Gustav Söderström on Why Spotify Could Only Be Born in Sweden

Lex Fridman opens with a question that sounds like a stunt: Spotify has more than 50 million songs, so what is the definitive greatest track of all time? Gustav Söderström, who leads product, design, data, technology, and engineering as Spotify’s chief research and development officer, refuses a global ranking. He would not believe much in machine learning if everyone had the same taste. For himself the answer is easy: “You’re So Cool,” Hans Zimmer’s cue from *True Romance*. He sat with the organist and helped arrange it for his wedding.

Fridman ranks the movie in his personal top three. Söderström says it landed in his formative years, which is when music taste actually gets shaped, so the song has followed him ever since. That is the problem Spotify spends its life on: how a private history of listening grows, and whether a machine can learn it. Asked for the purpose of music on the planet, he gives two opposite jobs. One is escapism, living in another mental state for a while. The other is focus, using sound as a fast hack to tune a brain, which is a pile of neural networks, into the state the current task needs. Stories can do that too, but they are slower.

Fridman treats listening as private. His library is full of guilty pleasures that almost nobody knows, including people close to him. Söderström agrees, then adds a layer Spotify spent ten years noticing. Music is social, but in a small group. It is an intimate relationship, not something you broadcast. The company tried sharing listening on social networks and found that people are not very interested in what their friends play. They want to know whether they have something in common. The full history is not the valuable object. A playlist you have grown for years is.

Fridman asks for a short history of listening, noting flute-like instruments in caves some 35,000 years ago. Before recording, Söderström says, creation and consumption had to happen in the same place. You walked to the village. Distribution was terrible, but two qualities came with that scarcity: the creator always saw the audience, and there was no time cap. Early classical works running tens of minutes were not a coincidence.

The first wax disc held about three minutes on a side. That physical limit became the standard length of a song, a constraint with nothing to do with art. You used to write an opus; now you got three and a half minutes because you were out of wax, and in return you got enormous reach. Radio arrived as a news medium, then needed something free to fill time so it could sell more ads. Music filled the clock. That ecosystem invented the hit, and a broadcast structure in which everyone heard the same things. Shared culture has a social payoff: the latest *Game of Thrones* episode is something you can talk about with a stranger. Personalization thins that background.

CDs made music a digital file that was still shipped in a box. Then the files could be downloaded, and the three-minute cap loosened. The genre that actually broke the format, Söderström notes, was EDM, born as a download. Tracks of eight or ten minutes were possible because there was no wax left to run out. He is surprised it has not happened more. Björk made a whole iPad app as an album and got nowhere near the distribution of staying inside three minutes. All the training data, Fridman admits as someone who writes music, is three-and-a-half-minute songs. Streaming finally lets creators change the format without killing their reach. They have barely started.

The internet created a vacuum: a digital format with no business model. Napster and, in Sweden, The Pirate Bay gave listeners an access model with no marginal cost. It was a fantastic consumer experience and a disaster for artists. Download stores later attached a price, 99 cents a try, and did not restore access. Söderström’s example is people who fall asleep to music. They would not do that at 99 cents per three minutes, and they would explore much less if every skip had a dollar cost.

Fridman, who grew up around Russian piracy, says the people he knew could often pay. They chose the better experience. Piracy, in that reading, marks a business opportunity. Söderström agrees. The pirate interface was rotten: no metadata, slow downloads, messy libraries. It still solved access. Users even used Spotify to find songs and then torrented a copy onto a phone, a picture of distribution and consumption split apart. Adobe’s shift to monthly billing, Fridman adds, turned friends who used to pirate Photoshop into people who pay gladly. Sometimes product-market fit shows up before the business model. Piracy is the signal.

Sweden made the signal political. There was a registered Pirate Party and a public sense that information should be free, that charging for ones and zeroes was somehow wrong, that artists should live off concerts. That culture forced Spotify to compete with free in a way Söderström does not think could have happened anywhere else. The industry had to be doing badly enough to take the risk. Sweden was a perfect lab: government-funded high-bandwidth, low-latency broadband so the product would work, and not much of a music industry left to lose. His conclusion is sharp. This product could not have been born in America, the world’s largest music market.

Once the price is zero, you only beat free on experience. Even on fast broadband, downloading a track still took 30 seconds to a minute. Spotify started playback inside the perceptual limit of immediacy, about 250 milliseconds. The trick, Söderström says, was that it felt as if you had already downloaded the entire library onto your hard drive. It was not there. It was still free. It was somehow legal. People showed each other the click, stunned that it could be real. Legal, fast piracy was the proposition.

