Video for AI Chat 1.42 — Smarter AI, Better Dialog, Longer Memory

Dev

>Take your voice and run it through AI to get the text There are local open source scripts that do this easily already. No need for server side. >Take her response and what you say, and run an AI to determine if a vamX action should be triggered in the scene, and if she should have an emotional change in the scene This sounds like prompt engineering, or context setting. Again can be done locally. The triggering can be done locally in the code. > Take her response and generate a voice for it Turning text to speech tts isn't too compute hitting. The LLM could depending on the model used. >So, to make it local, we need to solve some very tricky issues of how to handle all of these things locally, including the coordination of all of that, while also letting people run VaM, where generally, VaM itself pretty much occupies all of a computer's CPU and GPU, without MULTIPLE AI models running at the same time on that computer. I do agree if running vam and llm on the same system with full graphics and a large llm model could tax the compute on the system. Although there are now many open source llms that are tiny and getting better every day. Hardware is improving and some people already pay monthly for subscriptions for remote models. Adopting open source apis that can connect to these possibilities is only going to make modding the future easier and ideas from others you never thought possible when compared to your current system. I also feel the subscription plan on your side does gives more money, but over time will not make as much sense as things improve and focusing your time in areas that does not benefit everyone.

vamX

We do provide a save/load function for a single chat to be continued. Also you can press summarize (up top) and then save / load that text manually. The context is also somewhat dynamic, in that the base personality (whatever is in the personality section) is always present, plus the latest amount of conversation the context can handle. Aside from being able to save / load multiple different conversations more conveniently (without having to hit summarize) I'm not sure how this would help.

edwingrules

It'd be really cool to have them save memories into a vector db so that they can have long-term memory, as well dynamically handling the context so they can talk indefinitely. In could be stored as a local postgres db to stay private too. If you're curious about how to start, check out the RAG agent here as a concept https://docs.ag2.ai/notebooks/agentchat_RetrieveChat . This is the kind of stuff I work on. DM me if you want to collab.

Minor Flume

then the more important question is the level of security and animosity of user data. I do think it is a little sketchy for the back end to access or tie our data to a user. Especially around sexual fantasy topics.

PSF

With a small Q4 quality model and latest libraries running something like MeloTTS you could definitely do all of that locally on a solid gaming PC, I think, with pretty decent results. It might be some work to set it all up so it runs together, but I don't see why not.

vamX

It may or may not ever come. We've created a huge amount of custom code, plus the amount of things the server does is way beyond what a single computer can accomplish. For example, there are separate servers that all run at the same time that: 1) Take your voice and run it through AI to get the text, and maybe translation if required. 2) Take what you say and run the LLM/model (of course). 3) Take her response and what you say, and run an AI to determine if a vamX action should be triggered in the scene, and if she should have an emotional change in the scene. 4) Take her response and generate a voice for it. If you asked your computer to do all of that, plus run VaM, it just wouldn't, even if you have a 4090 card. You couldn't even load VaM + an LLM, + 3 additional GPU intensive models that handle those things, on a single a computer. So, to make it local, we need to solve some very tricky issues of how to handle all of these things locally, including the coordination of all of that, while also letting people run VaM, where generally, VaM itself pretty much occupies all of a computer's CPU and GPU, without MULTIPLE AI models running at the same time on that computer.

Dev

It'd be nice to try this out, but I will never do remote for this. When is the local or custom endpoint coming? I have a powerful pc where I can already run local models, so no reason for this to happen on your side. There are already open source loaders for llms that accept api calls where this could be connected.

vamX

If you haven't experienced in VR, you should. Many people eventually play in Desktop after having spent a while in VR, especially if they want to create something in particular or want a quick experience, but if you've never tried it in VR, you should.

Jayson Vazquez

is it better to play with vr or desktop

Michael Rusniak

TY for the help!

vamX

You'll find the password in the welcome note, it starts with "sw". This password is entered into the AI Chat web ui (make sure you aren't trying to unlock Virt-a-Mate with it). After entering the password, make sure you press Reset Chat.

unknowngreatone

Marc

Back on! Great stuff :)

vamX

Thanks! There were some DNS issues which are now fixed. Please try again!

Marc

Hmm is https://sxvr.com/ai down atm? Because I'm getting DNS error and just fails. I'm in the UK. Thanks!

vamX

If it's an 18+ non copyrighted personality, you can send it to use for general inclusion. Otherwise you can either: 1) manually select copy & paste it 2) Use the separate voice recognition deepgram app to manage custom personalities 3) host the personality list online privately (slightly more advanced) see https://sxvr.com/ai and search for CREATING YOUR OWN PERSONALITY LIST

serj tank

I've written a new personality, how can I save it and load next time?

vamX

Also, if you send a copyright clear AI generated thumbnail image at least 128x128 we will use it, otherwise we'll make our own. Please send to vamx@protonmail.com

vamX

We would love to have donated characters, and we have a section for user generated characters and would LOVE to have more. Characters must be 18+ and can't be copyrighted. If you have a Wizard, don't call him Harry, but Henry Pots would be fine.

Volmarr the Viking

Hehe! Love how she has no idea she is already an AI character and also how human like she is now! Maybe in the future with the new models and longer token length I might see if my characters work really good now in VamX. I tend to write very long characters with very complex realistic personalities. Mostly before I was pushing VamX AI to it's limits with my characters before this recent update. If they run smoother, who knows, might be interested in donating some of them to the project. What are the guidelines/rules about donated characters?