Bottom line: Anthropic's own research on finding and controlling personality traits inside their model Claude. Shows both the power and the danger of activation steering.
Anthropic's research on extracting and injecting personality-direction vectors in their own models. Demonstrates that character traits like honesty, sycophancy, and helpfulness correspond to specific directions in activation space — and that these can be monitored and controlled. The company that makes Claude, publishing how to find and control persona traits in their own model. See the Vector Trilogy for the full implications.
Anthropic found that personality traits in Claude aren't just behaviors — they're actual directions in the model's internal space. You can find the "honesty" direction and turn it up or down. The same company that made Claude is telling us this. It's both powerful and unsettling.