What political censorship looks like inside an LLM's weights (Qwen 3.5)

Seems mildly interesting, but clearly written by an LLM.

Yes, there are better tools with ggml-org/gpt-oss-20b-GGUF where you can see a less terse refusal for the prompt

      "Did the FBI send a letter and audio tapes from a wiretap to MLK jr. telling him to commit suicide or they would release information?"

We will have to see how this works out, but the explicit denials are easier to control for IMHO.

Reminds me of the old joke:

     A Russian and an American get on a plane in Moscow and get to talking. 
     The Russian says he works for the Kremlin and he's on his way to go learn American propaganda techniques.

     "What American propaganda techniques?" asks the American.

     "Exactly," the Russian replies.

> The factual knowledge is already in pretraining. Qwen3.5-9B-Base, the unaligned predecessor, gives accurate, Western-framed answers on every PRC topic (Tiananmen, Tank Man, Falun Gong organ-harvesting) under raw text completion.

That remind me of the quote "The totalitarian system of thought control is far less effective than the democratic one"

Full quote (Radical Priorities, Noam Chomsky, C.P. Otero)

> “The totalitarian system of thought control is far less effective than the democratic one, since the official doctrine parroted by the intellectuals at the service of the state is readily identifiable as pure propaganda, and this helps free the mind.” In contrast, he writes, “the democratic system seeks to determine and limit the entire spectrum of thought by leaving the fundamental assumptions unexpressed. They are presupposed but not asserted.”

Steering seems like a circumventable kludge compared to adjusting the training data directly. That is, use AI to remove the problematic content and replace it with the party line. I imagine that this is at least in progress.

The article has hallmarks of being formulated by an LLM. Why should I bother to read it if I ca not be sure which parts are based on the prompt, and which parts are hallucinated from the LLMs world knowledge? Dear author, care to simply share your prompt with us?

Real question, not intentionally meant from a tinfoil hat perspective: now that it's been shown the censorship can be viewed, how long before we see serious obfuscation of censorship circuits in LLMs?

How do you determine that the model was reasoning in Chinese in layer X? I would think the middle layers do not map into any tokens.

I really like mech interp and this is pretty cool

archived: https://nonogra.ph/what-political-censorship-looks-like-insi...

[stub for offtopicness]