various APIs to validate real data and the models are refactored and the date put into them cherry picked.
The fact that some result is generated pseudorandomly doesn't mean that there aren't any guardrails and quality gates.
And if the answers are nondeterministic the results needen't be either.
There really isn't a perfect way to phrase a statement or describe some fictional prop.
It is not like a fanfic's technobabel for example needs to include working orbital mechanics for instance, Arthur Clarke might have done that for short stories but most writes are not him.
And the A.I. can give me a few ways to rephrase "...she stared at him with bedroom eyes..." To make the peose sound more interesting then I will use it from time to time.
Not being a native speaker it can just give me more ideas how to do stuff, not that I use the suggestions verbatim, anyways.
Here is a little example.
deepmind in action. said:
Prompt:
Rephrase the below to sound more formal, generate five alternative rephraisals: And a lot of models muse vsrious forms of checking ansnpeuning to improve results and narrow down proper responses.
Fix spelling errors and provide rationale for the generated outputs.
And a lot of models muse vsrious forms of checking ansnpeuning to improve results and narrow down proper responses.
Response:
Here are five formal rephrasings of your sentence, along with corrections for spelling errors and a rationale for each output:
Original (with corrections):
"And a lot of models use various forms of checking and pruning to improve results and narrow down proper responses."
Rephrased Versions:
"Many models employ various forms of verification and pruning to enhance results and refine response accuracy."
Rationale: Replaces informal terms ("a lot of" → "many," "use" → "employ") and clarifies intent ("narrow down proper responses" → "refine response accuracy").
"Numerous models utilize different validation and pruning techniques to optimize outcomes and ensure precise responses."
Rationale: Enhances formality ("a lot of" → "numerous," "improve results" → "optimize outcomes") and specifies "validation" for clarity.
"A significant number of models incorporate various checking and pruning mechanisms to boost performance and filter responses effectively."
Rationale: More precise phrasing ("a lot of" → "a significant number of") and stronger verbs ("improve" → "boost," "narrow down" → "filter").
"Multiple models leverage diverse verification and pruning strategies to enhance output quality and restrict responses to relevant ones."
Rationale: Uses "leverage" for sophistication, "diverse" for variety, and "restrict responses to relevant ones" for precision.
"Various models implement systematic checking and pruning methods to improve accuracy and select the most appropriate responses."
Rationale: "Systematic" adds rigor, while "select the most appropriate responses" is more formal than "narrow down proper responses."
Spelling Corrections & Rationale:
"muse" → "use" (Typo)
"vsrious" → "various" (Typo)
"ansnpeuning" → "and pruning" (Miskeyed letters)
These revisions improve professionalism, eliminate informal phrasing, and ensure technical accuracy. Let me know if you'd like further refinements!
New chat
AI-generated, for reference only
I think it is pretty good, even better at detecting typos and turning them into something actually workable, too.
Sure bests grammarly in that regard.
Edit: Oh, yeah, forgot to mention, this is a lot like the typewriter vs. Dedicated word processor vs. PC with ancient dos stuff like WordStar vs. actually modern office software.
I am pretty sure at least one scifi writer bitched about muh computors not being the same as his old typewriter.
Cough-cough, I think it was Harlan Ellison, cough-cough.
And a bunch of scifi writers still had to use this ancient thing called WordStar.
And among a particular subset of it people there is always the old Vi vs. EMACS vs. Nano
omit:
What often gets those old farts to bury the hatchet is hatered for actual modern IDEs and advanced editors like VS code.
Also, Vim is best, the other CLI/TUI ones are crap, there is a reason the first hit for EMACS is "how do I exit it", 2 key combos beats 3.
Seriously, A.I. can be just another tool and should be treated as just another tool.
But do not let it do your thinking for you.
