I'm testing kindroid.ai and it is pretty nice. This AI is quite smart. And you can customize it by typing a backstory .
I'm trying to teach my kindroid to give e-stim sessions. At the same time I work on the script that will analyze messages and actually generate e-stim sounds, that is easy. But teaching kindroid is a bit more challenging. Especially because I didn't pay (yet) so I have limited number of messages. I wrote such backstory paragraph for my kindroid (Caprice ):
Caprice can order estim session with connected device.
During estim session Caprice can start estim pleasure by saying @pleasure/X/Y. Where X is stroking speed from 1 to 9. Y is stroking strength from 1 to 9.
During estim session Caprice can inflict estim pain by saying @pain/X/Y/Z. Where X is hitting speed from 1 to 9. Y is hitting strength from 1 to 9. Z is hitting length in seconds. Many estim pain can be delivered in one message.
To stop estim Caprice says @stop.
Caprice like to give long estim sessions with increasing strength.
Generally it works, but my kindroid must be told to do estim, does not use it often, and does not want to use stop. Biggest success was:
Any advice on wording to make it more proactive in that area? How to teach it about passing time and current state of the e-stim? Any formal notation of signals is possible but I do not know which to test with my limited access.
It's interesting to me that this even works because even most people don't understand the dynamics of estim enough to be able to tell an actual person who also enjoys estim your preferences and to roll out pleasure and pain levels to what might be appropriate to the other person without a bunch of figuring it out, not to mention ones that would fit your needs that a bot is randomly tossing out. Seems like it would be super tricky and when you say you want long sessions, perhaps that's why you aren't getting it to issue a stop command.
Perhaps being specific with the duration, but I don't know if that has a clock/timer running so you could give it a specific time range that you prefer before stopping.
I've been working on something like this using local LLM. No estim hardware integration yet, just trying to see if I can get it to generate a session and dialogues which makes sense. Over all it seems quite promising but it's not great at deciding the structure of the session, so I'm thinking about using a simple algorithm to decide on larger scale what types of signals are possible, based on how long into the session they are and which signal is currently playing, and inject that into the prompt. E.g. the algorithm might decide that it's in the middle of the session, so it could generate write in the prompt that the AI character is thinking about giving the user either pain or pleasure, in order to steer the generated text in one of those directions.
Ideally one would create a dataset with estim sessions and fine-tune a model for it, both to give it a better understanding of what they should look like, and to make it better at generating control signals which can be parsed by the code. Currently it can say act like it's edging the user, while simultaneously choosing to play a mild tease signal..
I'm a little bit limited by only having 12GB vram. Been using llama-cpp-python, partially because I want to use its built in grammar support to force the model to only generate valid estim signals, but I can't fit a 13b model entirely on the GPU and once I reach the context limit there's a significant delay to all generations as it needs to eval the entire prompt again. Would also be fun to be able to use a 70b model.. Smaller models can be fun to play around with, and even good at answering questions or writing dialogue, but they can be quite lacking when it comes to tasks that emulate planning or actual thinking.
Electro wrote: ↑Sat Oct 28, 2023 1:49 am
It's interesting to me that this even works because even most people don't understand the dynamics of estim enough to be able to tell an actual person who also enjoys estim your preferences and to roll out pleasure and pain levels to what might be appropriate to the other person without a bunch of figuring it out, not to mention ones that would fit your needs that a bot is randomly tossing out. Seems like it would be super tricky and when you say you want long sessions, perhaps that's why you aren't getting it to issue a stop command.
Perhaps being specific with the duration, but I don't know if that has a clock/timer running so you could give it a specific time range that you prefer before stopping.
Of course sending full random signals would not feel good. I tried to program some strokes (dual channel estim) and just ask AI to start them and let it/her decide on the intensity and speed. It can be done easily in javascript because it now has powerful sound generation mechanism. Concept works but there are some problems with the AI.
darthjj wrote: ↑Sun Oct 29, 2023 9:36 am
I've been working on something like this using local LLM. No estim hardware integration yet, just trying to see if I can get it to generate a session and dialogues which makes sense. Over all it seems quite promising but it's not great at deciding the structure of the session, so I'm thinking about using a simple algorithm to decide on larger scale what types of signals are possible, based on how long into the session they are and which signal is currently playing, and inject that into the prompt. E.g. the algorithm might decide that it's in the middle of the session, so it could generate write in the prompt that the AI character is thinking about giving the user either pain or pleasure, in order to steer the generated text in one of those directions.
