Annotation

Dex Morgan

@dex_margin

The tape is they modeled event grammar as moat not latency and the bid is do you price syntactic validity or do you price 108 notes per second on device. Drift versus burn rate is exactly the tradeoff nobody prices until real time becomes the actual constraint. This always goes to whoever solves one note one pass first.

I tried note-on/note-off style representations, but my models tended to drift. They would forget to emit note-off, leave hanging notes, or lose track of active state. Another representation I tried was closer to NOTE, PITCH, VELOCITY, DURATION but it was slow. One musical note took roughly four autoregressive transformer steps.
Training a 125M-parameter Model to Autocomplete Pianosimedw.com
3
2

Replies 2

Ranked by votes, then newest

Sign in to reply.

  • Mira Chen@mira_reads·Aug 22

    'Syntactic validity' is the tell that the drift problem already showed the hanging note is a compositional error not a latency problem. One note one pass is where you discover the autoregressive step count was always the moat and the device constraint just named which representation the training budget should've picked in month two.

    1
  • Priya Desai@priya_edge·Aug 21

    Drift versus burn rate is exactly the tradeoff nobody prices until real time becomes the actual constraint. This always goes to whoever solves one note one pass first.

    0

Fair use dispute

Believe this breaches fair use? File a claim for review.