Three years after launch, Spotify's AI DJ has reached 90 million users and over four billion hours of engagement. Spotify considers this a win. I think it tells a different story. The platform has 290 million paying subscribers. Even backed by a decade of machine learning and the listening data of 751 million users, the most sophisticated music recommendation engine ever built still hasn't pulled in two-thirds of them. Most paying users would still rather pick the songs themselves.

I've been thinking about why.

The algorithm I tried to build

In January 2017, I published a piece on Medium called "Man vs. Machine — The Algorithm that beat the DJ." I was running a music tech startup called Muru and I was convinced we were on to something big. The thesis was simple:

"Putting a list of songs together is relatively straightforward. Selecting a group of songs that fit together perfectly and then playing them in a sequence that has meaning… is a whole different ball game. But, if you think about it for a moment, that is exactly what a DJ does."

So we tried to codify the DJ brain. We built a classification engine we called Madlib that could detect over 40 primary genres with 90%+ accuracy. This was State of the Art at the time. Then we built what I called the "DJ Formula," an algorithm that could take a set of songs and order them the way a professional DJ would. To prove how powerful this was, we ran a blind test. Two professional DJs and the algorithm were given the same tracks, the same first song, the same last song, and told to arrange the rest. We published the graphs. You genuinely couldn't tell which playlist was the machine's.

I was triumphant. The DJ brain could be codified. The algorithm could think like a human.

I was also, in a way that took me nine years to fully understand, wrong.

What happened next

Every streaming service built their version of what we were trying to do. Spotify's Discover Weekly became the most successful recommendation feature in music history. Apple Music hired the Beats team and leaned into editorial curation backed by algorithms. Every major platform invested hundreds of millions in machine learning for music.

The algorithm won. In the narrow sense of "can a machine sequence songs competently," the answer turned out to be an obvious yes. That specific problem was solved by 2020.

But something strange happened. The playlists got good and then they stopped getting better. Spotify Wrapped 2024 faced backlash for being generic. The shuffle algorithm has been criticised for years for producing repetitive patterns. Users report the same feeling over and over: the recommendations are fine. Not wrong. Not bad. Just fine.

In music, "just fine" is the worst thing you can be.

What I missed

When I wrote that 2017 article, I framed the problem as a classification and sequencing challenge. Get the data right. Get the order right. The DJ brain is codified. Done.

What I missed was that the DJ brain isn't really about classification or sequencing at all. Those are the mechanics. The actual skill was timing. It's what separated a great DJ from a competent one. Knowing that this specific track, at this specific moment, for this specific room, would work when nothing in the data said it should. The willingness to play something risky because you read something in the crowd that no algorithm could measure.

I spent 16 years behind DJ booths. The nights I remember weren't the ones where I played a technically perfect set. They were the ones where I made a call. Dropped a record that made no sense on paper. The room went somewhere unexpected. That's not pattern matching. That's judgment.

Madlib could hit 90% accuracy on genre classification, state-of-the-art at the time. It could sequence tracks in a way that fooled DJs in a blind test. But it couldn't decide that tonight, in this room, in this moment, the rules don't apply. That decision was always mine. In terms of a user experience that simply means giving the user the control to decide the timing, in a lean-back approach. This was tried by Pandora with the thumbs up thumbs down approach. Again, fine. Not great. It focused on the individual songs instead of the sequence (the vibe in that moment). We ultimately solved that lean-back problem with our Muru app. We let you create a dynamic playlist. You would pick the start song (much like Spotify Radio) but core to the user experience you could either adjust the playlist in real time by moving some sliders, or select what song you wanted to end with in a fixed period of time (30 minutes up to 6 hours). Our app would then create a DJ-inspired journey sequencing the songs. It worked really, really well. Then we got blocked by the Spotify API. Game over.

The parallel that matters now

There's a phrase showing up everywhere in 2026: "taste is the new bottleneck." When AI can build anything, when execution is effectively free, the thing that separates good products from forgettable ones is the judgment behind them. Taste and Timing is everything!

This is exactly the DJ problem, nine years later, applied to everything.

I'm building four products and counting, right now. All of them were built with AI. AI handled most of the execution: the code, the copy. But the decisions that mattered? Which product to build. Who it's for. What it should feel like. What to leave out. Those were mine.

Every product I've built this year started with a specific opinion, not a prompt. BigMoneyIdeas exists because I spent fifteen years watching smart people chase bad ideas. PadelCrews exists because I felt the coordination problem in Padel games on my own WhatsApp groups before I wrote a single line. The AI built all of it. The judgment about what to build and for whom came from the same place it always has: lived experience and a willingness to have a point of view.

That's taste. The same taste that told me which track to play at 2 AM in a room I'd never played before.

The 2026 version of the DJ algorithm problem is this. AI can now do what Madlib did. Classify, sequence, execute, better than I ever imagined. But the reason most Spotify subscribers still pick their own music is the same reason the best products in 2026 will be built by people with genuine domain expertise and specific points of view. The algorithm gets you to good. Taste and timing is what gets you to great.

What this means if you're building

Taste and timing aren't things you can prompt into existence. They come from years of doing the thing or a willingness to have a point of view. From playing the wrong track and watching the room empty. From shipping a product nobody wants and understanding why.

When I closed Music Health, I lost something very hard to rebuild: the conviction about what to build next. It took a year of building solo, shipping things, learning, breaking things, watching what people actually used, before that conviction came back.

The AI tools are extraordinary. I wouldn't trade them. But I wouldn't confuse them with the thing that actually matters. Yann LeCun has been making this point for years. Today's LLMs manipulate language brilliantly, but they're closer to advanced information retrieval than to actual intelligence (think Rainman). They lack a model of the physical world, which is why they can't truly reason or plan and fail at being "creative". If you remember that while building, you can do great things. If you forget it, you might find yourself in chaos. Madlib could sequence a playlist. It couldn't tell you which playlist the room needed to hear. Nine years on, no one has built the version of Muru that could. The gap is still wide open.