AI that composes
Pictures done — now sound.
Sound is a completely different thing, so surely it needs a different method?
Let's make some first.
Everything you pressed sounded like music
That isn't the AI being clever.
There are only eight notes you can press. They all come from one scale, so nothing clashes badly.
We narrowed down what you can press at all. We did that to keep it fun — the AI had nothing to do with it.
What did the AI look at?
When you pressed Let the AI carry on.
Look at The table it counted, below the keys.
It's a count of which note came next after each pair of notes, with the numbers written out.
Notes that came up a lot get picked often; rare ones get picked now and then. That's all of it.
Wait — you've seen this before
Remember the fairy tales in 4.2.
It counted which word followed which, then picked by those counts.
Side by side:
| 4.2 Guessing the next word | 5.2 Making a melody | |
|---|---|---|
| Material | fairy tales | five nursery rhymes |
| What it counts | the word that came next | the note that came next |
| How it picks | by the counts | by the counts |
Not one part of the method changed. Only the material did.
This is what AI is like
Build one method and you can swap the material underneath it.
Guessing the next piece isn't only for words. It works on notes, on pictures, on video.
The from 5.1 is the same story. It isn't only for pictures — it makes sound too.
Method and material come apart. Want to make a new kind of thing? Take an existing method and point it at that.
It can only imitate what it learned
Choose London Bridge on its own.
The melodies start sounding like London Bridge. It has never heard of anything else.
Switch back to All five and things mix a bit more.
What you give it decides what comes out. Same point as 3.3.
But is that really composing?
Ours is one table. Counting, then joining things up.
Music-making AI today uses a neural network instead of a table, and hundreds of thousands of songs instead of five.
But the bones of the job are the same. Pick a likely next thing.
And those hundreds of thousands of songs were written by people. Everything 5.1 said applies here too.
When making things gets easy
You can make pictures now, and music.
Which also means convincing fakes just got easy.
That's the next chapter.