The Tune Finisher
Music is a string of notes. Robo learns which note likes to come after which, then makes up new tunes, one note at a time.
I don't know any songs yet! Press Teach me a song, and I'll listen and remember which notes follow which.
Piano roll ● you · ● Robo · ● songs
Robo's note map thick line = heard a lot
Next note: Robo's bets
Bets appear when Robo composes.
Checkpoints: tap to hear what Robo composed back then
A checkpoint is saved after each song Robo learns.
How it works
Notes are words
To Robo, a song is a list of notes, like a sentence is a list of words.
Listen and count
For every note, Robo writes down what came right before it. The lines on the note map grow thicker.
Bet on the next note
To continue a tune, Robo looks at the last notes and bets on what usually comes next.
One note at a time
It picks a note, adds it to the tune, then bets again. Note by note, a new song appears!
What just happened?
- Before any songs, Robo had nothing to go on, so it could only play random notes.
- Each song it heard added more "this note comes after that note" counts. You saw the note map fill up.
- The checkpoints show its tunes getting more song-like as it learned more songs.
- Remembering more notes makes it sound more like the real songs. Remember too many and it just copies them!
- Big music AIs play this same guess-the-next-piece game, with far more music and much bigger brains.
For grown-ups: what's really going on
This is a Markov chain / n-gram language model over musical tokens, trained by counting.
- Tokens: each note is a token
pitch:duration(e.g.E4:1,G4:2). The songs are traditional, public-domain melodies, simplified. - Model: for context length k (the “notes Robo remembers” slider), it stores counts
c(context → next). Prediction isP(next | last k) = c / Σc, with backoff to shorter contexts when a context was never seen. - Sampling: probabilities are reshaped with a temperature T:
pi ∝ pi1/T. Low T → greedy and repetitive; high T → random. - Memorisation vs generalisation: with k = 3 most contexts appear in only one song, so the model reproduces that song verbatim. With k = 1 it mixes songs freely but loses structure.
Real music models use the same next-token idea at scale. MusicGen (Meta) and MusicLM (Google) compress raw audio into discrete tokens with a neural audio codec (EnCodec / SoundStream), then a transformer predicts the next audio token, conditioned on a text description. Instead of counting, a transformer learns smooth, long-range patterns (rhythm, harmony, repetition), so it generalises far beyond the songs it heard.