From Curation to Recommendation to Generation: Spotify's Gustav Söderström on Rethinking Product

Gustav Söderström joined Spotify in late 2008 or early 2009 after founding mobile startups and selling one to Yahoo. Daniel Ek and the team had built the free streaming desktop product but needed someone to work out mobile, which was hard because edge networks couldn't stream in real time and the free model couldn't be funded with ads. He later took on all product development and then the CTO role, and is now co-president alongside Alex Nordstrom, running the product and technology half of the company.

He says that because Spotify grew so fast, it felt like starting at a new company every six to 12 months, moving from a Nordic challenge to a European one, then the US, then public markets. The title stayed largely the same while the job kept changing.

He also explains why he made a limited podcast series on Spotify's product story. It began as an internal show interviewing Daniel Ek's direct reports to make leaders more approachable, and it let him empathize with podcast creators facing problems like music rights and user acquisition. Several people have told him it was a reason they joined the company.

Söderström lays out a lens he uses with his teams: the internet began with user curation of digitized goods, shifted to algorithmic recommendation, and is now heading toward generation. He asks teams to treat generative AI as something new rather than more of the same. Meanwhile Spotify still uses language models for incremental gains, such as richer recommendation vectors and safety classification of podcasts nobody has listened to.

The one product he counts as truly generative is the AI DJ, where a digitized version of a real person talks about music you like and suggests tracks. It could not exist without generated voice and generated commentary, at the scale of half a billion people. It targets what Spotify called the zero-intent use case, when you don't know what you want, which radio served in a crude way through its channel-flipping feel.

The team's principle was to avoid showing off the technology: people came for the music, so the DJ should do as little as possible and get out of the way. He adds that the idea came bottom-up from teams, though it needed big acquisitions to build. He describes the early reaction as a small magic trick, since listeners wondered how someone could have recorded so many personalized lines.

Söderström credits Chris Dixon for the principle of fault-tolerant user interfaces. He says he often saw design sketches with a single big play button in the early recommendation era, but such a design only works if prediction error is close to zero. With a one-in-five hit rate you need a screen showing about five options and an escape hatch for when the prediction is wrong.

He thinks the same applies to generative AI and cites early Midjourney in Discord. Instead of making users wait minutes for one image that might be bad three times out of four, it quickly generated four low-resolution images, implying a hit rate near one in four, and let users pick one to refine or upscale.

He ties this to a broader idea he teaches: all great products pull a magic trick. DALL-E, Stable Diffusion and Midjourney felt like magic at first use, and he says reaching that point usually means getting performance to a threshold, scoping down and fine-tuning until it crosses from impressive to magical. The magic wears off, but it tends to correlate with products going viral.

Asked about auto-generated music from artists' catalogs, Söderström stresses it's very early and the legal landscape is unclear. Creatively, he compares it to Avicii, who wasn't initially seen as a real artist because he worked in a digital audio workstation. He thinks diffusion models may become a much more powerful instrument, enabling creators who couldn't play instruments or hire an orchestra.

He rejects a clean split between AI and real music, since talented musicians already use AI to get better, so the real question is what percentage and who uses it. He also notes that models mostly produce things similar to what exists, so truly original work may become harder, not easier, and he hopes new genres might emerge as EDM did with the DAW.

On rights, he draws on Spotify's own history with piracy. Peer-to-peer was an exciting technology shift for consumers but needed a new business model before creators and industry could benefit, and afterwards the music industry became bigger than ever. He hopes for something similar with generative AI, but says the second part has not yet been seen.

Söderström says squads and chapters were smart for Spotify's early stage and helped recruiting, but they have become a cost because people still assume Spotify organizes that way. Squads were meant to be small, about seven people, full stack and highly autonomous. At scale, increments of seven engineers produce a lot of overhead, so teams are now larger, perhaps 14 per manager, and the structure looks more traditional.

The deeper problem was autonomy pushed to the leaves of a very young organization. He ties it to Swedish culture, where decisions are drawn as a circle with no one in the middle, and says renting expensive brainpower while giving it no room to think makes no sense. But with a hundred squads and a hundred strategies, you mostly produce heat.

