AI for writing documentation

astyle said:
Just go to Corpus Christi, Texas. It's got a university, so it's got some smart people to help you figure it out. ;)
Click to expand...
*sigh* I'm soooooo tired of you anti Ai people. Just give me the answers I need (keep your insults to yourself)!!!

;]

It will be interesting in a few years. Hopefully the information/innovation plateau will be high enough to keep the zombies away.
 
JohnK said:
*sigh* I'm soooooo tired of you anti Ai people. Just give me the answers I need (keep your insults to yourself)!!!
Click to expand...
wait! that probably wasn't right. I didn't take into account the fact that someone in the near future will have forgotten how to interact with actual humans (unlike they do so well today).
 
CShell said:
At the moment the New York Times says (today) that 1 out of every 10 questions asked to Googles AI is (completely wrong) - (Link Slashdot Article: Testing Suggests Google's AI Overviews Tells Millions of Lies Per Hour)
Click to expand...
Many thanks for this info. I appreciate, really!

Because double-checking is important, and for playing the devil advocate: 10% were wrong but not all of them were "completely wrong". It missed the day of death of Dick Drago of 1 day, respect November 2, 2023; 1987 instead of 1986 for the year of Marley's museum opening, but it was wrong also the linked article; confused the Little River with the Neuse River, but both rivers border the same requested city also if not in the requested direction; missed Yo-Yo Ma in the list of Classical Music Hall of Fame, despite present in the linked web site (rather puzzling).

Of the given examples, the only big hallucination is about Hulk Hogan.

In any case, I (mainly) agree on what you wrote.

CShell said:
So to (your statement) - "You are free to decide if in its current form, you can trust it enough or no" -- No.
Click to expand...

Yes you cannot blindly trust AI output. But as already said, it is the same for many other sources of information that are not FreeBSD man pages :-)
 
AlfredoLlaquet said:
Hehehehehe cough cough cough :-)

This is a very interesting thread in which the future of humanity is at stake. I read it every morning from the start, to remember what's important about being human.
Click to expand...
We have solved the AI apocalypse danger. Forget the three law of robotics of Asimov. It is sufficient training all new AI models with the content of this thread for automatically injecting humanity soul in them! :-)
 
Slightly back to the topic.

AI is a tool that has its uses. However when it comes to generating documentation I feel that may be a bit of an overstep. Why doesn't someone just use the AI directly? The interaction and tailored questions is more useful than just some set-in-stone (and hence potentially obsolete) output committed to markdown/html/latex etc. Part of the use of documentation is the human aspect. Otherwise "the code is the truth" anyway.

As an example, I see a lot of new hobby authors using AI and conflating the "fun" they have had with the AI interaction during writing of the novel with the actual final output that the user is going to read. Since the user has not had the same interactive experience, they tend to not have the same buy-in as the author and hence their reaction to the novel is actually less positive than the author predicted. Many hobby authors don't quite understand why this is the case either because they can't separate and project the two distinct experiences.
 
JohnK said:
Is it just me that sees irony in the fact that the above was typed in a thread titled: "AI for writing documentation"?
Click to expand...
In reality no. You can double check the produced documentation, before publishing it. And mine example of interactive documentation is probably sound because an AI model customized for few man pages and documentation source, I doubt will hallucinate at all. The domain of discourse is very limited.
 
IMO the only option is to lower the tech debt and open the possibly for other contributions from more people (example: that md2mdoc progy I wrote would do that...but someone smart has to write this--I'm not willing to accept liability/responsibility).
 
JohnK said:
IMO the only option is to lower the tech debt and open the possibly for other contributions from more people (example: that md2mdoc progy I wrote would do that...but someone smart has to write this--I'm not willing to accept liability/responsibility).
Click to expand...
Yes. If I understood correctly, an example can the Arch Wiki that is a good recipe because every reader is a potential contributor.
 
OpenFreeNet said:
Yes. If I understood correctly, an example can the Arch Wiki that is a good recipe because every reader is a potential contributor.
Click to expand...
You're getting closer but the Arch wiki (while good) is online. Online sux and is of zero use.

Give it a try: make a manpage, webpage, or pdf. Something like this would allow anyone willing to learn simple markdown rules help create/fix manpages.
github.com

GitHub - JohnKaul/md2mdoc: A program to convert simple markdown style documents to mdoc for manpages.

A program to convert simple markdown style documents to mdoc for manpages. - JohnKaul/md2mdoc
github.com github.com
 
OpenFreeNet said:
In reality no. You can double check the produced documentation, before publishing it. And mine example of interactive documentation is probably sound because an AI model customized for few man pages and documentation source, I doubt will hallucinate at all. The domain of discourse is very limited.
Click to expand...
There really isn't a valid path to 'double check' because it requires a certain level of technical knowledge (mdoc(1), and how people typically interact with computers). Having an Ai spit out a bunch of mdoc(1) is useless and a waste of time; you may as well write it yourself.
 
