NSFW AI chatbots

[n3otidus]

i found The_Naughty_Novelist on poe, work very good, sometime get me error but if you refresh message he continue with topic

 

[deepsauce47]

I'll say it again:

Chub AI -> Mars Tier -> Mixtral 7B (or Asha 70B) -> "A decent Mercury" preset

...is a god-tier setup.

This chatbot fapping stuff is very sensitive to the quality of all the inputs: the bot itself, the LLM, and the preset instructions. All of these have to be excellent for the best results. I would encourage anyone who hasn't tried a truly good setup to do so.

I haven't played a single F95 game since discovering chatbots a few months ago.

 

[Astcit]

Giving my two cents: KoboldCPP (I use the

You must be registered to see the links

, 'cause AMD), using Tiger Gemma 9B v1a Q8 or Smegmma 9B v1h Q8 (16GB of VRAM so full GPU offload, 45 layers), and Adventure mode.

After making a little world building and describing my character, I can play pretty wild stories, with my darkest desires. The AI make descriptions and say what are the consequences to my char actions quite nice.
In Chat mode, it also works quite well giving it some little guidance on what I want from it.
And it generates answers instantly, all the way up to 300 or more tokens per answer consistently, then when the token limit is reached (using FlashAttention, no SmartContext nor ContextShift, both suck), it takes less than 2 seconds on average to keep generating (and that's when the GPU starts sweating =P).

AMD is fucking wild now on Windows with ROCm for LLMs and ZLUDA for Stable Diffusion. It's not Nvidia's level of performance yet, but it is quite close, and it took too long to get here imo.

 

[LogicalWizard]

Anyone does voice calls with AI? I feel it would be too gimmicky

 

[n3otidus]

You must be registered to see the links

You must be registered to see the links

I’ve recently come across these two AI chatbots and have been trying them out for a while. I’m curious if anyone else uses them and which one they prefer. In your opinion, which of the two provides more accurate and complete responses? What makes you prefer one over the other?

 

[Princess Groundhog]

I still find that Character AI still has the best memory, and its SFW filter seems weaker nowadays.

 

[Geigi]

I haven't played a single F95 game since discovering chatbots a few months ago.

Me too, they are incomparable.

 

[D0v4hk1n]

Hi guys does anyone know what's wrong with Janitorai?

Since a few days it's no longer possible to chat with multiple bots at the same time. Last week the bot was generating long responses despite having several bots at the same time.

 

[LogicalWizard]

Is there a free ai chat box where you can generate quick photos within the chat room itself

 

[abyss50055]

I'll say it again:

Chub AI -> Mars Tier -> Mixtral 7B (or Asha 70B) -> "A decent Mercury" preset

...is a god-tier setup. (...)

Thx for recommendation! I’ve just started playing around with AI chatbots on Chub. Isn’t mercury the lower subscription tier on chub? If you have a mars subscription wouldn’t it be better to also use a mars preset or will Mixtral/Asha perform better with a mercury preset?

 

[deepsauce47]

Thx for recommendation! I’ve just started playing around with AI chatbots on Chub. Isn’t mercury the lower subscription tier on chub? If you have a mars subscription wouldn’t it be better to also use a mars preset or will Mixtral/Asha perform better with a mercury preset?

All the built-in Chub models (Mars, Mercury, Asha) are similar enough that any of the presets should be interchangeable. I spent a long time looking for the best preset, and that was the one that I got the best results with, by far. It was much better than the top-rated popular Mars and Asha presets.

Lately I have only been using the Asha 70B model. I dunno what it is but some weeks ago Mars quality seemed to go way down.

 

[abyss50055]

All the built-in Chub models (Mars, Mercury, Asha) are similar enough that any of the presets should be interchangeable. (...)

Interesting! So far I’ve just tried out Asha and Mixtral with the popular / recommended mars presets. I’m already impressed by both models, but I’m still new to all of this, so I can’t judge how good these models actually are. I did get the impression though that Asha is more reliable and better at handling complex scenarios. I’ll try to combine Asha with a mercury preset next, thx for the explanation!

 

[growthhormone0722]

I've gotten into these recently, so I'll share my findings. I prefer to run everything offline, so my experience is entirely based on models you can actually run offline (so no 70B models. Most PCs aren't even remotely close to being able to run those).

SillyTavern with koboldcpp backend using NemoMix-Unleashed 12B Q8 has given me the best experience so far. The model doesn't have that annoying positivity bias that the llama trained models seem to have, it can be fairly creative at times without going too far off the rails and it's also completely uncensored. Using koboldcpp allows me to actually run it with somewhat acceptable speeds at around 2.5 T/s (compared to oogabooga which just gives up and shits itself). And sillytavern is a vastly superior frontend compared to the koboldcpp default when it comes to 1 on 1 character RP, allowing for a whole range of things including static character sprites, ai voice, vn-like ui, backgrounds, image generation, etc all built into it. Using these three combined with WebUI and Alltalk for image and voice respectively gave an experience that really put chatbots on the map for me. They're like visual novels juiced up with the most potent seroids imaginable.

If you're planning on trying out local models, I highly HIGHLY reccomend NemoMix. It understands context really well and seems to be good at both NSFW and SFW roleplay. I've tested out a couple of other 12Bs and 8Bs, but NemoMix is the model that is most consistent in my opinion. If you can't run that, some of the 8Bs are fine too. My favorites are Stheno3.2 8B, Niitama 8B, and Celeste 8B. All of them will start to repeat themselves after a while though. The main draw to these is that they can run on a lot more computers, but there's a tradeoff in quality. So if you're struggling with the larger 12B, try one of these. They're pretty good but once you try a 12B you're not going to want to go back to the 8Bs lol.

