Chain of thought models can flip back and forth between languages, sometimes producing gibberish before arriving at the right answer, which can be both fascinating and concerning.

These models are learning the principal axes of variance in music, identifying what's most important while having limited memory.

We're rethinking our music sequence generation work, moving beyond simple models to care about expressive timing, dynamics, and polyphony, aiming to create high-quality music generation tools.

I often see each project as an opportunity to explore a new style of music, which keeps my work fresh and engaging.

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.

I'm working on a novel that incorporates text generated by machine learning systems, aiming to create something worthwhile and interesting.

Artists have been sampling from our models and using them as a distance measure to explore creative possibilities, leading to a nice explosion of creativity in their work.

We can observe the model's thought process, such as its consideration of various ideas while composing poetry.

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.