In this talk, I will present the recent work of the music team at Sony CSL Paris revolving around music generation, representation learning and information theory. I will mention recent developments in Diff-A-Riff, a latent diffusion music accompaniment generation model that is particularly designed for music production use cases. It produces high-quality 48kHz audio stems that fit a given musical context and can be controlled by either audio references or text input. Diff-A-Riff is relatively lightweight because it builds upon Music2Latent, a Consistency Autoencoder that achieves a 64x compression ratio for musical audio. The generative decoder of Music2Latent produces high-quality reconstructions by filling in information that gets lost during compression. Meanwhile, a second generation of Music2Latent has been developed and I will explain its novelties. The representations of Music2Latent are also utilized by an autoregressive transformer model that can generate continuous-valued sequences. That’s based on an exciting new method that I will also include in my talk. Furthermore, I will present novel developments in Stem-JEPA, a self-supervised learning model that is trained to assess the compatibility between stems and mixes, potentially extending the range of data that can be used to train Diff-A-Riff. Also, he will introduce a method to control the information content of generated symbolic music using beam search, which is applicable to any autoregressive language model, as well as a method to estimate surprisal in musical audio.