Recommendation Engines Are for People Who Don’t Have Friends
Last week, Mike McGuire, a good buddy and a former Gartner analyst who specializes in digital music and media, threw out a line:
“Recommendation engines are for people who don’t have friends.”
It made me laugh because there is something true underneath the joke.
Some of the best music recommendations I’ve ever received came from people who knew both the music and me.
Recently, Joe Barham, co-founder and CEO of HIFI Labs, told me to listen to Angine de Poitrine, an experimental Québécois duo from Saguenay. Joe’s team works on technology, fan engagement and data-driven programs for artists including Linkin Park, Bruno Mars and RÜFÜS DU SOL.
More importantly, Joe knows I like disappearing into musical rabbit holes.
Angine de Poitrine is exactly that kind of rabbit hole. They perform anonymously in oversized polka-dotted masks and costumes, use asymmetrical rhythms and live looping, and play a custom double-neck microtonal guitar with quarter-tone fretting. Their song “Fabienk” got stuck in my head almost immediately.
After I mentioned them in my last post, five different people asked me about the band. They’re playing the Fox Theater in Oakland in December.
Joe made a great recommendation because he understood the music, my taste and the context.
No algorithm was involved.
What recommendation engines actually know
For most of music history, recommendation came from a friend, a DJ, a musician, someone working in a record store, or an older sibling.
Those people had context.
There is only one Joe Barham, however, and there is no way he can understand millions of listeners simultaneously.
That is where algorithms, and increasingly AI, come in.
One of the foundations of modern recommendation is collaborative filtering. Amazon’s influential 2003 paper on item-to-item collaborative filtering helped establish the approach: rather than trying to understand the content itself, the system looks at patterns in behavior.
People who bought this also bought that.
In music, every interaction creates signals: play, skip, save, repeat, playlist addition, sequence and so on. Collaborative filtering can discover relationships between songs without actually understanding anything about the songs themselves.
Spotify says its platform generates nearly two billion music discoveries every day, and that Discover Weekly alone drives more than 56 million artist discoveries per week, with 77% involving emerging artists.
That is an extraordinary achievement.
But there is another way to think about recommendation.
Pandora’s Music Genome Project took almost the opposite approach. Trained analysts listened to songs and described hundreds of musical characteristics.
Collaborative filtering asks:
What do people who like this also like?
The Music Genome Project asks:
What is actually in this music?
The next generation of recommendation will increasingly combine both.
The same person can want completely different music
Understanding behavior and understanding the song still leave out something important.
Context.
I love Miles Davis. But Bitches Brew is probably not what I want at 6 a.m. while trying to get through a workout.
Get Jeff Yasuda’s stories in your inbox
Join Medium for free to get updates from this writer.
I also love the Grateful Dead. A 23-minute version of “Dark Star” might be transcendent in one situation and completely wrong in the middle of a high-intensity interval session.
Same listener. Same musical taste. Different moment.
IFPI research has found that people listen to music across a wide range of activities, including driving, commuting, relaxing, doing housework, browsing the internet, exercising, cooking and working. Younger listeners report even more listening contexts.
This is where the work we are doing at Feed.fm gets interesting.
Our platform handles roughly 55 million API calls every day across real-world music experiences.
Those interactions contain much richer context than simply knowing that a song played. Different products serve different audiences in different environments. A skip during a high-intensity workout might mean something very different from a skip during passive listening.
The opportunity is not just collecting more data. It is understanding what that data means.
From recommendation to Music Intelligence
AI is obviously changing how music is created.
I think the larger opportunity may ultimately be how AI helps us understand music.
What experience does this song create? When should it play? What belongs immediately before or after it? Who is it right for? When is it completely wrong?
Matt Sandler, a senior software engineer at Feed.fm, has been researching some of these questions. His paper, Embedding Amnesia and the Partial Ear of Machine Listening, presented at AIMC 2026, explores an important limitation of modern audio embeddings.
Very sophisticated models can represent enormous amounts of information about sound while still missing things humans use to determine whether music is appropriate.
Language. Cultural meaning. Seasonality. Rights status. Listener context. Intent.
A recommendation can be acoustically right and contextually wrong.
This is one reason we have started using the phrase Music Intelligence.
Music Intelligence combines multiple forms of understanding: musical attributes such as mood, energy, metadata and similarity; behavioral signals such as affinity, skips, completion and repetition; and the context in which listening actually occurs.
A traditional recommendation system might ask:
What is similar?
Music Intelligence asks:
What is appropriate?
Those are not the same question.
Humans still matter
I do not think the answer is replacing human music judgment with AI.
Great music programmers understand pacing, familiarity, surprise, emotion and culture. Those things can be incredibly difficult to encode.
AI can analyze enormous catalogs. Behavioral systems can learn from millions of interactions. Humans bring taste, judgment and intent.
The systems that ultimately work best will combine all three.
Which brings me back to Joe and Angine de Poitrine.
Joe did not recommend the band because I had listened to something with a mathematically similar embedding. He recommended them because he understood something about the music, something about me, and something about what I might find interesting at that moment.
AI has the potential to do that at extraordinary scale.
But understanding the music is only half the problem.
The other half is understanding the listener.
And it is those special moments, when the perfect song hits at exactly the right time, that we never forget… and we keep on coming back for more.
