By the morning of September 26, the model on unit one was taking about 14 seconds to work through four seconds of signal. Kris Droste wanted the IntoMind One to run in real time, every window finished before the next arrived, and it was more than three times too slow.
He asked the training machine for two things: first the heads for age and sex, and then "a distillation ladder to see how small the foundation model can go while still beating or matching psd". PSD, the power spectral density, is the measurement band power is computed from. A ladder trains a series of ever smaller students the same way and stops at the first one that fails.
Two lines of work ran through the day. On the device, the firmware got faster: 14.13 seconds per pass, one run of the model over one window, then 9.97, then 8.63 by midday. It was still too slow.
On the training machine, an AI agent built the heads and climbed down the ladder at the same time, in a run of five hours and twenty-four minutes. A student of 240,480 parameters passed. One of 123,072 fell short, narrowly, at telling women from men. One of 170,772, trained in between, passed. That made 170,772 the smallest rung that still matched band power. It needed 14.6 million multiply-adds per window, the basic arithmetic steps of a model, against 43.5 million for the model on the device.
That night Kris ruled that it would ship. Nineteen minutes later it went onto unit one over Bluetooth, and after a restart early the next morning, each pass took 3.15 seconds. The device described a window every 4.2 seconds and lost no samples.
It was not free. Tests that followed showed that on every test of its heads, the smaller model did a little worse than the larger one.
On the firmware that units ship with, 1.4.5, the windows run back to back, and a pass takes 3.504 to 3.507 seconds, measured over 12 passes. It fits inside four.
The model was later named Mu 171K.