Chris Moody knows a thing or two about the universe. As an astrophysicist, he built galaxy simulations, using supercomputers to model the way the universe expands and how galaxies crash into one another. One night, not long after he’d finished his PhD at UC Santa Cruz, he met up with a few other astrophysicists for beers. But that night, no one was talking about galaxies. Instead, they were talking about fashion.
A couple of Moody’s astrophysicist pals had recently left academia to work for Stitch Fix, the online personal styling company now valued at $2 billion. Moody gawked at them. “They were like, ‘You don’t think this is an interesting problem?’” he says. Indeed, he did not. But when his friends described the work they were doing—sprinkling in phrases like “Bayesian models” and “Poincaré space”—predicting what clothes someone might like started to sound eerily like the work he’d done during his PhD. Quantifying style, he discovered, “turns out to have really close analogues to how general relativity works.”
Four years later, Moody works for Stitch Fix too. He belongs to a growing group of astrophysicist deserters, who have stopped researching the cosmos to start building recommendation algorithms and data models for the tech industry. They make up the data science teams at companies like Netflix and Spotify and Google. And even at elite universities, fewer astrophysics PhDs go on to take postdoctoral fellowships or pursue competitive professorships. Now, more of them go straight to work in Silicon Valley.
To understand what’s driving astrophysicists into consumer product startups, consider the recent explosion of machine learning. Astrophysicists, who wrangle massive amounts of data collected from high-powered telescopes that survey the sky, have long used machine learning models, which “train” computers to perform tasks based on examples. Tell a computer what to recognize in one intergalactic snapshot and it can do the same for 30 million more and start to make predictions. But machine learning can also be used to make predictions about customers, and around 2012, corporations started to staff up with people who knew how to deploy it.
These days, machine learning drives everything from Stitch Fix’s curated boxes of clothes to Netflix’s personalized movie recommendations. How does Spotify perfectly predict the songs that will surprise and delight you in its weekly personalized playlists? That's machine learning at work. And while machine learning now constitutes its own field of study, because scientists from fields like astrophysicists have been working with those kinds of models for years, they’re natural hires on data science teams.
“We were already in Big Data before Big Data became a thing,” says Sudeep Das, an astrophysicist who now works at Netflix.
Das got his PhD at Princeton, where he researched cosmic microwave background—basically the electromagnetic radiation left over from the Big Bang. Afterward, he spent a few years studying data from the Atacama Cosmology Telescope in Chile. The telescope collects nearly a terabyte of data from the cosmos every night, and from this massive data set, Das detected an elusive astrophysical signal. It was a rare payoff after years of painstaking work. The discovery earned him the attention of the University of Michigan, which offered him an assistant professorship.
But Das turned it down and moved to Silicon Valley instead—first to work as a data scientist at Beats Music, then at OpenTable, and now at Netflix.
The decision to leave academia came down to a few factors: The pay was certainly better, and the jobs were more plentiful. “There’s a bottleneck of getting into tenure-track positions,” he says. And being in the Bay Area meant he and his wife—who is also an astrophysicist—would never have to worry about both finding jobs. But the real surprise, he says, was that the work in tech companies was actually interesting. At Beats, he says, he found “like-minded people who were working on problems that didn’t take away the intellectual high.” Same math, different application.
Das says he’s noticed that more and more physicists are trading the difficult slog of academia—which can involve a decade of financially dicey postdoctoral work—to take cushy jobs in tech. “Only two people from my PhD batch at Princeton are not working in industry,” he says. “You have to be a die-hard academic to stay.”
This big bang extends across the industry. “Astrophysics is our number one domain,” says Eric Colson, Stitch Fix’s chief algorithms officer emeritus. “Most folks have a PhD in a quantitative field, but if we did a histogram, I think astrophysics is number one. They teach math really well—a lot of physicists are better mathematicians than mathematicians. They also teach coding well. They’re better computer scientists than most computer scientists.”
