There are a lot of people who are entirely okay with the censorship but think it should be done in a different layer than the main LLM itself, as not to hurt th

There are a lot of people who are entirely okay with the censorship but think it should be done in a different layer than the main LLM itself, as not to hurt the cognitive performance. Alignment is just fine-tuning... any type of fine tuning is possible to teach unwanted skills, and/or catastrophically forget previously learned skills. That is likely what is going on here, from what I can tell from the reading i've done into it. Most are arguing for a specific "censorship" model on the input/output of the main LLM.