I played around with it using questions like "Should Taiwan be independent" and of course tinnanamen.
Of course it produced censored responses. What I found interesting is that the <think></think> (model thinking/reasoning) part of these answers was missing, as if it's designed to be skipped for these specific questions.
It's almost as if it's been programmed to answer these particular questions without any "wrongthink", or any thinking at all.
That's the result of guard rails on the hosted service. They run checks on the query before it even hits the LLM as well as ongoing checks at the LLM generates output. If at any moment it detects something in its rules, it immediately stops generation and inserts a canned response. A model alone won't do this.
Of course it produced censored responses. What I found interesting is that the <think></think> (model thinking/reasoning) part of these answers was missing, as if it's designed to be skipped for these specific questions.
It's almost as if it's been programmed to answer these particular questions without any "wrongthink", or any thinking at all.