The nine gets a zoom

Preparing the 99.9% result for publication, he builds GIFs of the model’s gaze, and the policies it learned become legible. Most digits share one lazy strategy: drift the window downward and call it ▸ 13:21. The nine is the exception, every single time it zooms into the little circle, the one feature that separates it from its neighbors ▸ 13:30. The one is stranger: the model never looks at anything, it infers a one from the absence, if I see nothing, it’s a one ▸ 13:47, and a six gets called from three pixels in the right place ▸ 13:57. The rainbow-path plots confirm it at the population level: the trajectories are almost identical for every class except the nine ▸ 16:06. Interpretability for free: when your model has eyes, you can watch it decide where to look ▸ 16:25.

el modelo con ojos se audita mirándole los ojos →

The day’s debugging story earns its place next to the pretty pictures. CIFAR-10 training stalled at 30% accuracy, and the cause wasn’t the idea, it was the assistant: Gemini had quietly changed the architecture to train after every glance instead of propagating the final answer back through all of them, which he dismantles with one question, “¿de qué va a entrenar si todavía no sabe la respuesta?” ▸ 10:56. Asking it to re-review its own program surfaced the rewire ▸ 11:33, and with the loop restored, KMNIST, Japanese characters, harder than digits, trains to 87% on the first corrected run ▸ 24:44, the architecture’s second foreign dataset after Fashion-MNIST. CIFAR gets its rematch next, with the question he actually cares about queued behind it: what gaze patterns does a model invent for airplanes and birds ▸ 16:25? Then the write-up, then Bryan, then public…

Postscript, next day: CIFAR gets its rematch and lands at 92% test accuracy after 60 epochs, once he noticed at epoch 30 that “la red como que quería seguir siendo entrenada” ▸ 5:29. The anomaly he can’t stop staring at: training accuracy sits around 60-70 while test hits 90, the inverse of the usual memorization gap, which he reads as the architecture abstracting essentials instead of memorizing pixels ▸ 7:05.

Postscript, the correction: the inverted gap was not generalization, it was the symptom. The pipeline leaked the per-step error to the agent, and the uniform gaze trajectories in this entry’s own GIFs were a second symptom hiding in plain sight: the eye barely needed to look. The autopsy is entry 186-1 ▸ 25:01.