IdeaWell

Project ideas from Hacker News discussions.

Show HN: I trained a 125M model to autocomplete piano on-device

Original Article

Hacker News Discussion

📝 Discussion Summary (Click to expand)

1. Enthusiasm & “mind‑blowing” reaction

“This is really awesome thanks for sharing” – yoouareperfect
“I find your application absolutely magic.” – pil0u

The thread is dominated by astonishment at being able to generate music on‑device in real time and share it instantly.

2. Desire for richer musical control & planning

“I think adding some kind of bar/measure token might help with rhythm, and perhaps some form of longer‑term planning for the overall composition.” – isoprophlex
“Yes, some kind of planning step is on my TODO list.” – simedw

Users want the model to go beyond single‑note prediction: handle measures, structure, and parallel continuations.

3. Future directions – accompaniment, MIDI export & richer expressivity

“I would love something like that, except that I play the melody, and it produces proper 3‑4‑part accompaniment, preferably in good baroque style.” – heikkilevanto
“Another thing I want to try is generating a few continuations in parallel, picking the one that looks best, and then continuing from there.” – simedw

The community is already thinking about exporting MIDI, adding velocity/attack, and making the system react like a real‑time collaborator.

🚀 Project Ideas

Generating project ideas…

Real‑Time AI Accompanist for Live Performance

Summary

  • Generates on‑the‑fly chordal/harmonic accompaniment that adapts to user‑played melodies, handling rhythm and phrasing.
  • Provides seamless MIDI output for integration with any synth or DAW, eliminating latency and format‑conversion headaches.

Details

Key Value
Target Audience Amateur musicians, piano teachers, live‑performance hobbyists
Core Feature Real‑time style‑aware accompaniment using a lightweight transformer that predicts chords, voicings, and rhythmic patterns from incoming MIDI notes
Tech Stack Core ML + Core Audio on iOS, SwiftUI UI, Core MIDI, TensorFlow Lite for inference, optional WebAudio fallback
Difficulty Medium
Monetization Hobby

Notes

  • HN commenters explicitly requested “proper 3‑4 part accompaniment,” “reacting to what you’re playing in real time,” and “MIDI to a device” – this directly addresses those calls.
  • Enables instant musical experimentation without needing a second player, fostering deeper engagement and community discussion.

AI‑Assisted Musical Idea Planner (CoT Generation)

Summary

  • Adds a “chain‑of‑thought” planning layer that expands simple melodic seeds into complete arrangements with thoughtful harmonic choices.
  • Prioritizes quality over speed, trading note‑per‑second for richer, more musically coherent continuations.

Details

Key Value
Target Audience Composition hobbyists, educators, creators seeking higher‑quality AI‑generated material
Core Feature Two‑stage generation: (1) seed‑melody input, (2) planning module that proposes 4‑8 bar structures before melodic expansion
Tech Stack PyTorch Lightning, Hugging Face Transformers, ONNX export for desktop/web, optional Gradio UI
Difficulty High
Monetization Revenue-ready: Subscription tier: $5/mo for premium planning credits + $0.01 per generated bar

Notes

  • Users suggested “CoT for better quality” and “generating a few continuations in parallel”; this solution implements those ideas while giving a clear upgrade path for power users.
  • Sparks discussion around AI‑driven compositional heuristics and educational use in music theory curricula.

Mobile‑First AI Melody Composer with Offline MIDI Export

Summary

  • A lightweight iOS app that captures hum/singing via microphone, converts it to MIDI, and offers AI‑driven accompaniment with offline MIDI file export for editing in standard DAWs.
  • Optimized for on‑device inference to run at >100 notes/sec on modest hardware.

Details

Key Value
Target Audience Mobile musicians, casual creators, educators wanting on‑the‑go ideas without cloud dependencies
Core Feature End‑to‑end pipeline: mic → pitch‑track → tokenized melody → AI‑generated continuation → MIDI export (Standard MIDI Files)
Tech Stack Core ML (quantized transformer), Core Audio, SwiftUI, MIDIKit for file generation, Metal acceleration for speed
Difficulty Low
Monetization Hobby

Notes

  • Directly fulfills requests for “whistling or playing via microphone” and “audio as midi notes to a device,” providing an offline, shareable MIDI output that integrates with existing music‑production workflows.
  • Positions the project as a hobby‑first tool that could later expand to a revenue‑ready tier with premium soundfonts.

Read Later