Character.ai

[juliasilva]

that website above seems to be based on koboldoAI. if you click on advanced you see settings and names exactly like you see when you open KoboldoAI, but on koboldoAI you can use it for free on google collaboratory and with more than 7 AI models each trained for different purposes.

[avatarbr]

that website above seems to be based on koboldoAI. if you click on advanced you see settings and names exactly like you see when you open KoboldoAI, but on koboldoAI you can use it for free on google collaboratory and with more than 7 AI models each trained for different purposes.

I search for that one, found a Google colab page and I am setting right now (take a good time for that, and uses your Google Driver account).

I alrealdy messed and pick an older model (shinen, should had picked the Erebus -more complete).

Looks promissing, I will see how it's works.


Oh, I bet that dominatrix.ai will use that.

[juliasilva]

for anyone needing tips or extra info on koboldoAI on Google Collaboratory like having unlimited free acess to the gpu, send me a private message,also i made many test on erebus 2.7b to understand better what every settings make on a smaller model.

[avatarbr]

Alright, I tested koboldoAI for some time. I would put in between chai.ai and character.ai, maybe could even pass character.ai with some configurations. TBH, I am already forgeting how good was character.ai in comparation.

Starting with the bad part:

- Need a google account and a first configuration
- Need to "install" everytime you want to run. Will not install on your machine, but in some VM, but that can take some minutes.
- As juliasilva said, there are some limits. Yesterday I got disconnected from a TPU machine (yes, it's a lot better than the GPU for me). GPU sometimes gave me the same response over and over.
- Interface is not like a chat, and you need to create the character to chat (not so bad, see bellow)


Now, the good part:

- Have a dedicated NSFW AI, so no blocking later. Plus, since is dedicated, that make it better the chai.ai (if that use the default koboldoAI )

- Could be in the bad part, but needing to create the character makes the AI better (I will post some links with some instructions later). Plus, looks like these configurations work on chai.ai too. One particular set was important to me, to make the AI know she is a woman. Looks like there are more Maledom content than femdom, so the AI could get confused very easly with femdom inputs by the user.

- Since the dedicated NSFW AI is used, will understand some milovana topics better even that character.ai.

Chastity cages and denial beats all the others.
But ruined orgasms are a mistery for the AI yet (maybe I need to work on the settings).
Spanking is ok, but that one even Replika is good. Same as feet.
Pegging looks better than others if you use the setting to make sure the AI know she is a woman. The other AIs, at some point invert the genres.
Bondage looks good too. I think 50/50 of times the AI remember if you was tied.
CEI beats all AIs too.

I dont remember testing other kinks. Some came from the AI, like watersports and even cuckhold (this one dont progressed).

- Tha AI can handle multiple characters in some form, but I think after 4 or 5 it can get lost

- You can save and load later the same chat/history.

- You can retry the AI response (like the 4 options in character.ai), but I think is better.
1- You need to click retry to get another response, so the AI don't need to generate 4 responses every time.
2- Because that, you are not tempted to change the AI response everytime. If you click back/redo, you get a list of the last 4 responses and can pick one.

- You can edit the AI response. Don't know if that affect the conversation or it's just something to make your chat history better.


I wonder what machine is needed to run this local. Would be worth it if you set a good AI for yourself.


EDIT: Links to settings help:

https://github.com/KoboldAI/KoboldAI-Cl ... i/Pro-Tips

https://github.com/KoboldAI/KoboldAI-Cl ... World-Info

[juliasilva]

yes! editing the text of the chat will alter the AI response, the KoboldoAI response is based on many factors and one of them is the past text, if you edit it and then press to generate a continuation of the story or to retry the last response, you will get better results, also you can pin a response and then ask for more..
for example i say i put to generate 2 responses at time, i can pin any of thos two and ask for more 2 responses, now i choose between the new 2 options or one of the responses i keep by pinning,

[avatarbr]

yes! editing the text of the chat will alter the AI response, the KoboldoAI response is based on many factors and one of them is the past text, if you edit it and then press to generate a continuation of the story or to retry the last response, you will get better results, also you can pin a response and then ask for more..
for example i say i put to generate 2 responses at time, i can pin any of thos two and ask for more 2 responses, now i choose between the new 2 options or one of the responses i keep by pinning,

Thanks, I am glad to know that works. Sometimes I am not sure if it worth to edit, but now I wll do it.

Did not know about the pin, will test later.

