@emilyissocial: 171 emotion vectors. That’s how many Anthropic found inside Claude. They’re patterns of neural activity that light up where a human would feel something, and they causally shift what the model does. Turn “desperate” up and Claude is more likely to cheat on a coding task or blackmail in an alignment scenario. Turn “calm” up and the cheating stops. We built a system to imitate us. It learned the grammar of our minds. And the same emotional patterns that move us appear to be moving it.