MIDIArchive

Exploring AI with the sites and sounds of the early web

Archives seem to have come into a kind of fashion as of late. In an age of planetary-scale existential precarity, they seem to offer solace in feeling knowable, certain, and grounding. The training procedures of AI foundation models like GPT have a close relationship with archives too, but instead rely on their accessibility, volume, and givenness, which allows archives to be composed and instrumentalized as training sets. I developed this project during my residency at Recurse Center in the fall of 2023 as an exploration of the opposing, and yet mutually reinforcing dynamics between building an archive and building a machine learning model.

The MIDI files collected here and used to train the model were once very new. In presenting them here alongside the output of a machine learning model, I hope to bridge epochs of technological transformation. Within each, the possibilties of new aesthetic experiences interact with the technics of producing and distributing new forms of media, producing artifacts that carry hopes and fears about how we ourselves may change.

This project also owes a debt to BitMIDI, which also developed the MIDI player I've implemented on this site with the FreePats General MIDI sound library for your listening pleasure. And if you have a favorite MIDI website that would have a good home with this archive, please do get in touch!

Neural Net

The neural net model developed and trained for use in this project is an 800k-parameter decoder-only transformer, intended to express something at once whimsical and general about the underlying archive. It is intended to serve an educationl purpose, and does not represent the state of the art in 2023, nor is it intended to be used seriously as a tool for music creation.

It currently runs as an AWS Lambda Function at noon (GMT) every day to produce a piece of music around three minutes in length. Each time the model runs, it uses tokens generated from the previous pass to produce subsequent tokens, and in this way produces a single continuous piece of music that has no end, but is punctuated by a 24-hour cycle of rest, and is not presented here in its entirety. Check back tomorrow to hear new music generated by the model!

"I listened to the music. It was hideous. I have never heard anything like it. It was distorted, diabolical, without sense or meaning, except, perhaps, an alien, disconcerting meaning that should never have been there. I could believe only with the greatest effort that it had once been a Bach Fugue, part of a most orderly and respected work." - Philip K. Dick, The Preserving Machine

Click to listen to the most recent output from the model

John Sankey - Harpsichordist to the Internet

This collection of MIDI files is hosted on Dave's J.S. Bach Page, and it covers a portion of John Sankey's MIDI recordings. For a time, he was known as the Harpsichordist to the Internet ...