having spent some time unscientifically but persistently comparing LLM censorship in response to the same prompts with CN and US models I'd say that some CN models I've explored have galaxies of off limits outputs. however they are different. so you need to ask the kinds of questions that get these answers. Which means you need to focus on censored themes in China. Like the five poisons...

https://www.businessinsider.com/chinas-spy-agencies-five-poisons-taiwan-tibet-uighur-democracy-2023-8

Reply to this note

Please Login to reply.

Discussion

Makes sense

From what i understand they don’t have to worry about IP so their training data can be much larger