Moody, who joined Colson’s team in 2015, put his astrophysical training toward problems like mapping a client’s “latent style”—someone’s unique personal taste in fashion. Stitch Fix doesn’t ask its clients to self-identify with labels like “preppy” or “boho.” Instead, it collects data on what people like through their purchases and though tools like Style Shuffle, a Tinder-for-clothes where people can “like” or “dislike” specific items. In the aggregate, that data makes up Style Space—a map of all the things clients “like” and the way they relate to each other. (Items that are “co-liked” by a client are “co-located” in a similar space on the map; each region of the map represents an aesthetic.) Moody and his team use that model to make predictions about what else clients will like. Algorithmically, it can infer that someone who likes chunky necklaces will probably like beaded necklaces too, the same way Netflix’s algorithm infers that you may want to watch another comedy with a strong female lead.
Moody says those kinds of problems don’t look so different from the work he did during his PhD. That map of latent style? “This is a Poincaré space. It’s what Einstein used to describe relativistic spaces," says Moody.
Understanding latent style involves other physics principles too. Moody’s team uses something called eigenvector decomposition, a concept from linear algebra, to tease apart the overlapping “notes” in an individual’s style, sort of like “plucking a guitar string and listening for the multiple notes overlayed.” A client might like items with feminine silhouettes, but ones that skew casual rather than professional. Each person’s individual style contains many data points—few people are simply “preppy” or “boho”—and using physics, Moody says his team can better understand the complexities of the clients’ style minds.
"No one who studies physics ever thinks about going into clothing, but it turns out that it's phenomenally rich," says Moody. "It's fascinating to try to think of personal style as a science."
Colson says the many astrophysicists on his team are attracted to the company “because of impact that, in their theoretical work, they rarely get to see. Here, they can push something into production and see the impact.” When Moody does his job right, Stitch Fix is more likely to send its clients items they’re likely to keep—a metric he and his team can track, and improve, every day.
In academia, astrophysicists can spend years stuck on a singular problem. And many of the exciting problems have already been solved, says Amber Roberts, a machine learning engineer and former astrophysicist who now helps academics transition to industry at Insight Data Science. “We discovered the size of the universe. We measured the speed of light. We found pulsars. We found black holes,” she says. “A lot of those big things, like understanding how space-time works or how gravity distorts, are what get people interested in the study of space and cosmology. But what you’re really doing is contributing to a very small subfield where you’ll work about three years on a paper that about 10 people in the world are going to read.”
Das, the astrophysicist who now works at Netflix, says it can be difficult to give up the romanticism around studying the universe. “When I go and explain this to my parents, they’re like, ‘You were doing such amazing things with the universe, and now you’re making people watch stuff!’” says Das. But he agrees that the daily work has more to do with technicalities, like “trying to get the measurement of the parameter from a 50 percent error mark to a 5 percent error mark,” rather than contemplating the big picture of the universe.
At Netflix, the technical pursuit feels more or less the same. But when he lifts his gaze and considers the work he’s actually doing—connecting people around the world with films and stories that help them understand each other—he feels no less satisfied with his scientific contributions than he did as an astrophysicist. “It’s sort of like uncovering a different universe,” says Das. “The universe of human beings.”
Correction on 10/15/2019: An earlier version of this article inaccurately described an eigenvector decomposition. It is a concept from linear algebra, not quantum mechanics.
- The tech helping dogs learn to “talk” with humans
- How cities reshape the evolutionary path of urban wildlife
- The iOS 13 privacy and security features you should know
- The “smart kitchen” is very stupid
- Artificial intelligence confronts a “reproducibility” crisis
- 👁 How do machines learn? Plus, read the latest news on artificial intelligence
- 🏃🏽♀️ Want the best tools to get healthy? Check out our Gear team’s picks for the best fitness trackers, running gear (including shoes and socks), and best headphones.
