Writing


Mostly about evaluation — building the measurement before trusting the result, and reporting the one that kills your own idea. Every post links to the code.

I built it for the data center. They demoed it on a laptop.

Aug 2026 · ASR

Cutting 14.7% of a Whisper model's parameters with no retraining, by calibrating on the audio of the one speaker who'd actually use it — the sharp cliff at the fifth layer, and why two compression techniques that are each nearly free break the model when combined. Code: mgoldey/ablation.

More at matthewgoldey.substack.com.