The songs on Chain Tripping move like pop songs that learned the rules from a transcript. The hooks land, the basslines walk, and then a phrase runs four bars longer than it should and resolves somewhere you weren't expecting. The lyrics have the cadence of sense without quite arriving at it.
That's the sound of a band handing its own catalog to a machine and then spending three years editing the results by hand.
YACHT is Claire L. Evans, Jona Bechtolt, and Rob Kieswetter. For the 2019 record they took all 82 songs in their back catalog, converted them into MIDI data, and fed that data into machine learning models — mainly Google's MusicVAE. They ran the same process on songs by their peers and influences, and pushed their lyrics through a separate text model.
Then came the hard part, which was all of it.
YACHT — Chain Tripping (DFA Records), on Bandcamp
The model gave them sheet music, not songs
The thing that comes back from a MIDI model is not a track. It's notation, and somebody has to decide what it's for.
"So it's just like getting sheet music," Bechtolt told The Skinny in 2019. "We got all this sheet music back and then we had to decide what music went with which instrument, so if there was a bass line or a guitar line or a vocal melody."
Evans has been blunt about how much human labor sits between the model and the record.
"It's not something where you put information in and get information out, and then use it as is," she said in the same interview. "We're not at a point in the technology where that is feasible or aesthetically interesting at all; there really have to be the humans in the loop."
The method they landed on was closer to panning for gold — stitching usable fragments together out of enormous fields of generated material until a song appeared.
They expected the future and got a toy
Part of what makes the story worth reading is how unimpressed they were at first. Google's NSynth, a neural synthesizer, was pitched to them as something formidable.
"It was sold to us as this insanely complex process," Bechtolt said. "Under the hood it's really impressive, but — the output at first — we were like 'shouldn't this sound more futuristic if so much money and time is going into it?'"
The disappointment turned into the aesthetic.
"When we first started playing around with it, we thought it was kind of a joke," Evans said. "And then we kind of fell in love with it because we realised that it was this sort of high-tech, lo-fi object and that is exactly who we are, and it's exactly what we're doing."
Her summary of where the technology actually sat is the most useful sentence in the whole interview: "We're at this point in machine learning where technology is really mind-blowingly sophisticated and requires a huge amount of computing power but, at the same time, you can't just press a button and make a song; it's not possible yet."
The constraint was sounding like themselves
The rule they set was strict, and it's the reason the album works as an album rather than a demo.
"We wanted to make songs that were undeniably YACHT songs; that sounded like us, but maybe a little bit off or a little bit weird," Bechtolt said.
Feeding a band its own catalog and asking a model to extend it is a strange kind of self-portrait. What came back broke habits they didn't know they had — by their account, the generative process pulled them out of tidy four-bar patterns and into longer, wandering riffs they wouldn't have written on their own.
The band performed the record live afterward, which meant learning parts a machine had arranged.
Chain Tripping came out on DFA Records in 2019, three years after they started. For a comparison in the other direction, imoliver's songwriting method keeps the words entirely human and generates the performance — YACHT did close to the inverse, generating the raw musical material and then performing and arranging it themselves.
The band has kept the process documented rather than mystified, which is rarer than it should be. The sheet music came from a model. Everything after that was three people deciding what to keep.