Joining YouTube Music: the recommendation is the product
I've started as an engineering manager at YouTube Music. Why music, and why I think the next few years of music discovery will look very different.
I started at Google this month, as an engineering manager on YouTube Music. I’ll keep writing here in a personal capacity, using public information.
A few people have asked why I’d go from co-founding a company to managing a team inside a very large one. Scale, recommendation and music all pulled me toward this job.
I wanted scale. At Amanda we worked hard to get a model right for a few thousand people at an event. At YouTube, a small change in how music is recommended affects what a very large number of people hear every day. The problems at that scale are different in kind, and I want to learn them.
I wanted to work on recommendation. I’ve written here for years that in feed products the recommender is the product: Discover Weekly in 2015, TikTok absorbing Musical.ly in 2017 and 2018, Reels and Shorts last year. I’ve watched it from the outside and from the client side long enough. I want to work on it directly.
And I love music, which sounds like a small reason but isn’t. I’ve been listening to AC/DC since I was a teenager and to 万能青年旅店 since college, and I care about how people find the things they’ll love. Working on a product you use every day, for something you care about, makes the hard days easier.
Music is also an unusually interesting recommendation problem. People play the same songs hundreds of times, which is the opposite of news or video, where you rarely want the same item twice. Listening is often passive. The music is in the background while you work or drive, so a lack of skips isn’t the same as love. And taste is tied to identity and memory in a way that’s hard to model. The song you played on repeat in college means something to you that no embedding captures.
It’s also a good time to be here. Pretrained models that understand audio and text are getting good, and shared spaces for content and language, like CLIP, suggest a future where you can describe what you want to hear in plain words. Google I/O last week showed LaMDA, a model built for open conversation, which is another sign of where interfaces are heading.
I don’t know yet what I’ll be able to write about from inside. Probably less about my own work and more about the field. I’ll keep the blog going either way.