36 comments

[ 4.7 ms ] story [ 86.2 ms ] thread
If the DJ is removed the music quality would be the same.
Humans can set the input configuration for the AI and then filter the generated content. AI could generate too much content for anyone to hear, so there's value in human curation.
We might get a good idea of the kind of music AI's enjoy listening to /s
Probably, but DJing is a fun social activity. The DJ puts a bit of themselves in the mix. I think an AI for music mixing isn't really solving much of a problem. It's just a neat trick in my opinion. Making fun and skill obsolete is sad!
If you can't tell the difference between a dj and playlist or AI then the DJ is terrible.
It really depends on the style of music being played and the quality of the DJ.

I’ve heard some DJ really elevate tracks with the way they’ve mixed them. I myself would typically have two or more tracks playing at once (more than 3+ deck mixes when playing on vinyl turntables or up to 5 channels when playing from Ableton) and using that to layer percussion loops over older electronica and industrial tracks to make them sound fresher and more club friendly.

I’ve also heard some DJs completely ruin songs by badly mixing a sparsely engineered track (ie something set at a particular tempo but sounds a lot slower) after something heavier and more pounding, thus sucking the energy out of the second track.

Even with pop songs where there isn’t a whole lot you can actually do with them, I’ve heard some DJs take average or even bad pop songs and drop the hook between some really good songs and do so in a way that really compliments both. Resulting in the otherwise average pop song sounding better than it normally would if it were played in full.

Not to mention there’s DJs who spin their own remixes of tracks. DJs who seek out bootlegs. DJs who bring fx and synths to gigs. And all the other creative stuff going on beyond just spinning other peoples records.

I remember spending a few weeks trying to track down a song I’d heard on a mix tape one weekend, only to discover it was actually two different tracks. an instrumental electro with an a cappella layered over it. It sounded perfectly mixed as if it were originally released like that.

I also remember spending months tracking down a remix of Leftfield’s Space Shanty, which I heard then play on a Radio1 essential mix. Turned out that was a one off pressing they made for a few gigs in ‘93.

DJing is a really diverse scene and those that play the in bars of your local high street pub or the guys that entertain your drunk uncles at your sisters wedding are really only the upper most tip of the iceberg in terms of DJing.

Thanks for the insightful comment, I enjoyed reading it.

That being said, I don‘t think azaras was talking about DJs in general. Just this specific example.

Have a go at mimicking a daft punk setlist, they make it look easy - it's not. The techniques are often simple but art is not about technique
This was... really not good. As anyone who's played around with a grid beatmaker, honestly almost anything you put in there kinda sounds good and groovy, yet honestly most of the stuff there felt less good than average, no better than random.
The way to get good results from AI is letting it randomly generate stuff, then have a human pan for gold
You can say that with randomly generated content, or simple heuristics+random based content, it's not specific from AI
No, that‘s the way to get your paper published. Actually good AI systems produce convincing results without curation.
I agree that the "AI as assistant" model is AI's sweet spot now and likely for the next few decades.

This model is already common now in systems like Tesla's badly-named "Full-Self Driving" Autopilot, which Tesla itself notes "is a hands-on driver assistance system that is intended to be used only with a fully attentive driver".

Similarly, "AI as creator" mostly leads to uninteresting outcomes. In contrast, "AI as creative assistant" will lead to actually-fascinating advances in the creative arts in the near future. AI will help creators think differently, and in turn, creators will weave good and/or interesting output into coherent, higher-quality pieces.

You can probably teach it to start producing higher gold/crap ratios with enough human planners training it?
I assume you are talking about a less obscure app like Live, Bitwig, or FL Studio, but in case not: I vaguely remember a 90s Windows application (possibly shareware) that had a 3D pre-rendered UI with keys shown on screen that looked like a payphone's keypad. You could assign beatmatched loops and hits to any key and toggle or trigger them in a synchronized manner (either with the mouse of the physical keyboard). I had loads of fun as a kid recreating Music Instructor's archetypal/self-aware cliché dance song "Hymn" with it.

Do you or anyone else know which software I'm talking about?

I doubt it was Future Beat 3D, but that's the only one I can think of right now that sounds like what you're describing.
Wow, either that was it, or one of its precursors by a different company, Beat 2000 (I found a review on Future Beat 3D that mentions it): https://archive.org/details/beat-2000

I've been looking for it for quite a long time now, thanks!

