Donuts and gearboxes

He nearly fell in again, another experiment, another suspicious number, “y ahí decidí: no más, por favor” ▸ 2:10. So the response to the whispered exam is structural, not behavioral: strip the project to its essentials and redesign it so cheating is impossible ▸ 2:52. Inference now runs in a sealed room: the model receives images only, no labels, no errors, nothing, and moves its window until it commits to answers; only afterwards does a separate stage, “donde está el profesor,” pull out the answer key and grade ▸ 3:49. No channel for the whisper to travel.

The clean baseline is humbling, around 40% at 20,000 parameters, and the climb back happens by his own systems law, stated once more for the record: a system grows as much as its environment and tools permit, and the bottleneck here was movement ▸ 9:53. Fix one: the image becomes a donut, Pac-Man topology, exit left and reappear right, so no movement is ever wasted against a wall and the window never stares at empty repeated pixels in a corner ▸ 6:24. Fix two, the winner among the options he had the AI propose: not the saccade grid he rejected ▸ 8:11, but “gear movement,” a gearbox action added to up-down-left-right-zoom, letting the eye shift speed and cross the image in a few strides instead of twenty single-square steps ▸ 8:37. Result: 77%, same tiny network ▸ 9:07.

no agrandes la red: agrándale el mundo y las herramientas →

A bigger network also scores higher, but that’s the boring axis; the discipline is holding size constant and hunting the elements that unlock the architecture ▸ 9:13. Next levers: the curriculum, then a better eye ▸ 10:49. The numbers are smaller now and mean something, which is the whole trade…

Postscript, two days later: 80% clean, and a new honest problem, policy collapse. The agent picks one lazy trajectory, always left-and-zoom, or always down, and repeats it for every image ▸ 13:09. Two hypotheses tested and dead: training classification and movement as separate networks ▸ 14:36, and richer 30%-jump movements, after which it still repeats the same actions “una y otra vez” ▸ 17:22. Back to comparing against the original implementation, whose gaze, he remembers, actually looked at things.