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 ba

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.