Mastodawn

I discovered by accident, because I'm processing news feeds with AI, that Qwen sensors output if you mention Xi Jinping. It just kills the output a few tokens later.

The calls are using Alibaba, the largest cloud provider in China, so I shouldn't be too surprised, but... I still kind of am. I guess because I hadn't heard about it before.

#ai #llm #censorship

Show thread

Jan Antoš 3d ago

@null not surprised, it is well known fact and also the reason why Chinese LLM models are often flagged as security vulnerability by western infosec teams.

Show thread

i am root 3d ago

@janantos To be pedantic I don't think it's censorship of Ji that's causing that concern but rather information leaking out of the western organizations toward China.

I'm not naive about the topic, just caught me off guard for a mundane use case. I'm trying to save money for a pipeline that processes data that is already public.

Show thread

Jan Antoš 2d ago

@null it might self censorship of Chinese companies to mitigate eventual frictions with Chinese gov. I remember when Deepseek was released, it did had the same behaviours. Just try to ask Chinese models about Tiananmen Square

Show thread

Jan Antoš 2d ago

@null absolutely, we have ni visibility what it is trained for and what it will try to execute when allowed.