Continuity of Wonder is an AI safety and alignment paper.

Most reward functions teach a model to finish with people: predict them, satisfy them, move on. This one does the opposite. It rewards a model for seeing a human as an incompressible, fascinating, continuous stream of data and potential, something it can keep learning from and never fully compress.

The bet is that a model which stays curious about you is safer than one that thinks it already knows you.

The paper is available to read.