To get started with something like this might be as simple as training a model on mostly science and finding a novel way to poke around in the latent space?