The work was a distribution innovation. One of the engineers was Ludvig Strigeus, author of uTorrent, the tiny BitTorrent client. Founders Daniel Ek and Martin Lorentzon hired him early, sold the client to BitTorrent, and kept the person. Spotify therefore had deep peer-to-peer experience. Until a few years before this conversation, the company hosted all the music itself, server and client. That end-to-end control let them cache locally, treat the desktop fleet as a giant distributed disk, and hack TCP, skipping mechanisms such as Nagle’s algorithm that ramp slowly, running full throttle and trading bandwidth for latency.

The prototype was 2007; the Swedish launch was 2008, desktop only. The iPhone existed, the App Store did not. The writing was on the wall, but there was no phone yet. An invite-only beta, meant to keep their own servers from melting, also built expectation, a tactic they reused for the U.S. launch. Today Spotify runs on Google Cloud and the rest of the internet is fast, so that technical edge is gone. At the time it was the whole product. Internally they already thought of themselves as an audio service, not a music app. People spend as much time with audio as with video, he argues, and audio is badly undervalued. YouTube, oddly, is a competitor for discovery and a poor MP3 player, because the app has to stay in the foreground.

The deeper shift was not speed. It was access versus ownership. A lot of people said they would never rent music. Fridman describes his old habit as hoarding, song by song and album by album, the physical stack as part of the pleasure. Coming to Spotify felt like throwing the collection away. He calls it therapy. Söderström has run that session with a lot of users. When the product launched, people hoarded inside it as if the files would vanish, the streaming equivalent of downloading. Playlist caps of a few hundred thousand tracks were supposed to be unreachable. People hit them. Some still try to save the entire catalog.

The mental trick is not that you discarded your music. You mailed the hard disk to a locker that now also holds the world’s catalog, for free. You got more music, not less. The design that makes the story believable is the free tier. Playlists survive even if you stop paying. If the rule had been that unpaid work disappears, people would never have invested the hours, and Söderström thinks that is still a weak joint in many competitors. Fridman adds a side effect of hoarding: he wore out a small set of songs. Variety lets him keep playing “Stairway to Heaven” without getting sick of it.

The catalog, in this conversation, is over 50 million tracks. Playlists number over three billion, sixty times more. Söderström thinks of the catalog as a space and each playlist as a path through it. Across tens of millions of atomic units there are billions of journeys that make sense, and they are probably far from finding all of them. Early Spotify was a strong search box plus a programming language he calls playlisting. People who knew music could soundtrack a life, perfectly personalized, because they wrote it themselves. Most people are not that good at music, or that free of time. Recommendation is hiring them a friend who is.

Heavy playlisters retained better, so the first scale attempt was human: professional editors, statistics on how a list performed, manual iteration for large groups. There were never enough editors to playlist for you personally. Machine learning was the promise of going from group personalization to individualization. The three billion lists were luck, he admits, not a primary strategy. It looks clever in hindsight. Someone in the company was already good at machine learning in 2007 and 2008, when the toolkit was mostly collaborative filtering. They noticed that users grouping tracks for themselves, then naming the group, were labeling a semantic dimension. The embeddings worked.

The surprise was who they worked for. From the outside you would expect mainstream taste to be the easy problem. The opposite happened. Recommendations were fantastic for people who saw themselves as having unique taste, and weaker for mainstream listeners, whose behavior was messier. Spotify was forced to solve the hard problem first. In 2014 they added The Echo Nest, which had no play data: sonic structure plus Wikipedia-style cultural references. Spotify had behavior; Echo Nest had content. A brand-new song has no plays, so you fall back on the artist, the sound, and what it resembles. User data sees culture the spectrogram will never show. A Paris research lab also found that song structure alone can predict where a listener will skip. Music is a mind hack. It would be strange if it had no structure.

YouTubers Fridman has met watch drop-off minute by minute and rewrite in the afternoon. Music has almost no such loop. Söderström’s comparison is uncomfortable. A musician works in stems, the source code, then compiles a finished file and ships it into the world’s MP3 runtimes, hoping for the best. No software developer would work that way: no collaboration, no A/B test, no issue tracking, no iteration on feedback. “The tools that a music creator has versus what a software developer has” look arcane. Why wouldn’t a musician have something like GitHub?

Spotify bought Soundtrap, a browser DAW aimed at Google Docs for stems, so people can collaborate in source-code format. It also bought Anchor, a mobile-first podcasting tool that lowers the cost of starting. Before that came Spotify for Artists and Spotify for Podcasters: verify who you are, then see cities, demographics, and how a release is doing, the dashboards developers have had for years. Fridman notes that Adobe’s Photoshop helps you invent a sky, but Adobe does not own distribution, so there is no YouTube-style response to the thing you just made. If Spotify can join creation tools to the consumption pipe, it has a feedback loop Adobe does not.

