YACHT turned three years of machine-generated fragments into an album

What happened

On 30 August 2019 the Los Angeles trio YACHT released Chain Tripping, their seventh studio album and third for DFA Records. The label's own release notes state the band spent nearly three years building a machine-learning songwriting process around their own 82-song back catalogue, working with outside engineers and deep-learning specialists rather than an off-the-shelf tool. The album's Bandcamp page credits the three members, Claire L. Evans, Jona Bechtolt and Rob Kieswetter, and places the recording in Los Angeles and Marfa, Texas.

What the documents say

The DFA notes quote Evans saying the band wanted "to interrogate technology more deeply" by working "from the ground up". They describe the mechanism: custom models generated "extensive, seemingly endless fields of machine-generated music and lyrics", and the band then spent most of the three years, in the label's words, "painstakingly stitching meaningful fragments of plausible nonsense together". That separates two labours a loose account of "AI music" tends to merge: a model proposing raw material, and a band selecting and rewriting it into songs. The notes name one external tool, Google's NSynth, used on two tracks, and quote Kieswetter crediting the process with breaking the band out of "a mold of concise, formal, four-bar patterns" into longer riffs. Evans is explicit: "We didn't set out to produce algorithmically-generated music that could 'pass' as human. We set out to make something meaningful." Neither document names a chart position or award.

Why it matters for makers

The mechanism worth naming is curation as authorship. A model trained on a catalogue does not write a song; it proposes fragments at a volume no single writer could produce alone, nearly all unusable. YACHT's account states the creative labour was building the process and then choosing, cutting and reassembling its output against the band's judgement of what counted as meaningful rather than merely novel. For a working songwriter, training a model on your own material does not remove the editing pass; it relocates it, from a first draft to a much larger, stranger set of drafts that still needs a human ear.

What to check before you use it

A maker attempting something similar should establish, from project notes rather than assumption: what material trained the model (here the band's own catalogue, sidestepping rights questions a model trained on someone else's recordings would raise); which parts of a track are selected-and-edited generated output versus separately written material, since that affects who can be credited; and whether a named tool such as NSynth carries its own terms for commercial release. This is an editorial checklist, not a requirement stated by either source.

• Whose recordings or compositions actually trained the model being used?
• What share of the released track is selected-and-edited generated material versus separately written or performed material?
• Does any third-party model or synthesiser used in production carry licence terms that affect commercial release?

Chain Tripping resists the two easy stories about AI music: that a machine wrote it, or that it was irrelevant. The band's materials describe a process where generation supplied volume and strangeness, and three years of human selection supplied the song.

Continue in Listening & research, or NSynth shipped as an open dataset and model, not just a demo.