The digging is automatic.
The taste is yours.

Language models are one kind of AI. Audio models work with a different input: sound. One useful idea connects the two— turn messy human things into long lists of numbers, so that meaning becomes geometry.

Those lists are called embeddings. Language models embed words. Vision models embed images. Brew.fm uses a model that embeds sound itself - no lyrics, no genre tags, no play counts. That audio signal is one part of Brew.fm’s recommendations, alongside release information, shared-playlist signals and editorial selection.

Here are two songs Spotify files under the same label. Play them both.

“Indie pop” contains multitudes; “pop” means even less; half of what you love sits in genres you’d never click. Any system built on words inherits the lies words tell. So we don’t describe songs. We measure them.

A neural network trained on millions of clips listens to thirty seconds of audio - the texture, the tempo, the space between the notes - and writes down 512 numbers.

Those 512 numbers are the song’s fingerprint. Two songs that feel alike get numbers that are close. Two songs that feel like different planets get numbers that are far apart. Nobody tells the model what “alike” means - it learned by listening.

Read the 512 numbers as coordinates and every song becomes a point in a 512-dimensional space of pure sound. Distance means similarity. That one move - sound becomes geometry - powers everything below.

Your screen has two dimensions; the real map has 512. The math doesn’t care - the distance between two points works the same in 512 dimensions as it does in two.

Take every song in a playlist - every point - and average their coordinates. You get one new point: the centroid. It isn’t a song. It’s the sound the whole playlist orbits. We call it the heart.

The heart is how a machine can hold a vibe without understanding a single word about it. Anything close to the heart belongs; anything far doesn’t - and you can hear that boundary with your own ears, above.

Plot everything you’ve ever saved and it doesn’t make one cloud. It pools into a few dense regions - because nobody is one person all the time. A clustering algorithm finds those regions with no labels and no hints, just density.

Each dense region is one of your alter egos. The sound is discovered by geometry; only then does a language model step in for its single job - giving the cluster a name worthy of it.

You are a cluster with a heart. The artist you’d kill to work with is another. Draw the line between the two hearts and walk to the exact middle. The songs already living there - close to you, close to them - are the playlist you’d make together.

Not “people who like X also like Y.” No taste graph, no popularity contest. Geometry: the literal midpoint between you and your favorite artist, in the space of sound itself.

Saving the collab isn’t the end - it’s a birth. While it’s alive, candidate tracks are compared with your playlist’s sound. Taste-matched modes use that signal to help decide what belongs. It is a filter, not a guarantee: an artist can take a new direction, and a recommendation can still miss.

That is the audio-matching part of the process. Release tracking and shared-taste discovery supply candidates; the matching helps narrow them down.