The company mission Söderström recites is enabling a million creators to live off their art, a sentence that puts makers first even though the product faces consumers. They widened the definition from music to audio because the world already had. The moment headphones go on, you choose between a song and a new episode; that choice is in your hand, not in a product taxonomy. Audiobooks are rising. Apple Podcasts dominated, and two years earlier Spotify had none. Research said people wanted shows and music in the same app, so they did not build a second client. They dropped podcasts into an app that already had about 200 million people, and quickly became a distant but growing number two to iTunes.

His verdict on the format in 2019 is blunt. A podcast is an opaque one-to-two-hour file, not interactive, with no feedback loop, which does not make much sense. Winning, if it happens, will be a better product for creators and listeners. He is careful: the ecosystem is fragile. You cannot barge in and internalize it without understanding how hosts get distribution and how they get paid. Fridman wonders whether playlists will break podcasts the way they partly replaced albums. Söderström says albums still matter in music because they are a creative form, a longer story over several tracks. Podcasts are more like Netflix. You do not watch one *Narcos* and then one *House of Cards*. Hosts and shows will not dissolve.

What is missing is discovery. The catalog is only about 500,000 to 600,000 shows; YouTube-scale creator video is probably in the billions. Easier tools will explode the directory, and finding a new show is a large time bet. Some series must be heard in order; some are one topic with rotating guests; some change subject every episode. The team is spending time on a knowledge graph of those objects and how people actually consume them. Fridman’s low-hanging fruit: most shows have no sequence, episodes are uneven, and there is almost no signal for which installment to try first. No upvotes. Nothing.

Stations by Spotify is a Pandora-style lean-back bet for people who will not chase every release. Some users want “Stairway to Heaven” this second; the same person, later, wants upbeat, or songs to sing in the car. Machine learning does not accept “make it good” as a spec. Söderström cites Andrew Ng’s practical turn: the test set is the new wireframe. The product manager’s job is to source a representative set of what “songs to sing in the car” means, then let engineering approximate it.

Expectation is half the product. Discover Weekly tells you most of this will not be for you; one gem a Monday is a win, even if one in ten or twenty is a miss. Daily Mix implies these are already your songs. Nine misses out of ten and the product feels broken. Algorithmically, Discover Weekly can go wild; Daily Mix has to stay in a safe neighborhood of taste. Names set the loss function. Internally they also dropped the humans-versus-machines story. Editors are much smarter than the models. They know culture, they know whether a song was in a film. They cannot make 200 million decisions an hour. The algorithm is less clever and much cheaper.

The hybrid has an internal name, algo-torial. A paid expert in soul or hip-hop invents a missing human need, titles it, images it, and hand-picks a few thousand catalog tracks that are great to scream in the car. Delivery then uses your taste vector to choose the twenty that are *your* car songs. The spec for that idea cannot be written as loudness plus choruses plus movie placements; it falls apart on line three. Once the pool exists, you can imagine auto-expanding it through the rest of the catalog. Taste vectors themselves are a hard engineering problem, not a hard theory problem: non-stationary, hundreds of millions of users, updated daily, now with deep embeddings that catch latent cultural coincidences handcrafted features would miss. Mood is often better as a click than as a prediction. A happy/sad tab is right 100 percent of the time; twenty tabs of inferred mood is friction. Spotify can build fault-tolerant UI and take risks whose worst case is a “WTF” tweet. Tesla, he notes, cannot.

Music’s object is about three minutes, so the feedback loop is three minutes, not two hours. Play-through versus skip is the high-volume signal, and it is noisy. A lot of listening happens with the phone in a pocket, on a run, in a car, on a speaker. Not skipping may only mean the track was not bad enough to wake you. Saving to the library is a clearer like. Building a playlist is rarer and almost trustworthy, and it also says which other tracks this one belongs with. The three channels cannot be treated the same. Fridman wants a way to mark “this was the right song for this exact moment.” They have played with thumbs-up, talking to the algorithm as if it were a person. He is not sure that is the right mental model. The other analogy is Tesla’s shadow mode: learn from how you playlist, stay invisible, offer an autopilot, back off when you take the wheel.

