Definitely one of the more interesting approaches to AI safety I've seen recently

→ View original post on X — @datachaz