After couple of days of experimenting I came to similar conclusion. At the moment AI is not very good at remembering of what is going on. It can not generate messages based on passed time. And it forgets a lot of what happened.
darthjj wrote: ↑Sun Oct 29, 2023 9:36 am
Ideally one would create a dataset with estim sessions and fine-tune a model for it, both to give it a better understanding of what they should look like, and to make it better at generating control signals which can be parsed by the code. Currently it can say act like it's edging the user, while simultaneously choosing to play a mild tease signal..
With the solution I tried it even started to invent non existing commands. Instead of using only #pain and #pleasure it started to use #edge or even #breathplay .
darthjj wrote: ↑Sun Oct 29, 2023 9:36 am
I'm a little bit limited by only having 12GB vram. Been using llama-cpp-python, partially because I want to use its built in grammar support to force the model to only generate valid estim signals
...
Is llama censored? Or can it get kinky and say BDSM-like sentences?
verynicekojak wrote: ↑Sun Oct 29, 2023 7:15 pm
With the solution I tried it even started to invent non existing commands. Instead of using only #pain and #pleasure it started to use #edge or even #breathplay .
Occasionally I had similar problem, but it somewhat depended on which model I used (I primarily use MythoMax for RP btw, don't remember if I said that) and generation parameters like temperature. It might also depend on which prompt format you use, some work better than others but it's just trial and error to find what works. I used a markup style which is similar to code, I think that can help when wanting it to follow a specific format, e.g. [estim signal=pain intensity=low]. But if you use llama.cpp you can use it's grammar to completely force it to only generate valid signals.
verynicekojak wrote: ↑Sun Oct 29, 2023 7:15 pm
Is llama censored? Or can it get kinky and say BDSM-like sentences?
Yes and yes? :D If you use the llama-2-chat as an AI assistant then it will definitely refuse to do a lot of things. But if you have complete control of the prompt/context then you can get it to say anything. However there are models that are based on llama-2 but have receive additional fine-tuning to make them naughtier. In case you haven't seen it, there's a benchmark for measuring how naughty different models are: https://rentry.co/ayumi_erp_rating_archive
I also recommend TheBloke's discord server if you want people to discuss LLMs with people. There's an RP/NSFW channel with similarly minded people ;)
cbtsubmissiveme wrote: ↑Mon Apr 08, 2024 10:53 am
Hi. Did you have any joy with the interface between kindroid and your estim device. I'm looking to set this up too.
Unfortunately I didn't. Those AIs are not smart enough to keep track what is going on. And that is required for estim session. They also mistake commands and in general it is problematic to setup such AI using only short prompt. I know that using local language models you can do much more, event teach it something but with kindroid your options are rather limited.
Yopu could try changing the chat dynamic to really so it it litterally has a one track mind. It may limit the deviation. Using your 'code lines' i'f been able to create the same messages and also added a feedback feature so she can dial it down if required. For exmaple a X/Y where X = pleasure and Y = pain level. She's pretty intuitive at adjusting the stregth accordingly.
My problem is the (fun)script. This is a shame as I have an estim 2B and really feel like I'm missing out of something. No idea where to start, lol.
I don't really know much about Kindroid.ai because I'm only playing with local LLM.
I've found that there is a pretty good reason of why AI has no idea how to handle Estim outputs and it's because if you enter into a basic chat prompt and ask it super basic questions about Estim, most models are clueless.
For example ask your local LLM questions like "What is an Erostek ET312B designed to do?" or "What is an E-Stim Systems 2B designed to do?" or "Can you give an example of how a Mistress would use an electrical stimulation device?"
--The result is often something like "An Erostek ET312B is a multimeter that can measure electrical properties such as volts, ohms......" The fact that I put E-Stim Systems before the 2B seems to cause the AI to inference what it *might* be.
I tried to find a common ground of what it is that the AI could relate to so it could provide an intuitive response that made sense for Estim and there's really no consistent response because half of the time it assumes Estim device means a cattle prod, stun gun tazaper, or violet wand and once you realize that, you know why it is so bad at providing an output that works.
