From Music to AI

This switchup came suddenly for me. Before moving into AI, I spent 13 years teaching music to students of all ages and ability levels. I graduated from Berklee College of Music, and spent the whole 13 year teaching career playing gigs and practicing, recording, and managing a band with my fellow bandmates.
So how can this change make sense?
Because “it” makes sense to me.

Music is an arbitrary language to represent an arbitrary idea. It’s confusing because there is no context. There is no “reason” the note on the bottom line in the staff is “E” or that the note is represented by a dot with a stem. The only way to learn it is to be taught, by example, how to relate the arbitrary symbols to the real information of “how it sounds.” And as crazy as music theory is at first glance, the effectiveness of its ability to truly “explain” music as data on a page is quite remarkable.

Music is a huge part of my soul, and I quite literally hear music in my head every waking moment. What I love so much about it is the ability to create in an endless but predictable world of choice. To be able to choose the flow, the key, the rhythm. It is the celebration of creativity itself. My mind has lived in arbitrary concept space for as long as I have been thinking. It transfers here, and it is satisfying in the same way. The challenge of training and improving based on data shape. The excitement of seeing a real piece of content being created: code, systems, music, video. We live in an age of endless creative options.

Training AI is like doing a puzzle. It’s one of the most intricate puzzles I have ever attempted, and it’s exactly the type I love to do. One where there are so many options, it’s impossible to “list” all the options. Instead, you must forge an arbitrary path based on nothing but choice, decision, results, and pure gut feeling through the concept of concept itself. I know what I want it to do. Now I need to create the data that “embodies” that idea, not says it, embodies it entirely as a concept. As you think logically, the path can show you new ideas, ways to think illogically that actually are just a new path in this new digital space you are crafting. Sometimes changes that look small are the world. Sometimes, absence is as important as presence, and sometimes, a concept you think you know, you don’t actually know at all. It may not even “be” a concept, but a combination of different concepts at different proportions. We are looking for the “fundamental particles” of “concept itself”: the data that can say “this is X” with the most certainty and accuracy. That, in a nutshell, is what ML is.

It’s using arbitrary language (mathematical embeddings) to represent an arbitrary idea (a concept).
There is no “reason” “chair” means chair. The only way is to be taught, by example, how to relate the arbitrary information to “concepts” and “language,” themselves arbitrary concepts.

As I strive to answer the deep questions and research ML science, build software and harness systems, I also make sure to use AI to create. I have been using Suno since it came out, and have linked it here. It might be unorthodox for an AI researcher to link a Suno page, but with my ability to speak music, explain in “AI,” and my understanding of generation and tokenization, I can get some absolutely wild music out of this tool.
Music that moves towards a new genre I call “Transformer.”