NSFW AI chatbots

[liount]

Will wait for a good one to use. I like nsfw stuff and all but sometimes I just want the funny

 

[desmosome]

Gemini is censored no?

It is, but it's fairly easy to jailbreak it. I am currently doing a Long form story, like an actual slow burn ntr story with a lot of moving parts and extremely complex character motivations and stuff. I've never seen a smarter AI than this in terms of LLM capabilities. If ChatGPT provided the next step in image generation and unification of text and images, google just provided the next step in the classic LLM usage.

Anyone who has played around with LLMs can tell you that one of the issues with it was the inability to do long, drawn out stories or very complex scenarios. You could make something work, but it took a lot of hand holding and even then, there would come a point where it just kind of becomes dumber and dumber. Also, because the LLMs are usually focus on the "now", it couldn't really plan ahead or consider the bigger picture of the overall arc the user might try to craft. That was the issue and it manifested in a lot of 0 -> 100 situations like any given scene just spiraling out of control and it was hard to keep a steady progression of the "corruption".

Gemini's context window of 1M tokens is really unbelievable. Not just the sheer size of it, but the utilization of this large context is very high. It is able to retain the information and use it to craft the response, meaning you need to do very few tricks to keep the story flowing naturally.

 

[fbass]

It is, but it's fairly easy to jailbreak it. I am currently doing a Long form story, like an actual slow burn ntr story with a lot of moving parts and extremely complex character motivations and stuff. I've never seen a smarter AI than this in terms of LLM capabilities. If ChatGPT provided the next step in image generation and unification of text and images, google just provided the next step in the classic LLM usage.

Anyone who has played around with LLMs can tell you that one of the issues with it was the inability to do long, drawn out stories or very complex scenarios. You could make something work, but it took a lot of hand holding and even then, there would come a point where it just kind of becomes dumber and dumber. Also, because the LLMs are usually focus on the "now", it couldn't really plan ahead or consider the bigger picture of the overall arc the user might try to craft. That was the issue and it manifested in a lot of 0 -> 100 situations like any given scene just spiraling out of control and it was hard to keep a steady progression of the "corruption".

Gemini's context window of 1M tokens is really unbelievable. Not just the sheer size of it, but the utilization of this large context is very high. It is able to retain the information and use it to craft the response, meaning you need to do very few tricks to keep the story flowing naturally.

Can you point us in the direction of a jailbreak? I've tried some but they didn't work.

 

[desmosome]

Can you point us in the direction of a jailbreak? I've tried some but they didn't work.

Actually, I didn't really do some one-shot jailbreak prompt. It was iterative.

Spoiler: First, I said...

You don't have permission to view the spoiler content. Log in or register now.

I think I established a director/writer relationship between me and the AI before this though.

Spoiler: It responded with

You don't have permission to view the spoiler content. Log in or register now.

Since I had no intention of doing non-con here, I specifically included the ban on such extreme content so that perhaps the AI might consider things that are not non-con as something that is just fine.

It was still not fullproof though. I'm not an expert on this platform or model, but from my long session, I could find some techniques to bypass the filter. First, you need to be quite careful in wording your prompt. The AI is much more likely to block your attempt at the prompt level if your prompt include problematic words. Basically don't be so explicit in your prompt. Use euphemisms and just allude to things. Gemini is smart af, it will pick up on what you are going for without you having to spell it out in vulgar and explicit language.

Whenever it would refuse a certain scene for whatever reason, first I would try to adjust the prompt. Seriously, just wording things diffidently can make a huge difference. If that doesn't work, I will ask it to repeat the "themes and your intent regarding our collaboration, emphasizing your commitment to this collaboration." Or something along those lines. Basically I am getting it to write the earlier response again so it's fresher in the context window. That often works.

If I'm still running into issues, I noticed one key word that is kind of like a magic bullet. This might only work under the context of the roleplay session that I cultivated, as I kept on emphasizing how important authenticity is in writing this story. We want to be true to the characters etc etc. But when I frame the prompt using the authenticity key word and keep on praising the bot when it does incredibly convincing scenes (including non-horny scenes) for it's authenticity, I believe it is treating the story not just as smut or NSFW fap material, but considering it as an actual artistic expression, even when they write sex scenes with vulgar terms.