 

[Princess Groundhog]

Uh that’s a lot of words. I just go to a website and look for nice-looking characters

 

[fbass]

I've gotten into these recently, so I'll share my findings. I prefer to run everything offline, so my experience is entirely based on models you can actually run offline (so no 70B models. Most PCs aren't even remotely close to being able to run those).

SillyTavern with koboldcpp backend using NemoMix-Unleashed 12B Q8 has given me the best experience so far. The model doesn't have that annoying positivity bias that the llama trained models seem to have, it can be fairly creative at times without going too far off the rails and it's also completely uncensored. Using koboldcpp allows me to actually run it with somewhat acceptable speeds at around 2.5 T/s (compared to oogabooga which just gives up and shits itself). And sillytavern is a vastly superior frontend compared to the koboldcpp default when it comes to 1 on 1 character RP, allowing for a whole range of things including static character sprites, ai voice, vn-like ui, backgrounds, image generation, etc all built into it. Using these three combined with WebUI and Alltalk for image and voice respectively gave an experience that really put chatbots on the map for me. They're like visual novels juiced up with the most potent seroids imaginable.

If you're planning on trying out local models, I highly HIGHLY reccomend NemoMix. It understands context really well and seems to be good at both NSFW and SFW roleplay. I've tested out a couple of other 12Bs and 8Bs, but NemoMix is the model that is most consistent in my opinion. If you can't run that, some of the 8Bs are fine too. My favorites are Stheno3.2 8B, Niitama 8B, and Celeste 8B. All of them will start to repeat themselves after a while though. The main draw to these is that they can run on a lot more computers, but there's a tradeoff in quality. So if you're struggling with the larger 12B, try one of these. They're pretty good but once you try a 12B you're not going to want to go back to the 8Bs lol.

I'm doing the same but I got a 70B to work, pretty amazing.

 

[TheHighSpire]

I took the most expensive NovelAI (25$ per month) for a spin and was really impressed. On a scale of 0-10 where 0 is role playing with an orange and 10 is your favorite porn author is role playing with you and helps you relive your nastiest fantasies, I'd rate NovelAI's chatbot an impressive 7.

It does have its typical AI glitches (like forgetting stuff) but it never tried to block my sessions (nasty as they were) or steer me towards a SFW scenario. Additionally, from time to time it actually surprised me with good "plot initiatives" although I'd estimate that I still have to edit/delete about 20%-30% of its output. So: a 7/10 which amazing IMHO, especially since I'd rate its earlier incarnations from less than a year ago no more 4-5 out of 10, very frustrating affairs.

I do miss the old text adventure commands Look (describe the scene to me) and Examine (describe a specific object/person to me), but that's beside the point.

Yeah agreed on that. Though one of my biggest problems with NovelAI was its lack of details. It doesn't really describe what is going on in a room, and it is also unclear exactly where it wants you to go. That probably stems from the fact that it don't want you to do anything necessarily, but it can mean some pretty slow play which could've been better.

 

[TotalHaydium]

I've gotten into these recently, so I'll share my findings. I prefer to run everything offline, so my experience is entirely based on models you can actually run offline (so no 70B models. Most PCs aren't even remotely close to being able to run those).

SillyTavern with koboldcpp backend using NemoMix-Unleashed 12B Q8 has given me the best experience so far. The model doesn't have that annoying positivity bias that the llama trained models seem to have, it can be fairly creative at times without going too far off the rails and it's also completely uncensored. Using koboldcpp allows me to actually run it with somewhat acceptable speeds at around 2.5 T/s (compared to oogabooga which just gives up and shits itself). And sillytavern is a vastly superior frontend compared to the koboldcpp default when it comes to 1 on 1 character RP, allowing for a whole range of things including static character sprites, ai voice, vn-like ui, backgrounds, image generation, etc all built into it. Using these three combined with WebUI and Alltalk for image and voice respectively gave an experience that really put chatbots on the map for me. They're like visual novels juiced up with the most potent seroids imaginable.

If you're planning on trying out local models, I highly HIGHLY reccomend NemoMix. It understands context really well and seems to be good at both NSFW and SFW roleplay. I've tested out a couple of other 12Bs and 8Bs, but NemoMix is the model that is most consistent in my opinion. If you can't run that, some of the 8Bs are fine too. My favorites are Stheno3.2 8B, Niitama 8B, and Celeste 8B. All of them will start to repeat themselves after a while though. The main draw to these is that they can run on a lot more computers, but there's a tradeoff in quality. So if you're struggling with the larger 12B, try one of these. They're pretty good but once you try a 12B you're not going to want to go back to the 8Bs lol.

Would you mind helping me? I don't understand how to run nemomix with koboldcpp, I thought only specific ggml models worked on it, like wizardlm. I also don't understand how to load the nemomix model with koboldcpp, as it seems to have come in 5 parts?

 

[TotalHaydium]

Would you mind helping me? I don't understand how to run nemomix with koboldcpp, I thought only specific ggml models worked on it, like wizardlm. I also don't understand how to load the nemomix model with koboldcpp, as it seems to have come in 5 parts?

Nevermind! I downloaded the wrong one I think.

 

[LogicalWizard]

Has anyone gotten a subscriptions for Spicy Chat AI? Is it worth it? I lowkey wanna do it

 

[Daba]

Has anyone gotten a subscriptions for Spicy Chat AI? Is it worth it? I lowkey wanna do it

I haven't used it for a long time, but Spicy Chat AI was solid enough as a free version, you just have to wait for your turn.
If you are thinking of subscribing, I recommend Kindroid AI which is much better. Both as LLM and image generator. Only downside of Kindroid is 10 character slots, but we pressure dev to give in more.