lumpCuts: AI Poetics

lumpCuts is a series of handmade short-run poetry pamphlets produced using a GPT-2 instance trained on ‘The Lump’, a messy dataset comprising ‘every poem I’ve ever written ever’. To date four sequences have been made, which you can read in their entirety in pdf form below:
lumpCuts001: night burns (16 degrees) (edition of 16)
lumpCuts002: hbd sebastian (12 dumps) (edition of 2)
lumpCuts003: short dance (20 steps) (edition of 12)
lumpCuts004: spring event (28 events) (edition of 12)
Sooner or later I’ll get round to writing some critical reflections on the process. In the meantime, here’s a text from an Instagram post describing my completion of The Lump:
A few weeks ago I decided to gather together ‘every poem I’ve ever written’ into one big dataset and train an AI language model on it so as to generate ‘new’ ‘poems’ ‘by’ ‘me’. One of the largest AI models is trained on a vast dataset known as ‘The Pile’; I’m affectionately (and a little ominously) calling my little dataset ‘the Lump’ (or, in my more pretentious moments – and if the Lump is anything to go by, these moments are numerous – ‘der Geist’). To assemble it, I first combed through all my hard disks and cloud storage for everything vaguely poem-y and collected it all in one document, then I transcribed ~70,000 words from ~80 notebooks spanning ~22 years and added them to the document too. And as of the end of last week, the Lump is now COMPLETE. It comprises ~2300 individual ‘pieces’, totals ~210,000 words, and runs to 894 A4 pages in 12pt Garamond. The earliest extant piece is some lyrics I wrote for a digital hardcore band I was in when I was 14; the most recent is a commission I’ll be performing at Push the Boat Out at Summerhall next weekend (which, fittingly, includes some AI-generated text – well meta). Most of the contents are, inevitably, ABSOLUTE PISH. I had zero quality control – basically, if it had line breaks, it went in, regardless of its cringe/awfulness (and why I used to put arbitrary line breaks in my journal entries like a pound shop Allen Ginsberg I do not know, but I’m glad I grew out of it, typing that shit out took aaages). It’s still a *tiny* dataset – less than 1/800,000 the size of the Pile, it very humblingly could fit entirely on a single 1.44mb floppy disk. But aye, the idea isn’t to outsource my poetry creation to the machine (though I find that idea fun and funny), but ultimately use the outputs as prompts in a kind of present-self + (past-self × machine) hybrid collaboration – less AI-as-posthuman-rapture than AI-as-elegy/false-memory-appendage! It will probably be a colossal failure but that’s ok it’s a learning experience – and poring over the (so far completely unedited) results and ‘studying its habits’ I’m finding waaay more compelling than is healthy lol.