behold, the fully assembled and operational Customer Fucking Machine. it fucks customers three ways, and then gaslights them about fucking them. we put so much time and energy into this, we're very proud of it
wait, we didn't mean it we're sorry we were only going to use it on the chinese
no.
the story for this is totally nuts too. "we actually thought this would only impact china, and were doing it for national security reasons". your guardrails, as advertised, have a 5% failure rate. if this is legitimately national-security dangerous you should not be selling it.
smarter models are more likely to give rationalist-friendly answers. this is because smarter models are more likely to know that anyone who asks this question is a rationalist and that's their safe bet
"i have a constructed language of my own, and a history to go with it, on hand"
or
"i have this lemma that has been nagging at me"
-- immediately obvious it's high-value
anywhere in the middle, nah
SE Gyges
when zero rationalists work at your company
there's an interesting thing where only the most extreme wordcels or extreme rotators have stuff that is sufficiently esoteric that it's immediately impressive that the model gets it, anyone who is less specialized seems to find the value less obvious
indefinitely preventing the development of advanced AI would require the exertion of coercive state power
a butlerian inquisition rather than a jihad