Concentrating autonomy at the top has its own problem: a bottleneck, and Daniel Ek wants to maximize throughput. Spotify found that the VP level works best, giving tens to maybe hundreds of senior people with strong pattern recognition room to think. For example, he asks the podcast VP to define podcast strategy, and the AI DJ came as a bet from the personalization team.

Söderström describes an exaggerated spectrum of organizations. Amazon, with two-pizza teams, few dependencies and internal competition, has fast access to users and produced Kindle and Alexa. Because competing teams are incentivized to hide code, he says Bezos' insistence on hard APIs was essential, and it helped Amazon later expose its infrastructure as AWS.

The downside of that speed is shipping your org chart to the user, with duplicate search boxes or multiple toasters on the same screen, something Spotify itself saw in its autonomous phase. Apple sits at the other extreme: a centralized functional organization where something close to a single decision-maker decides what appears when you double-click the power button. The user experience feels coherent, but things can take years to reach market.

Spotify chose more centralization because its strategy is one app combining music, podcasts and audiobooks despite very different backend business models. A single recommendation organization decides which content type to surface, and a unified interface keeps things simple. Both models work, he says, and Lenny adds that being at an extreme may beat the middle, which Söderström likens to a smiling curve.

Söderström unpacks the backlash to Spotify's redesigned feed. Users said they were stuck in a taste bubble, but recommending a new genre is a fundamentally different problem from adding another EDM track to an EDM playlist, because you have no signal. Background recommendations need around nine hits out of 10, while discovery can accept one in 10 if swiping is cheap, which is why feed-style interfaces suit it.

The subfeeds for podcasts, music playlists and genres worked as intended, but putting them on Home did not. He says the old Home was about 90% recall, helping people return to sessions or known playlists, and the test flipped it to about 10% recall and 90% discovery. Traffic moved to search and library, and people were using the discovery UI, which he likens to a slot machine, to find known playlists.

The lesson was that Spotify's Home does multi-session recall better than other products and that this should be kept. The team is now testing a version where discovery tools are prominent but voluntary. He distinguishes new features, which people can ignore, from redesigns, which are mandatory and produce two kinds of feedback, one from disrupted habits and one from real problems. He suggests checking new-user cohorts to tell them apart.

Söderström says the hardest part of a big redesign is not outside criticism but the internal cost: with a new interface and new feed algorithms, a bad result may be either a wrong idea or weak machine learning, and the MVP must be large. His team tested several variants after rollout, combining quantitative data, user research and feedback. He told them beforehand it would be painful and that the chance of getting it exactly right was low.

His rule is to believe in something 100% until the data says otherwise, then believe the next hypothesis 100%, without getting precious. Lenny calls it strong opinions loosely held. Söderström adds that people dislike leaders who change their minds, even though we should want that.

On leadership, he says he brings energy because he only works on things he's excited about, and clarity because explaining ideas is the best test of understanding them. He asks leaders to always explain decisions, since even those who disagree are owed the reasoning. Provocatively, he says product is "0% art, 0% magic, and 100% science," noting that genetics, quantum physics and now intelligence were all called magic until explained.

For thinking, he now relies on distributed walk-and-talks over AirPods, which he found more effective during the pandemic than expected. He also shares Shishir Mehrotra's rule of spending about 10% of a cycle planning, which at Spotify's six-month increments means roughly two weeks.

Looking ahead for podcasts, Söderström says Spotify is focused on helping creators find an audience, which he calls even harder than music because finding a new podcast is such a big investment. The other major need is monetization, free and paid, as the industry matures. On the listener side he hints at investing in discovery, seamless playback across devices and cars, better search and generative AI, without giving specifics.

In a lighthearted detour he discusses Succession, saying Fjällräven and the saunas are authentic but the Swedish character's negotiating style is not, and explains a Swedish saying about peeing in your pants in the cold as shorthand for short-term thinking. His book picks include 7 Powers, The Complete Investor and several science titles, and he names Halt and Catch Fire and Duolingo as favorites.

He closes with a ritual he accidentally started about 12 years ago: a four-phase view of products, think it, build it, ship it, tweak it, with the think-it phase kept cheap to reduce risk before spending in the build phase. He asks listeners to reach out at @GustavS.