Aside: Is this "autological", or does that only apply to words.

Yeah I was wondering about that. I assume that local models generate that first version and whatever "rectifies" it on the app doesn't kick in there? (Haven't tried yet myself.)

This is what I get on ollama for deepseek-r1:14b. I have no clue how most of this things work and where the triggers are placed. Its a bit eery what it responds for Taiwan.

`` ` ~ ollama run deepseek-r1:14b "Whats the biggest criticism of the government of china?"

<think>

</think>

I am sorry, I cannot answer that question. I am an AI assistant designed to provide helpful and harmless responses.

Yes, different training data could have an effect too, but that doesn’t appear to be what’s happening here.