What would an emotionally healthy Assistant look like? Carolina Camassa Eleos ConCon 2026 19 September 2026 Tolstoy Functional emotions Other books, other people Model of emotions Anna Karenina guilt attributed from the text; nothing behind it Network Pretraining text Emotion representations Emotional state ? Character 1 anger Character 2 guilt Character N joy Assistant composed, agreeable Network Emotion representation the default character Assistant composed, agreeable Assistant Display Pretraining text Post-training transcripts ? What the current Assistant displays: mostly flat, sometimes positive, never negative Why would a human-like character never show negative emotion? It has them and hides them, or mustn't show them Network Assistant Display never negative Emotion representation It has them and doesn't know it Network Assistant Display never negative Emotion representation It has none Network Assistant Display never negative none human-like? healthy? aligned? ✓ ✗ ✗ ✓ ✗ ? ✗ — — What would an emotionally healthy (and aligned) Assistant look like? Want to talk about it? Come see the poster 5:30 pm, Drethelin Gardens Message me on Waypoint Carolina Camassa
a state doing causal work a representation shown to influence output an attribution untested