This paper relates an early art-research collaboration between two practitioners in machine learning and virtual worlds toward new embodied musical experiences of Artificial Intelligence (AI). Instead of a digital music instrument or a music-generating agent, we propose to craft a soundwalk experience where a human person moves through a three-dimensional virtual world to explore a latent sound space generated by deep learning. We report on the diffractive prototyping and iterative crafting of three such soundwalks through/out deep latent spaces, using nn~ and New Atlantis as computational platforms for AI audio processing and virtual world experimentation. We share critical perspectives emerging from our latent soundwalking practice, with the hope that they contribute to ongoing community-wide reflections toward new AI for musical expression.