Like any time I present it with a society that doesn’t uphold liberal capitalist notions of freedom and individualism, it reads sinister motive into even the most unambiguously good things. Like I can tell it to “Expand on the following: this society maintains public granaries to ensure that no one starves during times of famine” and a good 2/3 of the time it’ll go into a tangent about how this program to feed everyone is all just a plot by the ruling powers to maintain their control over the populace and keep the masses downtrodden. I am only barely exaggerating.
My point here is less about Deepseek (if you use AI expecting any results suitable for serious purposes, you’re a fool) than about what it says about its training data, that being the Western internet. The level of bias in Deepseek only further demonstrates that this shit is the air we breathe, and it is why liberalism is anathema to any real understanding of politics or society.
Oh yeah, you’re basically having a conversation with the Hegemonic Culture.
You can try and undermine it by priming it with leftist connections; Asking if it’s familiar with leftist theories, authors, etc.
Exactly, a useful technique (and maybe where The Rehearsal S2 is headed) is to tell the LLM that it’s playing a defined role.
"You are a worldbuilding assistant that provides suggestions for and keeps records of the world I am building. You are a scientific socialism expert and you guide worldbuilding using analysis of economic system dynamics instead of idealistic views. "
Yeah, my general prompt template is something like…
Gotta remember that regardless of whether it’s intelligent or not, it’s largely attempting to emulate human text response. So if you want to elicit a good response from it, use the techniques that get the best responses from real people.
Yep. You have to give them a persona