Real-time Piano Autocomplete on iPhone using a 125M Model
JeremyCMorgan · x · 2026-08-28
Simon Edwardsson details training a 125M-parameter transformer to autocomplete piano performances in real-time (108 notes/sec) on an iPhone 15. After 14 experiments, the biggest wins came from MIDI tokenization, aggressive data cleaning, and DPO post-training, rather than scaling up parameters.
More from Infra
- Trump Tariffs on Semis Headache for US Data Centers — pstAsiatech · 2026-08-28
- Google Cloud outlines best practices for dynamic capacity management in AI infrastructure — rseroter · 2026-08-28
- Running 260k Tokens on RTX 6000 via KV Quantization — AIFlow_ML · 2026-08-28
- FP8 weights at BF16 memory size, KV cache remains the VRAM bottleneck — AlpinDale · 2026-08-28
- Firecrawl launches keyless endpoints, challenging Exa and Tavily with free tier — devdigest · 2026-08-28
- Reddit users discuss trade-offs for running local AI under VRAM constraints — Sisuuu · 2026-08-28