#mechanistic interpretability

If the AI Industry Followed Its Own Research, It Might Have⭐10

Anthropic's CEO says AI safety requires understanding how models "think." Interpretability research suggests they don't—and that by its own logic, the industry ...