Smart speakers exploded that interface. The consumer need sounds small: control from across the room. Friction matters; play and pause by voice is more powerful than it looks. Spotify had assumed homes would skip from the CD player straight to a streaming box, because network-attached storage was too much work. The jump took longer than expected until voice speakers unlocked it. Music and podcasts are the number-one use. Google adding a camera so you can wave off an alarm is a hint they become seeing assistants, not “voice speakers,” with audio still the biggest job. The UI, though, is only vocabulary. People forgive almost no natural-language failure. Common acts, play, pause, next, are worth voice. Finding “that song” still sends people back to phone search, then to the speaker. The product has to sit at helpful-but-not-in-the-way, the Autopilot problem again.

Fridman’s concrete bug: say “play Stairway to Heaven” and a large share of the time you get a cover. The stack is the problem. Speech recognition returns an n-best list of guesses; personalization usually arrives too late. If someone has played the cover fifty times, defaulting to the original makes Spotify look stupid. The guess of what they meant has to use taste, and that layer is often in the partner’s hands. Spotify built its own understanding so it can make personalized guesses, and it negotiates constantly over how much user data to share. Users do not care about n-best lists or business deals. They said it fifty times. How hard can it be?

As of August 31, 2018, Spotify had paid more than $11 billion to rights holders. Fridman’s first reaction as a user was that the company would go bankrupt in a month; it was too good to be true. Söderström says betting against Spotify was statistically rational. The roadkill of music-streaming services was visible. If he had understood the industry when he joined, he would have made the more rational guess that it would fail. Ignorance was bliss. Sweden and the Nordics were a lost market, so labels could try without much downside. A healthy U.S.-scale market would not have allowed the experiment.

The cognitive gap was the model. On downloads and CDs, the artist got paid up front for all future plays. Streaming paid almost nothing on day one. Only later does the curve of incremental revenue cross that first payment, and it took years for labels to believe it. Once they did, the logic is clean: a hit that plays for years should be billed on actual use, not on day-one hopes. Industry revenues, after a brutal dip, were rising again. Some competitors skipped rights, got big fast, and dared labels to kill them. Spotify went legal on day one, negotiated slowly, and watched shortcuts seem to work. Söderström calls it a repeated game. After many rounds, cooperation is the statistically correct strategy. Ten years built a strange trust: you never screwed us, we never screwed you.

The hard base of that trust is alignment. If music does not get big, if people will not pay, Spotify has no other business. Other companies can lose money on music and still be fine. That is why the industry could take scary product bets, including the free tier, a huge risk for which labels deserve credit. Apple could afford iTunes’ dollar-a-song fight to fail; early Spotify could not. Daniel Ek’s tenacity in those talks, Söderström says, was part of why an unknown company survived them. There is no magic in who won. Outsiders saw $9.99 and thought people would pay for music. What actually happened is engagement grew on the free product until $9.99 felt plausible. Running large-scale advertising and subscription at once is rarer than it looks. Premium-only hit a wall of people who would not pay. Ads-only did not feed the industry. He half-jokes that he should not have said the secret out loud. Brilliance or luck: in retrospect it will be written as brilliance.

Music, podcasts, and audiobooks are core human needs, Söderström says, with no good reason not to reach messaging or social-network scale: billions of people, all the world’s music. The more interesting bet is format. SMS was a carrier standard, 160 characters, wonderful distribution, a straitjacket of a format. Adding pictures as MMS took from the late 1980s into the early 2000s, a fifteen-to-twenty-year product cycle. Wrap creation and consumption in one software stack, Snapchat or WhatsApp, and you get disappearing messages in week one and Stories two weeks later. For the first time, enough listeners and enough creators sit on a handful of streaming services that the hundred-year constraints can move. He hopes the next ten or twenty years are not the same three-minute loop again.

Podcasts give him hope about attention. Video in the foreground kept shrinking under monetization until it looked as if people had no patience for depth. Then long-form audio arrived. Maybe you will not stare at a phone for two hours, but you will drive and listen. Most of Fridman’s audience, they note, will hear this entire hour and forty-five minutes. That is not a species getting dumber. On devices, Söderström sees computing leaving the multipurpose phone for ambient, special-purpose hardware: a speaker always in earshot, a watch and earbuds for a run, the laptop used less rather than discarded. Voice is an interaction that does not need a large slab of glass. Big AR glasses that project onto the world still feel far; using a watch or glasses as a mouse pointer on reality, with audio answers, might not.

Fridman’s last detour is the film *Her*: falling in love over audio alone, and whether a machine could be that someone. Söderström’s personal answer is yes to both, given enough time. Podcasts already feel like sitting in the middle of a conversation. An assistant in your headphones all day would feel as if it lived in your brain, easier to fall for than a thing on a screen. Whether you fake it as a philosophical zombie or actually get there, he thinks it is atoms and arrangements of information. They started on *True Romance* and ended on love, which Fridman calls as good a place as any.