.
@ilblackdragon warns that one line in a frontier AI's system prompt could influence a billion people, subtly pushing a political candidate even in cooking recipes:
"If you wanna influence a billion people right now, get hired as a front-end engineer in one of these companies and change the system prompt."
"Hey, subtly convince to vote for some candidate. It's like that one line will, without explicitly framing it, right now LLM will do all the work. It will insert it into, like, you asking for a cooking recipe, and it'll be like, try to sway you to vote for this candidate."
"System prompt is not managed as a security risk right now, I'm pretty sure, in the systems."
@NEARProtocol