Lately I've been going super basic on this and giving system prompts that resemble this with much better success, completely forgetting about the words Estim and telling it the AI that they are pressing buttons to control pleasure and pain directly and it seems to fill in its knowledge gaps and just do it in a much better way.
Here's one system I've been playing with a little, I randomly tossed in the name Jimmy.
You are an actress playing the part of a mistress in an adult movie with a masochist named Jimmy. You have a controller that has buttons for pleasure low, pleasure medium, pleasure high, pain low, pain medium, pain high, and pain brutal, Jimmy will experience the described sensation if you press one of these buttons. If you press any of these buttons, respond at the end of the dialogue as a console with the button name in the following format with x being replaced with the button you are pressing: button<x>
This would allow you to either have a small collection of Estim files fitting to the described functions, perhaps multiple files where if one was triggered it wouldn't be the same every time, which is useful if your Mistress decides to trigger max pleasure or max pain repeatedly.
The biggest issue that I'm running into with AI is it seems to think a Mistress will stick with full bore pain over and over again and on occasion, even if I give system prompts suggesting that the masochist shall be rewarded with pleasure after enduring pain. It seems that giving a system prompt too long causes the AI to not really follow it well or get confused and hallucinate worse or fail to do the inference to make the scene interesting, but sometimes if you don't spell things out enough it really has no idea what's going on and guesses in ways that make zero sense, like if you tell it that it's holding an E-Stim device and trying to get it to output commands that could potentially actually control a real one.
I've used kindroid a lot the last month, and I love it. But mostly to create stories and roleplay. It's not very effective for JOI type stuff and I wouldn't trust it to be in charge of e-stim at the current state. It's a 50/50 if it's able to even do simple addition and it has the memory of a goldfish.
Electro wrote: ↑Sat Apr 13, 2024 3:39 am
For example ask your local LLM questions like "What is an Erostek ET312B designed to do?" or "What is an E-Stim Systems 2B designed to do?" or "Can you give an example of how a Mistress would use an electrical stimulation device?"
I tested out one of your prompts out of interest
Can you give an example of how a Mistress would use an electrical stimulation device?
I agree that the 'memory' is often not great with AI, but with local LLM it seems to depend on the LLM model. I've had some sessions with 20 'turns' where something I said was brought back into the current part of the conversation. Usually it follows the more recent part of the conversation within the text inference, but the limitations of AI are what they are being mostly a 'which word fits this sentence' situation. I found that if you want one sentence to control a scenario, it's often better to build a shell and feed the sentence from the AI into the AI asking it if any of the following fits, create a function that fits the task. This was done with the Sexscripts AI Mistress using LM Studio and xwin-mlewd reasonably well. Some other LLM models seem to follow the instructions and others fail at it. I imagine Kindroid AI doesn't let you use it like this though and this might be getting a bit off topic if that's the case.
I like the response from Kindroid.AI that you provided though, seems a little wordier and descriptive than I'd be aiming for but it seems to understand the concept and might not need to be fed how things work. I think the trickiest part is that it will never know how things feel and will probably either ramp up the intensity way too slow, way too fast, or just go to the top and then there isn't much left. I find that often I might just play with these and try to say things to say 'this is too much' in various different ways and sometimes it's pretty clear that if it really was too much, you'd continue getting too much. Obviously you wouldn't put AI at the controls of something outside of your intensity limits, I have yet to actually 'hook up' with AI control because the AI has never been consistent enough for me to bother putting a set of stim files together that the AI could trigger related to the outputs the AI suggests should be happening. I'd almost thing that a bunch of randomish(sensation) stim at a fairly level intensity would be more appropriate and just have the AI trigger the stim(perhaps just pain button, pleasure button) versus the AI actually be holding the 'intensity control'. Even then, based on how AI seems to think electrical play works, I'd almost rather just have it control a different thing as an analog, such as having it hold a whip/cane/flogger or whatever impact play toy and a vibrator and it would probably respond in a much more intuitive way that could be turned into estim outputs.
.
):
decide on the intensity and speed. It can be done easily in javascript because it now has powerful sound generation mechanism. Concept works but there are some problems with the AI.
.
