September 22, 2026 · 5 min read
Generators make songs. Instruments make musicians.
Three kinds of tool are sold with the same two words. One makes songs. One makes practice. One, we think, makes musicians. Here is how to tell them apart and why we built the third kind.
Type "melancholy lo-fi beat with rain" into a generator and, a minute later, you have one. It is good. It has the vinyl crackle and the seventh chords and the lazy snare. You could put it under a video tonight.
Now try to make the snare a little later.
You cannot. There is no snare. There is a three-minute audio file, and inside it the snare is a shape in a waveform that the model decided on. Your options are a different prompt and a different song. The thing you wanted to change is not a thing; it is a pixel in a picture.
This is not a complaint about generators. It is the definition of one. A generator is a machine for producing finished songs, and finished is the point. But "AI music" now covers three products that have nothing in common except the two words, and if you want to make music, it matters which one you pick up.
The three kinds
A generator takes a description and returns audio. Suno is the famous one. You get a song; you do not get the notes, the parts or the reasons. Editing means generating again, or, on the top plan, cutting the audio up in a timeline. The model keeps the music theory. You keep the file.
A tutor listens to you play a real instrument and scores you. Yousician and Melodics are the well known ones, and they are good. They make your hands better. They never ask you to write anything; you play along with their songs, and when you finally want one of your own, you are back at a blank DAW with the same questions you had before.
A co-producer lives inside an instrument. You are making the music: a beat on a grid, a bassline in a key, chords from a recipe. The assistant changes that project when you ask, using the operations you use, so the change lands on the same grid and can be seen, edited and undone. The new chat assistants for Ableton Live are this kind. So is ours.
Only the third kind can answer the question "why does this work?" about your song, because only the third kind is working on your song.
Hear what a generator cannot show you
This is a beat. Not a picture of one, not a description of one. Press play, and then look at where the snare lands.
The snare is on the two and the four. That is the backbeat, and it is why your head nods. Now here is the same beat with the kick moved to fight the snare instead of leading it:
It stumbles. You can hear exactly why, because you can see exactly what changed: one pad, one step to the left. A generator could make either of these beats. It could not show you the difference, because it has no pads to show.
That is the whole argument, and it fits in two loops.
What a co-producer does with this
Ask the co-producer here to make the first beat lazier. It will move the kick on the second bar a step late, or add a ghost note before the snare, or push the swing up. Then it says, in a sentence, what it did and why: "the kick now lands late on bar two, which is the lean that makes boom bap feel unhurried." The change is on the grid. If you do not like it, one press takes the whole turn back, and you can try the same thing with your own hands.
That loop, ask, see, read why, undo, try, is how people have always learned production. It used to need a friend who produces and a great deal of their patience. Making it a feature meant three design rules:
- The assistant uses the same operations you use. It has no private way to change the project, so everything it does is visible and editable.
- Every turn is one undo. However many things it changed, one press restores the project. Bold requests become safe.
- Every change comes with a reason in the words of the music, not a log of edits.
None of those is possible for a generator, because a generator's output is audio and audio has no operations to share. They are possible for an assistant inside a DAW, but the DAW makes them hard: forty edits through Live's API are forty undo steps, and the explanation has to make sense to someone who already knows Live.
An instrument small enough to see whole makes all three easy. That is why ours is small.
Which one to want
If you need a finished song with vocals tonight, want a generator. It is the right tool and it is astonishing.
If you want to play guitar better, want a tutor.
If you want to be able to make music, and to understand what you made, want an instrument with a co-producer. It is slower to a finished song and faster to knowing what you are doing. A year from now, the generator's user has a folder of songs and the instrument's user has a folder of songs and can make another one.
Generators make songs. Instruments make musicians. We think the second thing is worth building, and it is free to start.
Questions people ask
What is the difference between an AI music generator and an AI co-producer?
A generator writes a finished song from a description and hands you audio. A co-producer works inside an instrument on the music you are making, using the same operations you use, and the good ones explain each change and can be undone in one step.
Can you learn music from an AI generator?
You can learn what genres sound like and how to write prompts. You cannot open the notes, so you cannot learn why a bassline works or what makes a chorus lift. That knowledge stays inside the model.
Is an AI co-producer better than a generator?
It is different. If you want a finished song in a minute, a generator is better. If you want to be able to make music yourself, an instrument with a co-producer is the one that gets you there.