Cartographic exploration of steering directions within large language models. Linear directions within the latent space of large language models are investigated by clustering and rendering hundreds of maximally activating prompts. Select a direction below to explore a map of activating concepts centered on the strongest activations.