(Most of the content here is cribbed from a talk my partner in crime Chris Kucharczyk and I gave to the Chicago Chapter of the American Statistical Association Conference)
A happy hour conversation in which a colleague asked "how could you use AI to make a band?"
set in motion a number of experiments which would change the world be a lot of fun.
To explore this question, we decided to break down a band into 4 components
Band Name
Fortunately we had a head start on names. A number of our IDEO colleagues had running lists of potential band names (and even some Twitter accounts dedicated to putting those out into the world) that we were able to borrow, including such gems as "Dumpster Croissant" and "Meat Dreams".
Band Bio
With names in hand, we turned to the task of coming up with a story for each fake band. It’s here that we turned to generative language models. In particular, a state-of-the-art-at-the-time model that was still capable of running on our laptops: GPT-2.
We took a pretty simple approach. We decided to create a prompt with the name of the band, and then a brief descriptor of what that band might be. So for example, we prompted the model with:
Dumpster Croissant is a French avant-garde thrash band.
GPT-2 continued this train of thought with:
In 2004, they performed their first album of three-drummed, dark guitar riffs (a riff called "The Fetus Express", and a variation) to the French music magazine La Républications of Culture. The band didn't make it out of the recording process; that would be another matter.
This was pretty mind-boggling to us for a few reasons;
Things only got more weird from there: check out part 2 for more!