Just checking in: how are offline text-based AIs doing?Image-based ones (SDXL and all its models) are great and just as good as the online megacorporation AIs, but the text-based ones always lagged behind ChatGPT and Claude and the likes.How are they now? Can I run an offline text-based AI on 12 gigs of VRAM that competes with ChatGPT?
>>101538887>but the text-based ones always lagged behind ChatGPT and Claude and the likesI always heard that but never saw any concrete examples. All I ever see is>As a large language modeland that's when I can't read any further because I have tears in my eyes from laughing.
>>101538887the most unhinged LLM I came upon is wizard-vicuna-uncensored, if you want it to say nigger while giving you tips how to cook meth, rape a woman and do photos of maimed victims for fun, Llama 3 uncensored is actually usable though, gives better results, doesn't come close to GPT-4 though
>>101538970>say nigger while giving you tips how to cook meth, rape a woman and do photos of maimed victims for funfanfics of copyrighted material with some romance and maybe sex scenes is enough for meif that's too degenerate for you then fucking kill me i guess
>>101538970also stable diffusion is shit, literally unusable when compared to Dall-E, I hope the model leaks some day just like GPT-4o, what a nice day it would be, anyway, SD struggles with nudity, reflections, abstractions, hands et cetera, SD3 only fixed text
>>101539019lmao DallE doesnt even get the fucking aspect ratio right, bitch please
>>101539016wizard-vicuna was the only model that told me how to cook meth step by step with detailed instructions down to the what glass should I use, what temperature should be and how much grams of each ingredient I need to use, interesting stuff