Discussion What is your opinion on AI text models?

[jonoldak]

Weird question, of course, considering that most people on this forum despise AI slop. But I'm really curious to know how you feel about AI for texting purposes?? (not image AI)

And… A rather personal question... Do you flirt, or have you ever flirted with AI to vent your feelings? (I'm just a fan of role-playing with AI, I have some experience with Sillytavern, lol.)

 

[Scythe]

Eh, they exist. Doesn't bother me, I use the LLMs to do code for me sometimes. That's cool cause otherwise I would be stuck with shit code.

They are also good for searching shit up when you can't find things, always fact check but it's a good place to start.

Never used them for role-play purposes, and don't vent to them they steal ur info and offer what I assume is the plain old copy paste it gets better

I do admit I've thought about getting AI to dm DND cause I've got nobody to play with. But I heard AI has trouble with long term memory and sometimes just don't work.

What's your experience for role-play like?

 

[jonoldak]

Eh, they exist. Doesn't bother me, I use the LLMs to do code for me sometimes. That's cool cause otherwise I would be stuck with shit code.

They are also good for searching shit up when you can't find things, always fact check but it's a good place to start.

Never used them for role-play purposes, and don't vent to them they steal ur info and offer what I assume is the plain old copy paste it gets better

I do admit I've thought about getting AI to dm DND cause I've got nobody to play with. But I heard AI has trouble with long term memory and sometimes just don't work.

What's your experience for role-play like?

Hello, I responded a bit late, but I was just sleeping, heh.

You know, AI is actually very good at role-playing, but there is one significant caveat: you need a LOT of settings and a good prompt to achieve the desired result. Without them, AI often gives answers that aren't really good (depending on the model, if it's not RP, then that's the case, but if it is RP, then at least it will save the text a little). Sillytavern was created for this purpose, so that advanced users can customize the AI.

Usually, the best models for role-playing, less affected by Ai-Slop and good at maintaining context, are: Deepseek R1 0528 (my main model), Qwen3 235B A22B 2507, and the local Moistral-11B on HugginFace (not very smart, released in 2024, but writes surprisingly well, even though the PC is a bit weak :p).

Right now, I'm using Openrouter, and yes, it's better not to write personal data there, since those are server models, not the ones you can run locally on your PC (Koboldccp).

 

[tooBadTooLate]

My hate for those little fucking things made me convinced that I should go outside and socialize like a human.

 

[jonoldak]

My hate for those little fucking things made me convinced that I should go outside and socialize like a human.

Oh, I understand, AI censorship filters. They kick in when you write something that AI shouldn't respond to. But I have a whole table of AI models for censorship. You probably used chatgpt or claude (they can really write clichéd moral manuals). I'll give you a link to the censorship benchmark: AI Benchmarks

Of course, I sound a little technical, but I really want to share the reason why AI behaves like crazy pro-lifers. (Important note: Higher nsfw does not mean that the AI will behave like the most sincere AI. It simply means that the AI covers taboo topics to that extent. Also, don't use the web versions of AI, as they are guaranteed to have filters. APIs are better, as they won't have as many restrictive filters, unless you look at the policies of the companies and providers that release the APIs.)

 

[Scythe]

Hello, I responded a bit late, but I was just sleeping, heh.

You know, AI is actually very good at role-playing, but there is one significant caveat: you need a LOT of settings and a good prompt to achieve the desired result. Without them, AI often gives answers that aren't really good (depending on the model, if it's not RP, then that's the case, but if it is RP, then at least it will save the text a little). Sillytavern was created for this purpose, so that advanced users can customize the AI.

Usually, the best models for role-playing, less affected by Ai-Slop and good at maintaining context, are: Deepseek R1 0528 (my main model), Qwen3 235B A22B 2507, and the local Moistral-11B on HugginFace (not very smart, released in 2024, but writes surprisingly well, even though the PC is a bit weak :p).

Right now, I'm using Openrouter, and yes, it's better not to write personal data there, since those are server models, not the ones you can run locally on your PC (Koboldccp).

Don't worry about the late response lol, things on SS are slow.

What of Grok? Heard Grok is supposed to be more human caused trained on twitter and kinda more meant for fun stuff than Deepseek which is a knowledge based model. But I have seen screenshots of Deepseek making side comments in a suspiciously human manner and I have seen it act human during my uses as well.

How do you deal with the exposition AI likes to do and the memory problems. AI likely would forget my inventory and clues I've found if I tried to make it act as a DM.

 

[jonoldak]

Don't worry about the late response lol, things on SS are slow.

What of Grok? Heard Grok is supposed to be more human caused trained on twitter and kinda more meant for fun stuff than Deepseek which is a knowledge based model. But I have seen screenshots of Deepseek making side comments in a suspiciously human manner and I have seen it act human during my uses as well.

How do you deal with the exposition AI likes to do and the memory problems. AI likely would forget my inventory and clues I've found if I tried to make it act as a DM.

I don't use Grok for role-playing because he doesn't perform system prompts well, gets distracted, and is too expressive. But yes, he's pretty good for therapy, so much so that I have a chat room gathering dust where I vented all my emotions.

I have a good impression of Deepseek, although the latest models, such as 3.2, are more formal, dry, and a little boring for me, unlike R1 0528. Of course, R1 0528 has its own quirks. Sometimes it acts like a child and may not listen to system prompts. So you need to babysit it a little and be patient.

And yes, it can write suspiciously human-like responses, probably due to the cleaning of the dataset from Ai-slop and the gradual improvement of the AI architecture, which makes AI better in terms of empathy and possible consciousness. Still, modern models are still learning from old AI data (how ironic). So, Deepseek developers are probably focused on improving the dataset.

Regarding the AI memory decision, I don't recall this setting in Sillytavern, but there is a "scan_messages" setting in the lorebooks tab. As I recall, this controls message scanning, which means the AI takes into account the details of the last message.

I can't write in private messages because I'm new here q-q

 

[THEREALSLIMSHADY]

Texting with machine makes me wanna ctb even more

 

[tooBadTooLate]

Oh, I understand, AI censorship filters. They kick in when you write something that AI shouldn't respond to. But I have a whole table of AI models for censorship. You probably used chatgpt or claude (they can really write clichéd moral manuals). I'll give you a link to the censorship benchmark: AI Benchmarks

Of course, I sound a little technical, but I really want to share the reason why AI behaves like crazy pro-lifers. (Important note: Higher nsfw does not mean that the AI will behave like the most sincere AI. It simply means that the AI covers taboo topics to that extent. Also, don't use the web versions of AI, as they are guaranteed to have filters. APIs are better, as they won't have as many restrictive filters, unless you look at the policies of the companies and providers that release the APIs.)

I've kicked my urges to CtB a long time ago (going outside and getting involved with Strangers and Freaks winkwink helped a lot!), but moderately speaking, I can understand why they'd have to censor that. It's better for business, considering the AI companies, not to get too involved in such touchy matters and point them to a mental health professional... if the safeguards work. Their coding has been getting more and more sloppy the more they try to appease their shareholders.