Happy Wednesday! Hope you have exciting and/or restful plans for the end of the year, and aren’t getting too full and hungover from holiday party season. :)
This past month has been full of traveling, speaking, teaching, writing, and general cultural and intellectual fence-scaling. In early November, I spent a week in Kristiansand, Norway
for the Roundtable and Sørveiv
conferences, where I took my deepest dives to date into the complicated worlds of copyright, identity and mental health in the music industry. I recently wrapped up a week in Tokyo, Japan
, where I spoke about the state of Japanese streaming services at Tokyo Dance Music Event
and soaked in the local food, music and fashion scenes. I also brought my industry experiences to the front of the classroom for the first time, giving talks about journalism and the state of music/tech to masters students at the University of Agder
in Norway, then to undergrad music students at New York University
As this will probably be my last installment of 2017, I wanted to thank all of you for supporting me through such a wild year. Freelance writing isn’t the easiest job in the world, let alone covering an equally unstable (and exciting!) industry, and I’m so grateful for everyone who’s served as professional and personal sounding boards along the way.
Plus, I’m thrilled that the Water & Music community has now reached nearly 700 subscribers! I’ve had the pleasure of meeting many of you in person through my various travels, and look forward to hanging out with more of you next year. If this is your first W&M issue, please introduce yourself simply by replying below—would love to hear what you’re up to and what brought you to my lengthy rambles!
Today’s subject line is inspired by a piece
in the Harvard Business Review
about how to spot machine learning opportunities regardless of whether you have a data science background. TL;DR identifying opportunities for AI and supervised learning begins with asking what business outcomes are worth guessing
, and whether there’s enough data to do the job.
The authors outline a rough methodology for tackling this question on a company level. First, break apart work activities into daily, process-driven routines versus sporadic, independent assignments or tasks that require more patient deliberation. For the routine tasks, has your company kept a long-enough record of successful runs of these tasks that could be used as a training dataset for an algorithm? Which tasks could involve prediction or classification into different categories? How might a transition to AI/supervised learning change the products your company offers to customers?
Most importantly, what’s the accuracy threshold you’re willing to tolerate? Given a 65% or 70% accuracy rate, what’s the worst thing that an algorithm could do to your business?
I understand why no one in the music industry would want to ask themselves these last two questions. After all, there’s very little that the music industry wants to guess; there’s a lot that the industry wants to know. Who really is my fan? Who really owns what percentage of this song? How much ROI will we really see from this marketing campaign?
In trying to get to the bottom of things, though, I think it’s important for the music industry to realize that data does not always equal intelligence. There’s a perception across all industries that the more data our systems gather, the more intelligent they become. As a result, companies put a disproportionate emphasis on data gathering—and advertise their services as “AI-something-something” just because they gather swaths of data—without actually thinking through the problems to be solved and the methods/biases involved in doing the job.
Case in point: I have never encountered more unanimous skepticism around a music product in my (relatively short) lifetime than with the launch of United Masters
. I think underlying this skepticism is the argument that United Masters fails to distinguish correctly between routine and non-routine tasks in music marketing.
In an interview
, United Masters/Translation CEO Steve Stoute stated that one goal of his new company is to “operationalize independence” in the music industry, because “there needs to be 250,000 Chance The Rappers.” Put another way, Stoute claims that Chance’s meteoric, exceptional rise is something that can be repeated and automated at scale. But in my opinion, DIY music marketing shouldn’t be just about replicating someone else’s success in a process-driven way.
A lot of it is deeply personal and requires patient deliberation and creativity for which quasi-“AI-driven” social media analysis alone is insufficient.
To paraphrase Bob Lefsetz, Google can “take people out of advertising” because the ads in question are for static products that don’t have emotions and can’t talk back to you. My prediction is that Google, Translation, Andreessen Horowitz et al will have a much bigger challenge with trying to take people out of the music industry.
I think Spotify is facing a similar issue of applying AI to industry guesswork with mixed results, specifically around defining genres and music communities. As part of its end-of-year “2017 Wrapped” campaign, the streaming service published an enigmatic list of ten emerging genres, including cryptic phrases like “vintage swoon” and “jumpstyle” that not even the most passionate music fans have ever heard before. I was reminded of my own end-of-year Spotify recap from 2016, which listed “escape room” as one of my top genres; I was genuinely confused, and Spotify wouldn’t tell me anything about what “escape room” actually meant, so I had to put in the work myself
to find out.
Spotify is staying similarly silent this year with its 2017 genres, generally leaving fans more confused than excited (try searching “escape room” or “vintage swoon” on Twitter and you’ll see what I mean). There’s something fishy to me about imposing well-researched but nonetheless arbitrary cultural categories onto listeners who have no conception of what those categories mean or why they belong there. But yes, this is the same methodology behind the wildly popular Discover Weekly and Release Radar playlists—and perhaps this points to the importance of how we communicate the guesswork inherent in music-industry applications of machine learning.
I’m curious to hear where you’ve seen machine-learning applications work well in the music business, where you think they’ve totally missed the mark, where you think opportunities lie that we haven’t been thinking about, and/or whether I should stop being so harsh on United Masters. Let’s chat!