Tag
The author used Anthropic's Neural Linear Algebra (NLA) technique on Llama-70B to identify latent 'thoughts' (internal representations) that influence the model's behavior but are not directly visible to standard introspection, highlighting advances in mechanistic interpretability.