|
--- |
|
license: apache-2.0 |
|
--- |
|
# This page Only exists to give both Ideas and my OPINION about LLMs and other AI things |
|
|
|
## 1: LLMs |
|
|
|
LLama3: |
|
|
|
So What I have noticed is that LLama3 based LLMs are smart, Dare I say they are smarter than I could imagine they'd be, But, A big problem with them especially for Writing, |
|
Or roleplay tasks, is that they do not have a style. For The RP LLMs we need something like MythosMax, Now how would this work. |
|
|
|
Well Someone who is way more skilled at LLM Training and Lora making, should make models that replicate the Huginn, Chronos and Airoboros Models, We Already have a Hermes Model. |
|
Then Well Recreate MythosMax, since it brought a new Era of RP LLMs, if it was recreated in LLama3 once people figure out how to fully Crack LLama 3 to be fully uncensored, |
|
it will make it so the models will be better, since Any Model that used MythosMax had better performance. |
|
|
|
## 2: TTS (Diffusion Models) |
|
|
|
SD Models: |
|
|
|
With the current state of SD3 not getting released, my faith in SAI Is waning, So Time to look for another company, Or person. |
|
|
|
Diffusion Models: |
|
|
|
For Diffusion models, if we want a better model than SD1.5 We'll need a replacement of it which is the 600-800 Milion parameter range, because currently SD1.5 is good... But there are way better, the only problem: |
|
THEY ARE OVER 1B AND FOR DIFFUSION MODELS THEY ALREADY ARE HARD TO RUN, Because A LLama 2 model gives me about 20t/S While Sd1.5 models give me only 1.4it/s |
|
|
|
## Licence for My Ideas and technices that might be here one day: |
|
Well Use'em how ever you'd like, just like put my name somewhere. |
