it is possible to use chat gpt for create dialogs (1 Viewer)

[deru]

if someone take the "mindbreak" dialog from badinsomnia, can we extend it to 10 000 lines or even more ? maybe mixing real human writing with a chatgpt basis or something like this ?

chatgpt know how to code and know writing, and it's possible to get around the interdiction of "offensiv content".

 

[deru]

imagine a text with 50 000 lines, it would be like a real ia

 

[Wrecksler]

I am not sure about chatgpt but story generation is possible, take a look at AIDungeon2. I am also working on a "Fatality generator" as I call it, trained on kinky fight roleplay logs from my f-list bot currently residing at UFL MAIN PAGE

Current version is using simple markovify algorithm, but its just a proof of concept, I am researching better ways if doing it in the little spare time I have on it.

That said, I dont think ChatGPT will do anything kinky, it seems to have some hardcoded filters for such things, to not be offensive and whatnot.

 

[deru]

i'm a shit with the technical stuff, i was just thinking that there is a lot of AI who can "learn" what to do and one of them can maybe do the ultimate dialogue with all the dialogue code needed, or maybe it can deliver a good basis for humans, like if we do a story with a lolipop who squirt good juice if you suck it or something like this for "evade" the offensiv filters.

then you use another tool for replace all the "lolipop" by the word "dick".

maybe we can do it 1 thing after 1 other and do a compilation of all what we get.

first 1000 lines just for the "sucking", another for the "gag", another for the "drool" and with everything.

the only problem if it work is if there is a limit of text for sdt

 

[Wrecksler]

I, unfortunately, know practically nothing about sdt.

But I know a bit about AIs, programming and machine learning. And I don't know of any off-the-shelf AI that you can teach something by just dumping stuff at it. Usually it all involves custom code, data preparation, and other stuff, a fair amount of work.

So yes it can probably be done by someone with some data science knowledge, but no - there are no tools available for anyone who does not know programming at all, not UI tools to train things like that, etc. If there is - I don't know about them. If someone knows more - please correct me.

 

[Graceus]

potentially possible. tried with A Haunting Succxperience dialogue (tried to pick one that wouldn't immediately be recognized as sexual) and replaced "cum" with "ectoplasm" -- even with some careful directions to keep it SFW, the content warning gets triggered. not sure how sexy the end result is, and you'd still have to do some manual adjustments. Difficult to get 50000 useful lines following this process.

 

[BadInsomnia]

if someone take the "mindbreak" dialog from badinsomnia, can we extend it to 10 000 lines or even more ? maybe mixing real human writing with a chatgpt basis or something like this ?

chatgpt know how to code and know writing, and it's possible to get around the interdiction of "offensiv content".

Bro's trying to put me out of work xD
Just kidding, it would be an insane upgrade though. Considering how old SDT is, at this point you might be better off coding a new game entirely.

 

[ZipMip]

Bumping this. See my thread.

https://www.undertow.club/threads/it-should-be-possible-to-utilize-large-language-models-llms-ai-locally-to-enhance-dialogue-in-sdt.19835/

Ooba's textgen UI --> Sillytavern API

(or, directly into the Ooba API)

You can run local models uncensored, and get exactly what you want for super deepthroat. With Sillytavern, it can remember context of what is happening in the scene, and it's possible to give a scenario, among other things.

The main work would be a SDT-engine that is:

1) Time-based, meaning, a time-line that would make a new API request every, i.e., 10 seconds. Timer would be paused when generating and awaiting the result from the API, or, it could stream the tokens back to the application as they're generated
2) Would enable voice-creation via an online voice engine, or, a locally hosted voice model
3) Could be trained on SDT actions including Dialogue actions, to make the scene change dynamically, assuming creators had a way to provide some sort of "repository" that could be quickly queried and chosen from, whether by AI or by the scene.