The goal is to create tools that allow composers to offload interesting aspects of their work, like expressive timing and musical texture.

The challenge is to make our models good enough and easy to use so that even non-expert musicians can find them genuinely musically engaging, focusing on generating interesting sounds rather than just melodies.

We're gathering valuable feedback from artists using our tools, which is guiding us to improve the usability and fluidity of our music generation tools, especially for integration with software like Ableton.

To make music interesting, we need to incorporate hierarchical patterning and conditional models that can generate variations, such as chord changes, to create more engaging compositions.

Musicians with strong improvisational skills have been able to create interesting results using our primitive sequence generation tools, turning it into a collaborative game rather than just a tool.

We haven't yet created a model that can generate an entire song because we haven't perfected the conditional aspects of the LSTM yet, which is crucial for handling long-term musical structures.

We're currently working on a project called Henson, which aims to get deep learning models to generate new sounds, focusing on improving the quality and usability of the generated audio.

Generative models have evolved significantly, moving from basic attempts at generation to real-time applications that can create impressive outputs, such as artistic style transfer.

We're entering a new era of creation with combined audio and video generation that feels incredibly realistic.