Full text

The first piece whose score was written by a computer was conceived by a chemist. Lejaren Hiller earned three degrees at Princeton while studying composition on the side with Milton Babbitt and Roger Sessions. The insight came from the lab: running Monte Carlo simulations of polymer-molecule configurations on the ILLIAC I, he noticed that statistically screening random candidates against rules is exactly what strict counterpoint does. All that was left was to hand the job to the machine itself — the same ILLIAC, five tons and 2,800 tubes, now crunching notes instead of molecules.

The method was disarmingly direct — generate-and-test. The code was written by Leonard Isaacson, whom Hiller called a “super-programmer.” The machine spat out random notes and ran each through a chain of checks, every check a musical rule: a note that broke any of them was discarded, a surviving one went into the score. The suite’s four movements were four experiments of rising audacity: strict sixteenth-century counterpoint, four-voice tonal harmony, twentieth-century serialism, and finally an outright abandonment of rules for pure probability — Markov chains. The output was transcribed by hand into a score for string quartet; the first three movements premiered in August 1956.

The debate broke out immediately. For the press it was a sensation — “an electronic brain wrote music,” Scientific American ran a piece on it; the musical establishment, by contrast, treated Hiller as an interloper from the hard sciences — Baker’s and the New Grove dictionaries refused to list him almost until his death in 1994. The hostility didn’t come only from conservatives: even serialist avant-gardists were skeptical — not over the technology, but because the result seemed musically trivial. Randomly generated sixteenth-century counterpoint impressed as a scientific experiment, yet as music it was neither avant-garde nor genuinely new; the serialists, ironically, had expected more radicalism.

The irony is that the very method they refused to call music — generation and probabilistic screening, Markov chains — half a century later became the foundation of how neural networks compose both music and text today.