Shaping AI Benchmarks with Together AI Co-Founder Percy Liang

Musical Roots

Percy Liang shares his deep connection to music, particularly classical piano, which he has played since he was eight. He reflects on how his musical background has influenced his career in AI, noting that he has always sought ways to merge his passions for music and technology. Percy mentions his collaborations with postdocs Chris Donahue and John Thixten, who focus on music and AI, leading to exciting developments in music generation models 1.

For a long time I was thinking, how can I combine the AI work and the music works?

--- Percy Liang

Despite advancements in AI-generated music, Percy emphasizes that the personal, subjective experience of playing music remains irreplaceable by automation 2.

   

AI Music Creation

Percy discusses the technological strides in AI music generation, highlighting the development of the anticipatory music transformer. This model allows for more controllable music generation, enabling musicians to edit and change compositions rather than just generating music at the push of a button 1. He explains that even smaller models, like those the size of GPT-2, can produce compelling music, especially when working with symbolic music formats like MIDI.

We are training foundation models on music. And this is extremely exciting to me.

--- Percy Liang

Percy also touches on the potential for AI to augment human creativity, allowing musicians to translate their ideas into music without needing extensive practice time 2.