What's this about?
A Show HN describing training a 125-million-parameter model to autocomplete piano performances, running on-device. You play, and it continues in your style.
What exactly makes MIDI a good fit for a small model?
MIDI isn't audio — it's a compact event list: this note, this loudness, this moment, released here. That's structurally close to text, so language-model architectures apply directly, and the vocabulary is tiny compared to natural language. Music also has strong local structure — keys, rhythms, chord progressions — that a small model can capture, since you don't need world knowledge to continue a phrase plausibly. The result is a model three orders of magnitude smaller than a frontier LLM that does something genuinely useful, fast enough to feel instant while you play.
Why is this worth noticing?
HN 513 points, 107 comments
- It's a corrective to the week's economics. Amid trillion-parameter costs and memory prices up 500%, a 125M model doing a delightful job on a laptop is a reminder that model size follows the problem.
- On-device matters for music specifically, because latency is the whole experience — a round-trip to a server would make it unusable regardless of quality.
- It's the kind of project that used to define HN, and its score suggests an appetite for that in a week of acquisitions and outages.