The mental model

Imagine a human DJ sitting next to you, taking notes. Every time you nod, they note "more like this". Every time you grimace, they note "not this". Every time you say "play something jazzier", they pivot. After a week of this they know what you want by 8am on a Tuesday. That’s exactly what musen’s AI DJ does. It just notes a lot faster and forgets nothing.

Explicit signal #1: Love

Love is the heaviest single signal. When you Love a track, the model treats it as a strong positive example of music you want more of, right now and across this time of day. Loves compound: three Loves on a single artist in two days will measurably shift the next session’s opening. Loves never expire (artists you Loved a year ago still nudge the model, just less strongly than recent ones).

Explicit signal #2: Retune

Retune tells the model "not this vibe". The interesting bit: Retune is more interesting than Skip. A Skip says "not this track"; a Retune says "not the cluster this track belongs to". The model uses Retune to shift the next 5–10 selections away from the current acoustic / mood / era cluster. If you Retune three times in a session, the radio reshapes around your new mood within roughly 90 seconds.

Explicit signal #3: Request

Request is the most direct signal. Natural-language prompts like "Tuareg desert blues late-night" or "morning focus, no vocals" or "1970s Ethiopian jazz" are converted by an LLM into a multi-dimensional taste vector and fed directly into the next selections. Requests act as a strong positive signal for everything the prompt describes. Not just the next song, but the next five. And they decay slowly over the next 30 minutes unless you steer again.

Implicit signal: dwell and skip

Did you let a track play to the end? Did you skip in the first 15 seconds? These are quiet signals but enormously useful in aggregate. Dwell-through is a soft Love. Early skip is a soft Retune. The model weights them about 1/4 as much as the explicit equivalents, but you produce hundreds of them per session, so they end up shaping a lot.

Implicit signal: time-of-day shaping

The model learns what you listen to in the morning, at lunch, after work, and late night, separately. The same Love signal at 8am and 11pm pushes the radio in slightly different directions because the time-of-day cluster captures different intent. Most users don’t notice this consciously. They just notice that the radio "feels right" at the times they listen.

Implicit signal: device

Mobile vs desktop vs car (when we get there) are quietly different listening contexts. The model doesn’t change what you like across them, just which subset of your taste it pulls from. Car listening tends toward lower-density, smoother selections; desktop tends toward more rhythmic / focused tracks. Nothing you do has to opt in to this; it just happens.

What we deliberately don’t do

We don’t read your email. We don’t read your calendar. We don’t read your contacts. We don’t read your clipboard. We don’t share your listening with advertisers (musen has no advertising). We don’t enrich your profile with third-party data. Your taste profile stays inside your account; you can export it any time and delete it any time from Account > Privacy. If you want a daily audio briefing built from your calendar and email, use Huxe instead. It’s a great tool for that, and we’re happy to point you at it.

How long until your radio feels "yours"?

Within five minutes of pressing play for the first time the radio already feels reasonable, because the cold-start model is good. Within a session it starts to converge on your particular interests. Within a week of regular listening the model has enough Love/Retune/Request data that first-time listeners and long-term users get noticeably different streams from the same prompt. Within a month the radio is recognisably tuned in ways you didn’t consciously train.

Retraining cadence

The taste profile updates in near real-time. a Love right now will influence the very next selection. The underlying ranking model retrains in the background on aggregate (anonymous) data on a weekly cycle. There is no manual "retrain" button because there shouldn’t need to be one: the radio is always converging on your current taste.

FAQ

Can I reset my taste profile? Yes. Account > Privacy > Reset. Can I have multiple profiles? Multi-profile is on the roadmap (so partners with different taste don’t blend). Does musen sell my data? No. We do not, will not, and have no business model that requires it. What if I share my account? The model will blend the two tastes; you’ll get a worse experience than separate accounts. Use separate accounts if your taste differs.