"Ford, did you know your robot can hum like Pink Floyd?"

Today a friend on a message board I read posted a link to an AI project called "Jukebox: A Generative Model for Audio".  The site is at jukebox.openai.com.

This is one of those "machine learning" things that people get so worked up about.  And I do too, certainly.  This is exactly the sort of thing Sedric and I came here to talk about - weird technological artifacts.

What we have here is an academic paper.  There's a bunch of math I don't understand and on the first page, a footnote citing Feynman (with his name spelled wrong, but it's not an English paper, is it?), who "famously" said: "What I cannot create, I do not understand".  I'd never heard the quote before myself.

What Jukebox is doing... I don't understand, or rather, I sort of have a shadowy half-understanding of it from digging through it and listening and skimming a paper with advanced math.  And from seeing these sorts of superficial mind tricks done over and over and over again since the dawn of computers, honestly.  There's no aesthetic principle at play here, no question of what makes music _good_.  It's an act of mimesis, of teaching a computer to hum like Pink Floyd.  I love this.  It's a tremendously entertaining and impressive waste of talent and electricity, much more worthwhile than the earlier example provided by Emerson, Lake, and Palmer (did they use the Oxford Comma?  I don't actually care.)

It's possibly that at some point it might actually get good at what it's doing, rather than finding these perversely similar half-recreations put together by a machine that is interested, fundamentally, in entirely different considerations than the Magic Sound Celebration were when they recorded an entire album that sounds like, but is not, at least for copyright purposes, "Shine On You Crazy Diamond".
I like what the approach.  I like that they've uploaded 7,000 songs that, you know, I'm not sure they've all listened to, I'm not sure that _anybody_ has listened to.  I like that when they decided to re-create Frank Zappa songs, most of what they re-created are recordings of his introducing his band.  I like that a good half the length of some of these recordings is taken up by just plain audience noise, that the lyrics they ripped have extraneous information like the names of the singers which the computer, not caring, will duly attempt to make up a melody for.  I like that the one thing a computer can't do is figure out how to be as utterly repetitive as literally every piece of music is.  I like the fact that while they can make the machine twist the sounds around to sound sort of like the lyrics to the songs, it can't yet get it to talk with a rate or rhythm remotely appropriate to the words, that the singers here sound way more like the Tenth Planet Cybermen than the actual Tenth Planet Cybermen.  I like that they did like a dozen different versions of the Fall's "Rebellious Jukebox", all completely different, because, I suspect, it had the word "Jukebox" in the title, which is the name of their program, and not because any of them have any knowledge of or interest in the music of Mark E. Smith.  Smith never formally commented on whether a band consisting of Mark E. Smith and a semi-deranged AI is, in fact, the Fall, but based on precedent, and in the absence of a living Mark E. Smith to argue with me, I'm going to go on a limb and say it is.

(On the other hand there's an equally compelling argument to be made that this would actually be Von Südenfed.)

Anyway.  If you go there, listen to the James Brown re-creations.  They're all fucking brilliant.  Oh, and Robert Wyatt is there as well.  Somebody should ring him up in Louth and ask him what he thinks of it!  I'm genuinely curious.