Thank you for answering that, sincerely. You've knocked somebody's white whale out today, which is one of the most selfless things you can do. In doing so, you've probably freed up minutes or even hours of their thought for years to come.
I'm not sure it's the same software but your comment made me remember Dance eJay ( https://youtu.be/b1PpXcC8Ik0 ). It was distributed in Sweden in the 90s by a radio channel called NRJ, and I guess it was made like that in a few other countries in the EU. It was so much fun and those samples still bring back a lot of memories.
I wasn't actually talking about any specific app. I'm not a musician, and while I've touched Live/FLStudio, most of my experience is just with toy ones on the web. I've seen dozens over the years with various level of complexity, but the one I remember the most vividly was a flash one called Tone Matrix, here's a non-flash remake:

https://www.maxlaumeister.com/tonematrix/

Honestly just clicking random squares, adding and removing as you go can great unlimited good sounding patterns.

Audiotool itself (the site that hosts Tone Matrix, linked from Max's page) was also very impressive at the time.

I looked around a bit on the googles and it helped me remember a French person's name who had lots of similar Flash experiments on their site; apparently that was where TM originally started: http://lab.andre-michelle.com/

The fact that browsers never managed to come together to build and integrate a proper emulation layer is truly a shame. I know there's quite a few projects out there, but we really need something built-in the browser.
Generative music is absolutely nothing new either.

I have listened to so much generative music the last 25 years and there was one piano piece that was really good and something that felt like a unique style. Not good enough though that I even remember the name of the piece.

Generative music is really no different than what you get when a human starts noodling around with music making software. It is almost literally the same process.

I'm pretty proud of having created some good-sounding AI music. https://soundcloud.com/theshawwn/sets/ai-generated-videogame...

I tried to make it in as many styles as possible. "Crossing The Channel" is a favorite, because the tempo is so unique. I think the GPT made a mistake early on, and then completed the most likely "not-mistake," which caused the piece to have some pretty unique timing.

Maybe someday I'll dig up the model off my old laptop. I thought everyone would be making nice-sounding AI music, but I guess nobody can match a Gwern-trained model. He's the Hattori Hanzo of ML.

(comment deleted)
I have the minority opinion here: I thought the music was interesting, and love the hierarchy of models for generating music given latent random seed vectors and adding the extra layer of models to simulate DJs, under a layer of a human DJ. I once built and worked on a model to generate spreadsheet data given example data using a GAN (generated spreadsheet data could be used to train models that were accurate when tested against real data - that whole experience was an eye opener for me). The architecture of this system is more complex and much cooler than what I did.

I also look at this system as a solid indicator of how humans will interact with AI systems in the future: we will lightly affect and guide automated systems with an internal complexity way beyond what we can understand. There used to be an expression “a human in the loop” and I don’t think this expression really describes future AI/human interactions.

Musical complexity is hard to judge, and the relationship between complexity and "sounding good" is evasive. Consider Satie's Gnossienne #1:

https://www.youtube.com/watch?v=PLFVGwGQcB0

It's ... simple, right? It's just a few really good choices in an unbelievably enormous nearly-Hilbert space. I could listen to that stuff forever.

Most of Nirvana's songs have a similar quality. Judging from where OpenAI Jukebox was a couple years ago, I imagine in a decade we might have AI that sounds as good as Nirvana.

On the other hand, my favorite piece ever is Stravinsky's Rite of Spring. We seem to be very far away from any kind of Stravinsky algorithm.

Ha! I grew up listening to Stravinsky's Rite of Spring, my Dad loved it. And yes, it will be a while before any combination of DL models can write something as touching as Gnossienne, or ‘simple’ Bach stuff that I love like Jesu Joy of Man;s Desire.
Bach is a good example of music that's both simple and complex. He'll take a few (often just one) fairly simple theme, and weave variations of it into a lush fabric that's coherent and natural (yet always surprising) -- the nearest thing to justice one finds in music.
I saw a video with a guy analyzing how well Elvis could play guitar. And ultimately arrives at the conclusion "well enough to express himself musically". Elvis's guitar playing in the video below is pretty simple cords. But it has a purpose. It matches the story he's telling through his music. It's like cinematic movie, the camera moves in a way to help tell the story. So not sure if AI will get there because does AI really understand what the story is. Being simple but purposeful beats complicated and pointless which all of electronic music seems like to me.

https://www.youtube.com/watch?v=GMkyIJ0g5sg

Its like movie scene with a KONG FU FIGHT where the opponents puts on a wheelhouse show with impressive moves. But ultimately get knocked out with one simple effective counter punch when they clash. And ultimately you know who the real master always was.