Pandora has never had a more pivotal moment than this one. The 75 million user-strong service is synonymous with Internet radio, but a growing list of competitors (which now includes Apple) and its royalty dance with the record industry loom large.
It’s a classic tech business conundrum. The world looks at Pandora as a new standard, but its position is actually much more precarious: To keep thriving, Pandora needs to ink major deals inside and outside the U.S. This is the kind of stress that would crush a lot of founders, but Tim Westergren says he’ll fall back on an old standby: naiveté.
“The only reason I did [Pandora] was because I was naive,” Westergren says. “I would say that entrepreneurship in general requires naiveté. Were you not naive, it could be so daunting that you would either not try it in the first place or you’d give up if you began to see things getting difficult.”
When he was trying to sell the idea of Pandora to investors, he kept hitting a wall. Indeed, he pitched the idea 348 times before securing the company’s second round of funding.
Westergren heard the same criticism again and again: It doesn’t scale. That is, having trained musicians sit down at a computer and punch in descriptive attributes one by one–effectively building a recommendation engine by hand–might be a culturally smart way to do things, but it would never scale enough to cover all the music in the world. This was, after all, the early 2000s when automation and algorithms were shedding their academic labels and starting to transform industries. You want to use humans to do this? You can’t be serious.
“People thought I was out of my mind to do something that was so retro and slow,” Westergren says. “It was contrary to all the trends in technology. There was a strong belief that you need to build scalable models and that technology was the only way to tackle big data problems. There was a lot of naiveté that I needed to buck that institutionalized belief.”
And yet, like a gambler at the blackjack table, Westergren just kept trying. And trying. And trying.
“What makes people lose so much money in Vegas is they have a hard time saying ‘Enough’ and pushing back from the table,” he says. “I think there’s a degree to which that happens in entrepreneurship. You’re like ‘Oh man, I’ve gone this far. I’m not going to have nothing to show for it. I’m going to keep going.'”
So Westergren and his team pushed forward with their unique human-machine hybrid engine for music discovery. Their approach, called the Music Genome Project, used the human intuition of academically trained musicologists to break down music song-by-song, attribute-by-attribute and feed that insight into a complex algorithm. Over time, the resulting ocean of data would grow more and more complex as Pandora’s user base grew. And over time, Pandora’s tech team and data scientists have tweaked their processes in order to tame the data.
Today, the Music Genome Project still lies at the heart of what makes Pandora tick, but it’s far from the only factor at play when you fire up the Beyonce station and hear a Rihanna song. To be sure, without this hand-built data set, Pandora would not exist today.
Still, Westergren admits, those early critics had a point.
“It took us four years for us to build the Genome large enough to be an asset that was useful in a product and then to figure out that we should use in the form of personalized radio,” Westergren says. “For four years, it looked more like a university R&D project. Almost like a vanity project.”
To Westergren, the concept made sense. As a musician himself, he instinctively knew that computers alone couldn’t solve the problem of digital music discovery. Perhaps they would in some distant, cyborg-populated future, but certainly not anytime close to the early 2000s.
“It’s not something I uncovered through technology,” he says. “It was something that was born of a human process.” And so, in spite of the criticism and rejections, Westergren and his team plowed forward with Pandora radio.
“Had I listened to what everybody was telling me at the time, there’s no way I would have done that.”
Today, Pandora is far from alone in the online music discovery space. Not only have music subscription services like Spotify and Rdio launched their own Pandora-style radio features, but entrenched giants like Google and Apple have their own copycats, too. Then there’s music intelligence platform The Echo Nest, which was recently acquired by Spotify. Its algorithm powers dozens of online radio services with its own blend of human and machine smarts.
Of all of the Internet radio products on the market today, none of them relies solely on computers to fuel music discovery. Indeed, Beats Music, the newest entrant to the subscription market, relies quite heavily on human curation and markets itself by underscoring that focus. Then there are smaller music discovery sites like Songza, Shuffler.fm, and The Hype Machine, all of which are based on the expertise of human tastemakers.
Westergren appears to have been onto something.
In addition to just the right amount of naiveté, Westergren says Pandora owes its success to discipline. As complex as the underlying science behind the service is, users still see the most bare-bones possible interface when they use Pandora from the web, any number of mobile devices, car dashboards, or any other device with which they’ve integrated. There’s a reason for that simplicity.
“We’re not lost in the things we want to build or that we think are cool or are sexy or attention-grabbing or sophisticated,” Westergren says. “We really have developed a very strong sensibility of the user.” They’ve done this part by meticulously studying user behavior and running A/B experiments on their millions of listeners. And as much knowledge as they unearth through these processes, the interface remains dead simple.
“It requires a lot of discipline, because lord knows that with 450 attributes we could build a 747’s cockpit’s worth of controls on Pandora. And we’ve been really disciplined not to do that.”