i wanted to try machine learning. i built a convolutional network in pytorch that classifies fashion-mnist — sixty thousand labelled images of clothing, learned by example. the model doesn't read rules. it looks at pictures until it can tell a shoe from a shirt
the eyes learn
fashion-mnist was a good first try because it's honest and familiar. the network stacks convolution and pooling, learns shapes, proves itself on the test set. a computer that has never worn clothes recognizes them. the abstract stuff — tensors, loss, gradients — finally became a thing that actually saw
next to it, project-euler kept the math sharp, a problem at a time. same rhythm as the rust notes: a problem, a tool, a working answer
the turn of the year
new year's day, two rust things shipped close together. a server that queries logstash's esql and renders the data in the terminal — the server and the tui in one program. and a full terminal tetris, released for windows and linux, complete with levels and scoring
i wasn't aiming at anything specific. i just liked making things, and by then i'd made servers, a game, and a model that could see
视 is to see — and the machine could see. the terminal, the net, and the eye were all things i'd built. i didn't know yet what any of it was for