>>16213
If you're talking about the offline AI thread, as the guy who made that thread, I am totally fine having a more general thread for this stuff, and rming that thread.
Atm, all I've been doing is connecting the modules together. So now I've got it set up so the bots on my offlineweb just run overnight, interacting with each other on all the different sites. I've got their memories, emotional states, etc. set up in profile jsons, and they slowly change after sleep cycles based on their site activity in the previous day.
It's been interesting, but so far the biggest improvements I've seen have been to:
- Make their "brains" dual-modeled, following a Kahneman "1 and 2" style thinking.
- Giving them emotional states that change with each act of speech.
- Of course, upping the model weight sizes.
Regarding the two model style, the Jev models are kind of interesting.
>>16212 (OP)
>Since the usual suspects seem to be constantly gatekeeping this while the typical corporate inbreds are trying to destroy everything it's time to have a thread and generate at least some interest and informative value on recent AI/LLM developments without any bullshit.
Most of the anti-AI posts are written by bots. We got Uncle Remus Bots before returning to the Moon.
>You should also be saving some of these models, especially the historic ones, and keeping them in your digital library for safe keeping. This area moves fast and there's no telling what will happen.
I have my setup cloned onto an SD card now. Just the fact that something I'm pretty sure will be made illegal in the future I can stuff surreptitiously into a coin case and clone on a dime makes the entire concept of trying to make this "illegal" so absurd to me that I had to do it.
In other words, it's sneakernet capable.
>rtfm
You don't even have to do that. Once you get a barebones model going, you can then just ask that model to rtfm for you regarding what commands you have to run to get the next model up and running or configure your card better or cetera, and you're off to the races.
>download Comfyui and the generation model you want
Case in point, once you run basic ollama, you can ask it what you have to do to get ComfyUI going, and then optimize by directly running on llama.cpp, etc..
>We should be figuring out ways to train or adapt our own models too because LLM models are becoming increasingly censored and braindead because the mainstream cuck tech fags keep pozzing training data into newest models now gradually polluting ALL models(because they have billions and we are poor).
Even still, what abliterated models there are are a breath of fresh air.
>>16213
>modded 2080ti 22gb
Nice.