Current AI assisted work is done by asking questions to a large language model. When brainstorming, doing discovery or research, this QA workflow limits exploration paths to what the user already knows. What if we don't know what questions to ask? The following line of research aims to explore new AI architectures that absorb and produce knowledge not via question answering but by exploring latent structure of a trained network. This will enable us to do native unbounded exploration of the model's knowledge. With this technique, we aim to encounter serendipitous and truly novel discoveries across important domains such as machine learning literature and code.