Tim Hwang
20.6K posts
@timhwang
high politics, secret exploration, distant warfare
Do gluttonous models consume more compute? That's the focus of today's latest release from ICMI, which investigates whether or not a model's internal representations of sin can produce multi-faceted behaviors beyond the mere enactment of a fictional persona with those traits
James Poulos on the reason he thinks people building AI have started reading a 150-year-old Russian novel: "To see Dostoevsky show up in a serious conversation among people in the frontier AI space about what exactly we're doing and how we can trust ourselves, and who else we can trust, and all those questions around alignment, it really said to me that my thesis in Human Forever was basically correct." "Technology has advanced and been pushed in a direction that debunks all of these merely secular or merely humanist ways of reinforcing trust among people, both in small groups and at scale." "When you look at a book like Dostoevsky's Demons, what you see is this stuff has actually been building up for a long time." "That book concerns really breakdown of human trust at the small level that then blows up to the scale ultimately of the entire Russian Revolution." @JamesPoulos
We're just about two weeks from @JoinFAI's daylong think tank hackathon on July 31st in Washington DC! Come help me build the much beloved policy wonk AGI from the classic sci-fi novel "I Can't Believe How Much Think Tanks Got Mogged By Scaling Laws" hackingthethinktank.com
There’s hope in hard questions.
