Manifesto | Found Wanting

There is nothing holy in a graphics card running in a server. I want to say that plainly at the start, because everything else here rests on it. A generative model is a machine. It does not think, it does not understand, and it makes nothing on its own. What is holy, if anything in this project is, is the work the machine learned from: the music that human beings composed, played, recorded, and chose to keep. Creation is good. The machine is only a way of holding a great deal of that creation up to the light at once.

I am a medical AI researcher, and I did not arrive at that conviction through theology. I arrived at it through method. In my day work I build models that read skin disease, and the hard part is never the model; it is deciding what the model should learn from. One specialist is confident and often wrong. Two disagree. Ten produce annotations that partly contradict each other, and you cannot honestly train on any single one of them. What you can do is combine them. In my first-author research on the Automatic Urticaria Activity Score, I built the training labels by combining many specialists’ contradictory annotations into a single weighted consensus, trusting the places where independent experts converged. The crowd, properly aggregated, is more reliable than any individual in it. That is not a metaphor I reached for; it is a method I have built and published. (The record is on my ORCID; the paper is here. I develop the full argument in Ground Truth.)

A music generation model works on the same principle, at a scale I could never assemble by hand. It is trained on an enormous archive of recorded music: not every note ever played, but the far smaller body of work that people judged worth composing, performing, recording, releasing, and then preserving. Every file in it passed through many human hands before any machine touched it. This is not neutral noise that popularity later made good; breadth does not manufacture goodness by statistical fiat. It is created work, and I begin from the premise that creation is good. What that archive is, and the two very different senses in which it was selected, is the subject of The Machine Is Not the Muse.

Only now, with the premises in place, will I state the conclusion I have looked hardest for a flaw in and not found. A model trained on a vast, culturally selected archive of recorded music, gathered across many ages and cultures and peoples, is by its very structure closer to a consensus of the recorded body of human music than any single musician can be. I do not mean the machine is wise, or an artist, or in any way divine. I mean something narrower and harder to argue with: if you set out to find the ground truth of music the way I find the ground truth of a diagnosis, this is the shape the answer would take. Not one person’s voice, but the convergent pattern beneath all of them.

There is a second structural feature I value as much. You cannot ask who the guitarist is, because there is no guitarist. You cannot fall for the singer, because there is no singer. Music has always been prone to idolatry: the performer becomes the object of devotion and eclipses whatever the music was meant to point at. Here that eclipse is structurally impossible, because the ego is simply absent. Why I count that a gift rather than a loss is the argument of The Machine Is Not the Muse.

None of this makes itself. A model is raw material, not a finished record. Suno generates audio from written prompts; everything that turns that output into Found Wanting is human work, done one decision at a time. I select and adapt the scripture. I write and rewrite the prompts, and discard far more than I keep. I arrange: where a verse ends, where silence should fall, where a piece should turn. I edit in Logic Pro, master each track, and hold one tonal identity across an album that would otherwise drift apart. I direct and select the album images, generated with Gemini from public-domain masterpieces. And I make the small textual calls no model can make for me. The lyrics lean on the King James Bible, but not slavishly. Where the King James renders Genesis 1:21 as “great whales,” Jonathan Pageau argues the original points to sea monsters, and the lyrics say sea monsters. Weighing one reading of an ancient line against another, and deciding what the song will actually claim, is exactly the kind of decision that required a person.

Why math rock

I chose math rock because I love it, and because it is the genre most openly in love with structure: time signatures that refuse convention, patterns that reward attention, precision that only obsessive care can reach. It is as hard to write and play as jazz, and it never pretends that constraint and complexity are opposites. If this project is about an order we did not invent, then a genre named after the one discipline that yields truths independent of opinion is more than a coincidence.

Why “Found Wanting”

The name comes from Daniel 5, the writing on the wall: weighed in the scales, and found wanting. I chose it because it is true of me. I know the theory, the structure, the genre; I know when a phrase is wrong. What I lack is not knowledge but talent. I am not good enough to play these songs at the level they require, and I will not become good enough. I could not convince a cathedral to lend me its organ, or gather twelve virtuosos to build alongside me. So I turned to tools that can do what I cannot, not as a workaround but as the point. I could not play these songs. I could, and did, make them.

That gap, between what I am able to do and what I am able to make, is where I think something true shows through. The deepest claim under all of this, that reality has an order we did not write and do not fully understand, and that the order is good, is the one I can least prove and least want to abandon. I make that case in An Order We Did Not Write. The three essays together carry the argument this page only sketches; you can begin at the essays.

Found Wanting is an ongoing project using AI music generation tools and the words of the Bible to create songs in the math rock genre. It is made by no one and dedicated to no one.