So, while I have nothing to offer at the moment other than a question, I thought I'd start this string so that any discoveries/wisdom/finds made by our community can be shared.
Please, Mistress?
Moderator: 1885






Yes, memory is not great for almost all these AIs. KoboldAI at least have some option to make it better, like world info. Is not 100%, but can help if you create some triggers.zenlover wrote: Mon Feb 27, 2023 4:05 am Messed around with Chai this evening. The bot told me to cum after a bit of instruction. For a moment, until I reminded it I was male the instructions switched to treating me as female, and telling me to undress, even though it had already undressed me once already. But no puritanical filter.

Here's the git for FasterTransformer. NGL, this goes over my head by about paragraph 3, so I could be mis-reading the information here, but it was referenced elsewhere as an optimization to reduce the workload for a 175B model to a single GPU (3090 was referenced). I can't find the secondary reference, but I am certain that FasterTransformer is what it was referencing.Wen74 wrote: Tue Feb 28, 2023 9:52 am
I would really like to see the reference for running a 176B LLM in a single machine. Yes, you can theoretically dump quite a lot of vram in a single computer - if you are a millionaire.


The site changed some things, but you can find your bot here: https://www.chai-research.com/demo (if you are logged)zenlover wrote: Tue Mar 07, 2023 1:04 pm I had some fun for a few days with CHAI on my desktop. Tried to log ion tis morning and it looks like they are past the stage where sexual play is allowed. All of the various bots that others were built and testing have disappeared, and it looks like they are going to be on phones only.
What will be next?
Anyone else having a bit of luck without jumping through a bunch of complex setup challenges?