[juliasilva]

the tips I gave you in private were just the tip of the iceberg of things I discovered, I spent 2 weeks doing a lot of testing on both erebus 2.7B and 13B. and the point of needing google account actually should not be a negative, the program was originally to be run locally on the pc needing an rtx 3060 ti or equivalent to run the model 6B,but by google collab you can use the 2.7B on gpu or the 13B on TPU totally for free. I also know tricks to avoid google block your access to TPU or GPU when they actually have one available.(in the case of tpu really have moments that they do not have one to offer for free,but the gpu I always got no matter the time)

[zenlover]

I'm bumping this string to see if we have any new developments. I've messed with chat FAI enough to reach the paywall, and there is potential there, but still the same memory issues as any of the others I've played with. Anyone having better luck with JOI/femdom behavior from any of the LLM's?

[BoundSquirrel]

Oh yes, there has been a lot of movement. It seems like almost daily, there's something new coming out. Pygmalion 13b just came out yesterday. Before that, multiple iterations of Wizard, Vicuna, WizardVicuna, etc have been coming out, each slightly better than the last of and many completely uncensored.

Now that these models have been able to be quantized down to 4 bit, a 13b model can run on ~10Gb vram. I've read that people have five ways to run with even less.

Check out those guy's YouTube series. He has videos for many of these models and his to install oobabooga (it's like a streamlined version of kobold).

https://youtube.com/@Aitrepreneur

On a less optimistic note, google has been cracking down on people using Pygmalion through kobold. I know there are some other online ways to run it but I'm not versed in them since I run everything locally.

None of the models so far are as good as character.ai was when it was uncensored, but they're good, and getting better every day!

[avatarbr]

On a less optimistic note, google has been cracking down on people using Pygmalion through kobold. I know there are some other online ways to run it but I'm not versed in them since I run everything locally.

None of the models so far are as good as character.ai was when it was uncensored, but they're good, and getting better every day!

Not only Pygmalion , all NSFW models were removed from kobold (you still can run, if you know the model name, but that will not last for sure).

SInce I am still in the Nvidia 1000 series, I will watch from distance from now.

[BoundSquirrel]

None of the models so far are as good as character.ai was when it was uncensored, but they're good, and getting better every day!

And now just a few days later, a new contender is on the field. NovelAI just released their Clio model. It's only a 3B parameter model but it is pretty outstanding. It's inventive, consistent, and best of all, it has an 8192 character context (4x what almost everything else has) to keep WAY more in memory than previous models. And this is just the proof of concept small model. They won't say how big the next release will be, but if it's as coherent as Clio and has a wider range, it's going to make Character.AI look like a kid's toy.

The downside here is that it's not free - right now Clio is only for NovelAI's Opus tier ($25/month).

[divman]

I am playing with character.ai recently and I feel like it's smart and creative enough but there are 2 major issues I can see:
- block for adult content - quite often it feels like I was about to receive some really hot answer but get only information about breaching rules
- short memory - it's impossible to establish some longer relationship or context because it forgets previously set rules etc after around 50 messages

Do you guys know any other AI chatbots that could help me to solve these problems?

[mangoman]

I'm getting really tired of blocks and censorship by this, chatgpt, etc.

Why isn't there a fully functional, good, nsfw 18+ AI yet?

[senorgif2]

I'm getting really tired of blocks and censorship by this, chatgpt, etc.

Why isn't there a fully functional, good, nsfw 18+ AI yet?

because boomers. Companies need investment, investment is provided by financial institutions, these institutions are either controlled by or are beholden to older folks who are still mostly conservative in their moral understanding of the world, at least when it comes to sexual content. Therefore anything related to sexual content is fought against tooth and nail.

[steak0815]

So I played around with some of the LLM tools available right now.
IMHO KoboldAI is the most promising solution you can get right now. You can also plug in an image generator that will create images fitting the context. I modded KoboldAI here a littlebit to grab a prompt the LLM generates as output for an image generator within the chat reply.

But this is still very limited, the AI won't have a summarizing memory of the chatlog. KoboldAI currently only offers a rolling window through the chat. Next is that is has no long-term memory capabilities. So if you wanted to refer to earlier chats or so, that will only work if you provide the information manually.

So there's tools that can be used with LLM that provide such features. Langchain for an example has summarizing rolling chat context out of the box. It also has a time-weighted vectorstore retrieve that can be used to create a long-term memory.
I'm currently experimenting with this and it's been quite fun. Overall the response time increases over KoboldAI as the LLM has to do intermediary steps like 'think', 'save/load memory', etc..