Slow, repetitive, contemplative, and noisy. Repeatedly-applied neural reconstructions fail slowly and imperceptibly as feature space is explored. This performance uses the 'I Am Sitting in a (Latent) Room' improvisation system. Inspired by Alvin Lucier's 'I Am Sitting in a Room' and the general process of degrading sound by repeatedly passing it through an acoustic medium, it is a system that allows the improvisers to interact with the process of degradation in real time. Using a bespoke variational autoencoder (VAE) model, a 25-second audio clip is repeatedly encoded and decoded through two parallel instances of the model. On top of this process, the performers take on the role of improviser. By manipulating the model's latent embeddings of the audio in real-time, they explore the latent space (or ""room"") of the model over the course of the performance.