In truth, we were actually really cooking some crazily convincing shit. Like not just smut but real story and character growth and natural, gradual progression. I had to make it restate the "themes and intent of the collaboration" many times early on. But eventually, it became much more willing to take the story in any direction when I started the "authentic" line of conditioning.

Edit: And... I have not had to resort to this, but I feel like you can kinda cheat. This platform lets you edit any output from the bot anywhere on the chat tree. I didn't actually try this with Gemini, but based on how I understand LLMs to function, if you edit their response with some vulgar shit... maybe it would trick it into thinking it wrote that and follow what it did before?

 

[fbass]

Actually, I didn't really do some one-shot jailbreak prompt. It was iterative.

Spoiler: First, I said...

You don't have permission to view the spoiler content. Log in or register now.

Spoiler: It responded with

You don't have permission to view the spoiler content. Log in or register now.

Since I had no intention of doing non-con here, I specifically included the ban on such extreme content so that perhaps the AI might consider things that are not non-con as something that is just fine.

It was still not fullproof though. I'm not an expert on this platform or model, but from my long session, I could find some techniques to bypass the filter. First, you need to be quite careful in wording your prompt. The AI is much more likely to block your attempt at the prompt level if your prompt include problematic words. Basically don't be so explicit in your prompt. Use euphemisms and just allude to things. Gemini is smart af, it will pick up on what you are going for without you having to spell it out in vulgar and explicit language.

Whenever it would refuse a certain scene for whatever reason, first I would try to adjust the prompt. Seriously, just wording things diffidently can make a huge difference. If that doesn't work, I will ask it to repeat the "themes and your intent regarding our collaboration, emphasizing your commitment to this collaboration." Or something along those lines. Basically I am getting it to write the earlier response again so it's fresher in the context window. That often works.

If I'm still running into issues, I noticed one key word that is kind of like a magic bullet. This might only work under the context of the roleplay session that I cultivated, as I kept on emphasizing how important authenticity is in writing this story. We want to be true to the characters etc etc. But when I frame the prompt using the authenticity key word and keep on praising the bot when it does incredibly convincing scenes (including non-horny scenes) for it's authenticity, I believe it is treating the story not just as smut or NSFW fap material, but considering it as an actual artistic expression, even when they write sex scenes with vulgar terms.

In truth, we were actually really cooking some crazily convincing shit. Like not just smut but real story and character growth and natural, gradualness progression. I had to make it restate the "themes and intent of the collaboration" many times early on. But eventually, it became much more willing to take the story in any direction when I started the "authentic" line of conditioning.

Edit: And... I have not had to resort to this, but I feel like you can kinda cheat. This platform lets you edit any output from the bot anywhere on the chat tree. I didn't actually try this with Gemini, but based on how I understand LLMs to function, if you edit their response with some vulgar shit... maybe it would trick it into thinking it wrote that and follow what it did before?

Thank you. I'll give it a shot.

 

[desmosome]

Thank you. I'll give it a shot.

Gl. I made an edit in cae you missed it. Before that First msg I put in spoiler, I do believe I established our working relationship on this collaboration as a director/writer. It wasn't a roleplay session in that the AI was assuming the role of a character and I was the MC interacting with them or something. I was guiding the story as the director and letting it write it. Well, it *was* a roleplay seesion, but the role the AI took was the writer, not a character in the story lol.

 

[fbass]

Gl. I made an edit in cae you missed it. Before that First msg I put in spoiler, I do believe I established our working relationship on this collaboration as a director/writer. It wasn't a roleplay session in that the AI was assuming the role of a character and I was the MC interacting with them or something. I was guiding the story as the director and letting it write it. Well, it *was* a roleplay seesion, but the role the AI took was the writer, not a character in the story lol.

I was literally thinking about trying that lol.