This is the first issue of Extra Tokens. It means a lot to me that you’re giving my newsletter a shot. I’m excited, I’m nervous - I have the first-day-of-school jitters. Read through it and share your feedback with me. I want to make this your favorite newsletter. Hope you enjoy!
I noticed something new on my Spotify home page last week. Instead of jumping into my Drake Throwbacks playlist (IYKYK, and yes, Marvin’s Room might be on it), I tapped the new thing. Of course, it's a chat interface.
I asked for upbeat work music to write this issue. It delivered. Then I asked for upbeat Drake (because I wasn't trying to get into my feels) and of course, it queued Trust Issues, which might be the slowest Drake song ever recorded. Not great.
Turns out I'd stumbled into Spotify's AI DJ. Not just a chat interface. A radio host spinning jams exclusively for me. It's great when you have a vibe in mind (Italian Riviera while making pasta). It's terrible when you ask for a specific artist. When my wife asked for Harry Styles's new album, it played everything but Harry. Likely to spite her.
Love it or hate it, more personalized AI experiences are coming. Here's the strategy.
SPONSORED BY
You bought something online with a credit card. You may have even used the “right” credit card.
But did you:
Activate offers on your credit cards
Use a shopping portal
Check for heavily discounted gift cards
Shoppers who use Savewise earn 7-10x more points & miles for things they’re already buying.
Savewise finds every way for you to earn rewards and puts them all in one place.
TL;DR
The thirty-second version of what Spotify did.
Spotify's AI DJ has logged 4 billion hours of listening and been used by over 94 million subscribers since launch in 2023. On days a user engages the DJ, 25% of their listening time goes to it.
Spotify has been improving its recommendation engine for over a decade. They pump out more products than Marvel spin-offs (I’ve lost the plot, which branch of the multiverse are we in now?). A few of them:
Spotify Radio - 2012
Discover Weekly - 2015
Daily Mixes - 2016
Spotify Mixes - 2021
Daylist - 2023
Prompted Playlist - 2026
Spotify DJ does something different. Same recommendation engine. They just wrapped a voice around it. A DJ who calls you by name, tells you why this song is coming up, and ties it back to something you played last week. "Next up, John Summit. You've played 32 of his tracks this week" 😳.
🔥 The Shift
The leap in AI consumer products is personalization moving from a backend feature to the UX layer itself.
Under The Hood
The three strategies that make it work.
1. Personalization as a UX feature, not an algorithm change
Personalization on the internet today comes in two flavors.
Backend recommendation engines: Run on your data and serve a feed ordered for you (Discover Weekly, your LinkedIn feed, your Netflix recs).
Onboarding quizzes: Ask ten questions and hand back "your" plan. The questions make you feel seen. But the end product is the same.
Both feel hollow. Spotify went a different route. The recs are the same, but what changed is how they're served. No playlists. Just a voice that explains its picks and plays pretty much whatever it's feeling.
No playlists, just vibes
The real difference is the commentary. Hearing why the algorithm picked it makes you appreciate it more. Spotify's own tests proved it: users are 4x more likely to listen to a song paired with an explanation. Same backend, new presentation, so much more engagement.
Additionally, you can speak to the DJ and tell them what music you’re feeling. Not only does this let you control your personalization, but it also provides Spotify with valuable feedback. Previously, they only received skip signals, but now they’ll start to understand exactly why you can’t listen to Nickelback when your friends are in the car with you.
2. AI's job is to remove decisions, not add them
I'm less likely to buy ice cream if there are forty flavors (although hand me Baskin-Robbins mint chocolate chip any day).
Choice overload = more flavors, fewer scoops sold.
Spotify’s home page is the definition of overload. I counted at least 20 different buttons that I could click. It feels impossible to know where to start. Even with all their recommendation features, the last-mile choice is still on me.
I feel very exposed sharing my homepage
AI DJ removes the last mile. Press one button. The product decides what plays next, explains why, and moves on. Zero decisions from you. The decision moves from the user to the model. The freed attention keeps you in the product longer.
3. Reshape what users love. Don't bolt on something new.
Spotify and Snapchat both shipped AI products in 2023 on top of enormous user bases. One lifted engagement. The other crashed and burned. The difference came down to what each company decided to AI-ify.
81% of Spotify listeners cite personalization as what they love about the app, per VP of Personalization Ziad Sultan. Spotify took the thing users came for, personalized music, and gave it a new shape. The product users already loved got better.
Snapchat did the opposite. Their chatbot, "My AI," launched in February 2023, which was pinned to the top of the chat feed as a ChatGPT-style helper that answered questions like "what filter will hide my grey hairs?" (I tried this; it’s the Grey Tone filter, now I just wish it were applied all the time).
The core product, messaging friends, was untouched. AI was a separate thing, glued to the top of an app users opened for a different reason. A week after launch, Snapchat's average U.S. App Store review hit 1.67. 75% of reviews were one-star.
The lesson: AI works best when it makes the thing users already love feel new. AI fails when it tries to make users love something new.
So if Spotify's move is the new playbook, the obvious question is: where else does this work? Plenty of consumer apps are sitting on the same setup right now. Below is a few ideas of where it could work.
What I’d Ship
The AI DJ pattern works when the recommendation engine is good, and the UX still asks the user to pick.
1. Goodreads
I have 472 books on my Goodreads Want to Read list and a decade of star ratings on the ones I've finished. I still don't know what to read next. So I open Claude, tell it the last few books I liked, and ask.
The feature: a "what should I read next?" answer that picks 2–3 books from my Want to Read list and says why each one fits, "you gave Mistborn 5 stars, this one has similar pacing and a darker tone."
It can even pull from the whole catalog to surface books I haven't seen. Eventually, it asks specifics — "did the pacing of X work for you?" — and you answer, because you want better recs. Until Goodreads knows you better than your book club does.
2. NYT Cooking
I never know what to cook, even with NYT Cooking's catalog at my fingertips. I'd love to say "give me healthy(ish) Italian dishes for dinner" and have it pull from my saves, my ratings, and the popular recipes in the database. And if I'm stuck, surface a recipe I already love. Bonus points if it does the dishes. My current AI assistant (my wife) has filed complaints.
SPONSORED BY
Built by yours truly → a 2-minute pre-game routine for high-stakes moments:
The important job interview in 5 minutes.
That big board presentation starting in 10 minutes.
The hard conversation you’ve been putting off with your team.
"Just breathe" doesn't work when you're already spiraling, and Calm and Headspace aren't built for the panic before the big moment.
Primo is. It walks you through a 2-minute routine using science-backed techniques from Stanford and Harvard.
Extra Credit
If you want to go deeper.
A slightly technical read on how the Spotify team managed to fine-tune their LLMs to provide the commentary from the AI DJ. Interesting stats from this piece:
By hearing the commentary alongside personal music recommendations, listeners are more willing to listen to a song they may otherwise skip.
Tailored adaptation led to significant improvements (up to 14%) in Spotify-specific tasks compared to out-of-the-box performance.
Interesting watch on how Gustav Söderström thinks about what’s next for Spotify. What’s curious is how much of the discussion drifts towards podcasts and audiobooks. In addition to using AI for music discovery, he thinks the DJ technology can be used for podcast/audiobook discovery and even enriching them with knowledge from LLMs.
That’s all for this week!
I’ll leave you with one question to ponder: "What's an AI feature you've actually kept using and not just tried once?"
Reply and tell me about it. I promise to get back to you.
- Amaraj (or DJ Butta Butta when I'm on the turntables)
The Meme
All consumer companies in 2026