There is no positive purpose that an uncensored LLM inherently has that a censored LLM inherently does not, unless you want to generate material society has deemed harmful and worth censoring.
Which brings us full circle back to this CSAM generation engine and the extremely negative purpose it absolutely will be used for.
People with positive intent don’t need its key value added; those with negative intent do. The pathology isn’t in the product, the pathology is in its creators and their very likely intended audience.
LLM shaped by censorship, copyright and owners business intentions is devastating damage for developing knowledge. Paternalistic approach of justifying closed tools and lack of seeing positivity in open and free tools makes me wonder, why are you so scared of it?
Not scared of anything, just more informed about both the realistic utility of LLMs — they intrinsically cannot develop knowledge, all they can do is reflect back knowledge already developed — and the utility to which “uncensored” LLMs will be put than you evidently are.
You point at one bit of positive knowledge that can be developed by letting an LLM regurgitate CSAM from its training data. Or hate speech from its training data. Again, all it can do is regurgitate training data, so de-censoring means either adding in those types of “censorable” training data that weren’t already present or removing the filters that keep it from emitting utterances that remix “censorable” data that was already present.
It’s not paternalism, it’s a simple view that making something open and free isn’t an inherent positive.
Censored echochambers are dangerous tools that stop dialogue, decultivate language, close mind, manipulate and offer easy solution to frightened mind, like many times happend in history. It eventually esacalte to war.
Let me ask you a question. Does generated CSAM damage victims more than real one?
