Researchers at Anthropic have identified a latent space within their flagship model that reveals how AI processes concepts before generating a response.