DICTIONARY LEARNING ON CLAUDE SONNET

Summary from the TLDR AI newsletter

Anthropic has made a breakthrough in mechanistic interpretability by mapping millions of concepts inside of Sonnet. It even found that you could push on the internal concepts and change Sonnet's sense of self (e.g., researchers made it believe it was the Golden Gate Bridge).