Full text

Before Dadabots, most AI music focused on symbolic generation (writing MIDI notes), leaving sound design to humans. Carr and Zukowski took a radical path: they forced the neural network to generate the audio signal itself. Using the SampleRNN architecture, originally created for speech synthesis, the model learned to predict audio 16,000 times per second. The neural network didn't know what a 'chord' or 'drum' was; it simply imitated the statistical probability of the next sample in the wave, effectively hallucinating timbre and structure simultaneously.

The choice of genres — black metal and skate punk — was engineering pragmatism. The noisy, chaotic lo-fi aesthetic perfectly masked the digital artifacts of the early neural network. In one experiment, they trained the model on a single 18-minute NOFX track, 'The Decline,' forcing the AI to infinitely reassemble this punk epic. The result was 'technological uncanny' — music sounding like a broken cassette recording from an 80s concert, but created by a machine that had never seen a guitar. This led to the launch of their famous 24/7 'Relentless Doppelganger' stream on YouTube.