Skip Navigation

Pipe-dream of the day

OK, so Shuffle is great, and the so-called ‘Smart Shuffle’ is a nice riff on an old staple, but I think we’re missing the bigger picture here. We’re missing an opportunity to make our computers and portable music players more like our clever personal assistants than assembly-line drones, and an opportunity for math geeks to really shine.

Sounds weird, I know. Stick with me for a sec.

The problem with Shuffle is that you’ll eventually hit something that doesn’t match your expectations. In the old days of shuffling CDs you had pretty explicit control over what you heard; in fact, there was no other way about it. If I loaded Punk-O-Rama volumes 1 through 6 into a stacker and hit the Shuffle button I could be damned certain I was going to get a hundred and thirty four whiny punk rock tracks in a row; if I wasn’t in the mood for that I would’ve loaded something else. But Shuffle really doesn’t work that way any more: on any given occasion I’m likely shuffling my entire music library at once. I could get Pantera followed by Jewel… and who wants that kind of mix?

The simple solution, the ‘sensible’ solution (read: the unimaginative solution), is in carefully-constructed playlists. Create every conceivable interrelation of musical themes you could possibly enjoy and move through lists those depending on your mood. Upside: there’s no way it can go wrong. Downside: it’s no different to shuffling a handful of CDs… except this time you had to compile those CDs yourself. Too much work, not enough sexy.

The next solution on the ladder lies in faceted genres and better tagging systems for our music. This makes the relationships between tracks more complex, creating “track families” and “track peers”, and gives us the ability to filter through them without explicit playlist creation. It shifts part of the workload onto the programmer (tagging is internet buzzword of the year, next to AJAX, and we still don’t have it in our music libraries), but everyone would still need to expend a lot of energy organizing their music. So while I’m personally anal enough to do that and love every minute of it, it’s still a lot of work for normal people.

What would really make a difference, moving forward, is machine learning. Bayesian algorithms. Buzzword-compliant music exploration strategies. Marketing fluff for what is essentially the nerdiest thing in the world being used to facilitate the coolest thing in the world: a personal DJ that figures out your mood after the first few songs, and sticks with the theme until you change your mind.

Turn your iPod on and hit Shuffle All. Korn’s Y’All Want A Single starts blaring through the earbuds, which is great music for smashing bottles to, but you aren’t in the mood. Skip to the next track. That skip doesn’t just move you forward, though, it register’s with the iPod that you aren’t ready for that kind of music today and adjusts the upcoming playlist accordingly. Nothing too hardcore, no metal, and definitely no Korn, Limp Bizkit, or any of the myriad soundalikes that flood the charts. So how about the Dixie Chicks with Godspeed? No? Coldplay’s Clocks? Sure, let’s go with that. So now your iPod knows you’re in the mood for some mellow alternative …stuff, and it’ll keep working to make you happy.

The real magic happens when, after you’ve expended all the tracks stereotypical to your favored style (though honestly, just because I’m in the mood to hear Clocks right now doesn’t mean I want to hear three Coldplay albums back-to-back, which is something else it needs to account for), the player starts branching out. After a few hours of listening, I might be back to Dixie Chicks, but I chose the path that brought me from alt to country today: I’m ready to hear it now.

Far fetched, I know. Impossible, certainly not. Back-end it with information gleaned from your own hand-constructed playlists, song metadata, play metadata, and clicks to the all-important Next Track button —clicks that say “not this song, not right now”— and you have scads of song relationships to work with. Like all machine-learning applications it needs to be taught what you like, but the end result is magic: Kahlua goes with milk, and Kahlua goes with Coke, but milk and Coke don’t mix.