Anthropic found a hidden space where Claude puzzles over concepts

Anthropic researchers have identified a method to map internal neural activations in the Claude 3 Sonnet model to specific concepts, such as cities, people, or scientific ideas. By isolating these features, the team demonstrated they could manipulate the model's behavior to increase or decrease the prominence of certain topics during output generation. This advancement offers a new approach to model interpretability, potentially allowing developers to better understand and govern how large language models represent information internally.