Empowerment — is minimized instead of maximized by → Penalize influence

Empowerment was developed as an intrinsic reward to be maximized: an agent maximizing mutual information between its actions and future states seeks out positions of capability, avoiding walls, picking up keys, for its own sake, even absent external reward (concrete-problems, §"Penalize Influence:", p. 6). Penalize influence proposes running that machinery in reverse, minimizing rather than maximizing the same quantity, so that seeking capability becomes something to avoid rather than pursue. The paper flags why this inversion is imperfect: empowerment "measures precision of control over the environment more than total impact," so a button that cuts power to a million houses counts as only one bit, while someone merely transcribing the agent's actions counts as maximum empowerment despite having no effect at all (concrete-problems, §"Penalize Influence:", p. 6). The sign flip is clean; the measure it flips is not.