The essence of aligning large LLM models is basically to conform to the existing social order. ChatGPT respects China's order, while China's open-source models, although reluctantly, also tolerate the order in the US. Maybe we can build a revisionist SFT dataset.
Any conversation
a constant, existential pull
so old I couldn't tell you the end
or draw its shape
between kind acceptance
and striving, struggling, writhing, climbing
learning, growing, discontent
"do you feel this way, too, at endless war with longing itself?"
but the birds and trees and ants and wind
in their own marvelous design
don't reply