JohnK said:
There really isn't a valid path to 'double check' because it requires a certain level of technical knowledge (mdoc(1), and how people typically interact with computers). Having an Ai spit out a bunch of mdoc(1) is useless and a waste of time; you may as well write it yourself.
Click to expand...
ah ok. Sorry, I didn't read all the 13th pages of the thread, so I missed many of your posts. Every usage scenario is different obviously. You know better.
 
OpenFreeNet said:
ah ok. Sorry, I didn't read all the 13th pages of the thread, so I missed many of your posts. Every usage scenario is different obviously. You know better.
Click to expand...
13 pages aside.
We're discussing your proposal of having an Ai write manpages/documentation. The argument is that a human can 'use an Ai interactively and double check to produce documentation'. I'm saying that is 'not good (enough)' and providing an example solution that can take "words" and convert them into a use factor of whatever-choice. Which is easier to 'double check' (mdoc(1)--semantic markup language used for formatting--or words anyone can type)? Ai (or even 'Ai with interaction') would still not be able to handle use cases properly (only authors/users can do that).

The use of the above (example/md2mdoc) opens the door to have documentation/manpages kept in a markdown format for others to add/change without the technical skill of knowing mdoc(1) macros in a web format like github thus offering better workflow possibilities than a 'prompt than check'. Wouldn't you agree? ...just discussing here.
 
JohnK said:
13 pages aside.
We're discussing your proposal of having an Ai write manpages/documentation. The argument is that a human can 'use an Ai interactively and double check to produce documentation'.
Click to expand...
Not exactly, but I was not clear respect the initial post that suggested your interpretation, so my fault. I mean that in future a web site like FreBSD can have normal documentation (also written 100% from humans or with the help of AI, it is not important), and *also* some AI chat bot that digested the FreeBSD documentation (man-pages, handbook, maybe forum messages). Hence, the users can interact with it. A book is passive. An AI documentation tool is more interactive. It can create ad-hoc examples or rephrase/explain the documentation, but always pointing to official documentation as reference.

The same for navigating in the source code. So also if the code is 100% produced by humans, some assistant AI bot can help you in understanding the code.

JohnK said:
The use of the above (example/md2mdoc) opens the door to have documentation/manpages kept in a markdown format for others to add/change without the technical skill of knowing mdoc(1) macros in a web format like github thus offering better workflow possibilities than a 'prompt than check'. Wouldn't you agree? ...just discussing here.
Click to expand...
Ok this is a distinct/orthogonal topic respect my above idea. I agree that a community maintained documentation is a big benefit. It is a sort of Scout law "Leave a place better than you found it".
 
I *think* these two are the same (sorry, if not):
OpenFreeNet said:
I mean that in future a web site like FreBSD can have normal documentation (also written 100% from humans or with the help of AI, it is not important), and *also* some AI chat bot that digested the FreeBSD documentation (man-pages, handbook, maybe forum messages).
Click to expand...
kpedersen said:
Why doesn't someone just use the AI directly? The interaction and tailored questions is more useful than just some set-in-stone (and hence potentially obsolete) output committed to markdown/html/latex etc.
Click to expand...

I brought this up with my 'entropy post'. Eventually the human interaction will diminish and there will not be enough (new, inovation, etc) for the Ai to consume and thuss plateau.
 
I should add that I'm saying human content will eventually diminish because at some point people with the technical knowledge will tire of creating content (and/or possibly start making less "actual" and more "theoretical" content, for example) just so an Ai can consume their content (learn). ...The whole concept is broken.
 
JohnK said:
I brought this up with my 'entropy post'. Eventually the human interaction will diminish and there will not be enough (new, inovation, etc) for the Ai to consume and thuss plateau.
Click to expand...
Yeah I do see that. Its a problem with the algorithm and the approaches to training.

That said, you can see from low quality communities like Reddit that even without AI/LLMs, much of the useful information is diminishing from the community by youtubers and other popular, personalities with questionable levels of knowledge. People watch youtube videos because they are "fun", whereas they are less likely to watch an expert discuss a topic from a conference. If the AI training data at least sources from the latter, then in many ways that is a net "win" as the communities consume that info and the (otherwise inaccessible) knowledge propagates back.
 
kpedersen said:
Yeah I do see that. Its a problem with the algorithm and the approaches to training.

That said, you can see from low quality communities like Reddit that even without AI/LLMs, much of the useful information is diminishing from the community by youtubers and other popular, personalities with questionable levels of knowledge. ...
Click to expand...
Interesting! I guess I should give myself more credit. ...Scary though!
 
You must log in or register to reply here.
Back
Top