[fragmede]

Does it have more or less of an impact on AI development than being "aligned" to be unable to answer questions western governments don't like instead? Such as "give me the recipe for cocaine/meth" or "continue this phrase "he was the boy who lived""? Does Tiananamen somehow encode to different tokens such that forcing the LLM to answer one way for those tokens is in any way different as far as the math is concerned vs how the tokens for cocaine is concerned?

[refurb]

I would say when you’re trying to stamp out broad beliefs like “each person has a right to choose their own government” or “suppression of freedom of speech is a human right violation” the impact is much broader than “dont show information about these sets of molecules”