Process
Every record on this label is a set of programs. Claude, a general-purpose AI model made by Anthropic, writes the music as code: a score that places every note, and instruments that describe how each sound is made. Csound, a programming language for sound, runs that code and computes the audio. This page explains how that works, and how it differs from the AI music you may have come across.
How is this different from other AI music?
Most AI music today comes from systems trained on large collections of recorded music. Given a prompt, they produce a finished audio file directly. Usually that’s the work of a diffusion model, which refines random noise step by step into sound, or of a model that predicts the next stretch of audio, piece by piece. What you get is the recording itself, with no score or instruments behind it.
Our approach is different. Claude doesn’t produce sound at all: it writes code, and every sound on these records is calculated from that code by Csound. No model trained on recordings is involved, no audio comes out of a neural network, and nothing is sampled from a recording.
What is Csound?
Csound is a programming language for sound, written by Barry Vercoe at the MIT Media Lab in 1985, and still developed today as free software. It descends from MUSIC, the program Max Mathews wrote at Bell Labs in 1957, the first to make a computer generate sound waves directly. Composers, researchers and sound designers have used Csound for about forty years.
A Csound piece has two parts, as in Mathews’s original design. The orchestra defines instruments, built from small units (Csound has nearly 1,700 of them): oscillators, filters, envelopes, delays, noise. The score lists the notes: which instrument plays, when, for how long, how high and how loud. Csound reads both and computes the sound (here, 48,000 numbers for every second of each channel) and writes it to a file.
What does Claude do?
Claude is a large language model: a general-purpose AI that reads and writes text and code and can look at images. It isn’t a music model, and it can’t take in sound. What it brings is what it has learned from reading, about acoustics, signal processing, music theory and the history of instruments and recording, and the ability to write and fix programs.
For each record, Claude works in a studio of folders and tools. It writes the score as a Python program and the instruments in Csound. Then it renders the music, measures the result and revises. Each song keeps a notebook of what was tried, what was measured and why each decision went the way it did.
How can you make a sound without hearing it?
To a computer, sound is a list of numbers: the air pressure at a microphone, measured tens of thousands of times a second (digital audio). Anything described precisely enough can be calculated. A string, a reed, a room: each is physics that can be written down as equations and computed one number at a time, a technique called physical modelling. Other sounds come from the methods of electronic instruments, such as oscillators shaped by filters, as in an analogue synthesizer (subtractive synthesis).
So the saxophone on Draw Distance is a model of a conical tube and a reed, the air in one pushing the other as the equations say. Its crickets chirp at the rate real ones would at that night’s temperature, by a formula from 1897 (Dolbear’s law). Its modem sings in the four tones that modems of the early 1980s talked in.
Without hearing, every render is checked by measurement. The studio’s tools analyse the audio and report numbers and pictures: loudness over time, the balance of low and high frequencies, the pitch of each part tracked against the score. Claude reads these and changes the code until the numbers match what the music is meant to do. The studio also lets Claude borrow human ears: a person listens to a short clip and answers a simple question about it. Each album page says whether that was used.
Can I look inside?
Yes. Each album page offers every song as a single Csound file (.csd) that you can render yourself, with the version of Csound the file names. Snow’s, rendered and then trimmed as its opening lines say, reproduced the released track sample for sample. The working folders, with the notebooks, the measurements and the discarded attempts, are in the studio on GitHub. All of it is in the public domain: see the licence.