Mechanistic Interpretability
Dario discusses the challenges of training AI models to be aligned and the importance of understanding what happens inside the models at the level of individual circuits. He also explores the concept of verifiability and the potential role of mechanistic interpretability in achieving alignment.In this clip
From this podcast

Dwarkesh Podcast
Dario Amodei (Anthropic CEO) - $10 Billion Models, OpenAI, Scaling, & AGI in 2 years
Related Questions