AI can be cool. AI can be strange. AI can be promising, or frightening. Here’s AI at totally uncool and not frightening at all – bundled with a crappy MIDI keyboard, for … some … reason.
Okay, so TL:DR – Amazon published some kinda me-too algorithms for music generation that were what we’ve seen for years from Google, Sony, Microsoft, and hundreds of data scientists, bundled a crap MIDI keyboard for $99, and it’s the future! AI! I mean, it definitely doesn’t just sound like a 90s General MIDI keyboard with some bad MIDI patterns.” “The machine has the power of literally all of music composition ever. Now anyone can make musiER:Jfds;kjsfj l; jks
Oops, sorry, I might have briefly started banging my head against my computer keyboard. I’m back.
This is worth talking about because machine learning does have potential – and this neither represents that potential nor accurately represents what machine learning even is.
If at this point you’re unsure what AI is, how you should feel about it, or even if you should care – don’t worry, you’re seriously not alone. “AI” is now largely shorthand for “machine learning.” And that, in turn, now most often refers to a very specific set of techniques currently in vogue that can analyze data and generate predictions by deriving patterns from that data, and not by using rules. That’s a big deal in music, because traditionally both computer models and even paper models of theory have used rules more than they have a probability. You can think of AI in music as related to a dice role – a very, very well-informed, data-driven, weighted dice role – and less like a theory manual or a robotic composer or whatever people have in mind.
Wait a minute – that doesn’t sound like AI at all. Ah, yes. About that.
So, what I’ve just described counts as AI to data scientists, even though it isn’t really related very much to AI in science fiction and popular understanding. The problem is, clarifying that distinction is hard, whereas exploiting that misunderstanding is lucrative. Misrepresenting it makes the tech sound more advanced than arguably it really is, which could be useful if you’re in the business of selling that tech. Ruh-roh.
With that in mind, what Amazon just did is either very dangerous or – weirdly, actually, very useful, because it’s such total, obvious bulls*** that it hopefully makes clear to even laypeople that what they claim they’re doing isn’t what they’re demonstrating. So we get post-curtain-reveal Oz – here, in the form of Amazon AI chief Dr. Matt Wood, pulling off a bad clone of Steve Jobs (even black-and-denim, of course).
Dr. Matt Wood does really have a doctorate in bioinformatics, says LinkedIn. He knows his stuff. That makes this even more maddening.
Let’s imagine his original research, which was predicting protein structures. You know what most of us wouldn’t do? Presumably, we wouldn’t stand in front of a packed auditorium and pretend to understand protein structures, if we aren’t a microbiologist. And we certainly wouldn’t go on to claim predicting protein structures meant we could create life, and also, we’re God now.
But that is essentially what this is, with music – and it is exceedingly weird, from the moment Amazon’s VP of AI is introduced by… I want to say a voiceover by a cowboy?
Summary of his talk: AI can navigate moon rovers and fix teeth. So therefore, it should replace composers – right? (I can do long division in my head. Ergo, next I will try time travel.) We need a product, so give us a hundred bucks, and we’ll give you a developer kit that has a MIDI keyboard and that’s the future of music. We’ll also claim this is an industry first, because we bundled a MIDI keyboard.
At 7 minutes, 57 seconds, Dr. Wood murders Beethoven’s ghost, followed by at 8:30 by sort of bad machine learning example augmented with GarageBand visuals and some floating particles that I guess are the neural net “thinking”?
Then you get Jonathan Coulton (why, JoCo, why?) attempting to sing over something that sounds like a stuck-MIDI-note Band-in-a-Box that just crashed.
Even by AI tech demo standards, it’s this:
Deeper question: I’m not totally certain what has earned us in music the expectation from the rest of society that, not only is what we do already not worth paying for, but everyone should be able to do it, without expending any effort. I don’t have this expectation of neuroscience or basketball, for instance.
But this isn’t even about that. This doesn’t even hold up to student AI examples from three years ago.
It’s “the world’s first” because they give you a MIDI keyboard. But great news – we can beat them. The AWS DeepComposer isn’t shipping yet, so you can actually be the world’s first right now – just grab a USB cable, a MIDI keyboard, connect to one of a half-dozen tools that do the same thing, and you’re done. I’ll give you an extra five minutes to map the MIDI keys.
Or just skip the AI, plug in a MIDI keyboard, and let your cat walk over it.
Translating the specs then:
- A s***ty MIDI keyboard with some buttons on it, and no “AI.”
- Some machine learning software, with pre-trained generative models for “rock, pop, jazz, and classical.” (aka, and saying this as a white person with a musicology background, “white, white, black-but-white people version, really old white.”)
- “Share your creations by publishing your tracks to SoundCloud in just a few clicks from the AWS DeepComposer console.”*
Technically *1 has been available in some form since the mid-80s and *3 is true of any music software connected to the Internet, but … *2, AI! (Please, please say I’m wrong and there’s custom silicon in there for training. Something. Anything to make this make any sense at all.)
I would love to hear I’m wrong and there’s some specialized machine learning silicon embedded in the keyboard but… uh, guessing that’s a no.
Watch the trainwreck now, soon to join the annals of “terrible ideas in tech” history with Microsoft Bob and Google Glass:
https://aws.amazon.com/deepcomposer/
By the way, don’t forget that AWS is being actively targeted right now by the music community with a boycott. Maybe they were hoping for a Springtime for Hitler-style turn-around, like if this is bad enough, we’d love them again? Dunno.
Anyway, if you do want to try this “AI” stuff out – and it can really be interesting – here is a far more comprehensive and musically interesting set of tools from rival Google:
https://magenta.tensorflow.org
Now back to our regularly scheduled programming of anything but this.
AI: I am the button.
Deja vu all over again. (At least Microsoft Songsmith you can laugh at. I mean, through the screams.)
Just remember, what humans have that no AI can replicate is agency. So they take terrible tools like Songsmith, and generate ingenious ideas like – using them to make musical sonifications of the stock market crash. (Thanks, Johannes.)