What’s this then?

Playlisting algorithms are imperfect.

We’re fast entering an era where machine learning and algorithmic ‘prediction’ services are monitoring, shaping, and prompting our behaviour like never before. In the musical realm the role of ‘gatekeeper’ has swiftly transitioned from the preserve of record shop employees and music journalists to Spotify’s playlists like ‘Release Radar’ and ‘Discover Weekly’. Yet, if you’re a voracious consumer of music you’ll find that rather than finding your horizons limited as your recommendations nestle you inline with ‘similar’ users, your musical world has opened up. At first glance it would appear that finding a good playlist is easy and that playlisting algorithms have been perfected.

So why the downer on algorithmic predictions?

What sets human created playlists apart from machine learning? Only humans can ascribe value, because it is conferred through taste. Machine learning software can suggest “You might like ‘Love Is The Drug’ by Roxy Music”, and it can also suggest “You might like ‘Love Is The Drug’ by Grace Jones”. These suggestions are equally weighted. Only a human can suggest “You might like this Roxy Music song, but you should really hear the Grace Jones version. It’s one of the best cover songs ever!

Why is this distinction important?

Society is built on common knowledge and assumptions. The development of this knowledge requires taught histories that create shared assumptions and common truths over subjects that are often, at their heart, completely subjective. Gatekeepers, those with ‘the knowledge’, act as a conduit in the sharing of ‘common knowledge’ through their privileged positions (there is, clearly, a reason why working in your local record shop is cool but your local supermarket less so). It is through arbiters of taste that we build communities, subcultures, and fan groups. Without them, musical culture weakens even as boundaries and horizons spread.

Thus you find a human created repository of ‘taste’ here in Slow.Fine.Fast.

The finest machine-learning based tastemakers are currently predicated on human control. See the Music Genome Project that precipitated Pandora or the value conferred on Spotify’s algorithmic playlists that are given equal billing to the hugely successful ‘Rap Caviar’, one of the very best Spotify playlists, which was originally created by (the very human) Tuma Basa. Whether at the input or output stage, real people with real knowledge and shared histories have developed real taste that has gone on to define the finest playlists we listen to, share, and create.

I have long been fascinated by the way we can interpret subjective musical characteristics into objective distinctions. My own musical journey is marked most heavily by my time helping to develop the industry leading ‘Energy Slider’ at Open Ear Music. Through the creation of proprietary musical markers, users can choose their energy level and instantly get a new playlist, created automatically from their assigned music. Elsewhere I’ve found myself working as an under employed music journalist, a radio DJ, and in a major label marketing department. I have found myself, on a Thursday morning, asking myself what makes a good playlist.

So what makes a good playlist?

Ultimately, I find the same qualities that confer value in music to be the same qualities that determine what makes a perfect playlist. Each must meet expectations, but also defy those same expectations. This is the mark of any great playlist, DJ, or even song. Call it ‘flavour’, ‘depth’, ‘nuance’… musical taste begins when we do away with formulas and set patterns and throw emotion, instinct, and contradiction into the mix. It’s something machine learning has yet to grasp.

Until that time, I present: Slow.Fine.Fast. a repository of human created playlists from tastemakers I trust, grouped by whatever takes their fancy. There is no claim that these will meet your tastes, nor that the taste here is somehow better than anyone else’s. Only that the music collected here was put together with taste. Until machine learning can distinguish based on taste, playlisting algorithms will not be perfected. Until then, we have Slow.Fine.Fast.


Mike – Curator.