Austin's Nerdy Things

Checking Our Work: What Does the Raw Data Actually Say?

Those numbers above come from what the servos report. ptp4l prints a 1 Hz RMS summary. chrony’s sourcestats shows the standard deviation of its filtered, averaged output. Both are honest numbers, but they’re the numbers after each servo has done its best to smooth things out.

Background: What Is RYS?

Ng’s technique is called RYS and the core concept is surprisingly simple. A normal transformer forward pass goes: Process input through layers 0, 1, 2, …, N-1 sequentially Done With RYS, you pick a contiguous block in the middle — layers i through j-1 — and after the model finishes layer j-1, you jump back and re-execute layers i through N-1. Those middle layers run twice on the evolving hidden state.

Will This Actually Run on a Phone?

The whole point of this exercise was to run inference on an iPhone. So here is some napkin math: ModelGFLOPsEstimated Phone InferenceNano1.6~15ms, smooth 30fps easyNanoish3.2~25ms, still goodSmall26~80ms, might be sluggishYOLOv8n (for reference)8.7~27ms My app was already running YOLOv8n at 15fps with plenty of headroom.