>>17564339most people dont know you can run large language models on basic pcs and even phones right now (i assume just android right now because of its low-level linuxness), as you can guess running local models gives you more control to evade filters and yea some models have built-in biases, well technically all models do since essentially the only thing llms do is text completion, models are just a bunch of weights and math saying statistically this "word" follows this one and these "words" follow these ones, but anyway back to biases in the colloquial sense, the first non-shit local model available for people to use was a leaked model from facebook so you might think the way they trained their model may filter or bias it in the ways you mentioned like uncharacteristic positivity and things like that and the answer is kinda? first off there's a difference between filters that users experience and filters inherent to a model, for example with gpt4 it's been documented that using their online interface gives users much more filtered responses than utilizing the gpt4 api. secondly people can finetune models to dramatically reduce filters even more, not to mention jailbreaking local models is far easier than online ones. as for requirements, general guide is worst uses 8gb ram/vram, good uses 12gb ram/vram, and best uses 25gb ram/vram and consensus on /g/ is that the best local model right now gives responses indisputably superior to charai though with one small caveat. local models' most significant shortcoming right now is their limited context which is essentially their memory. right now its about 2000 words which is decent for chats and short stories, most people are horrendously shit at estimating wordcounts so they dont really know what 2000 words actually means but yeah, of course there are ways to "increase" context like giving the bot summaries when you get near the limit but currently online beats local for super longterm memory