Opus 5.5 writes a 169-line JS neural net that sorts a spiral at 99.2%
Claude Opus 5.5 wrote forward pass, backprop and Adam from scratch in 169 lines of JavaScript, training on two-spiral data to lift accuracy on unseen test points from 24.5% to 99.2% and rendering it as a 1080p animation.
This network had never seen a single test point and still got 99.2% of them right. Before training it scored 24.5%, below the 33% you'd get by picking colours at random. The task is the classic three-arm spiral. 600 training points, 3 colours, and you can't separate them with straight lines. The network has 2 inputs, two hidden layers of 32 neurons and 1,251 weights. Epoch 0: 24.5% on unseen points Epoch 100: 98.7% Epoch 150: crosses 99% Epoch 1,500: 99.2% test, 99.7% train The left side is the network's guess for every spot on the plane. The 16 tiles on the right show what individual hidden neurons react to. Each one draws a simple wavy cut. The output layer combines them into the spiral. Claude Opus 5.5 wrote the whole thing from scratch in Claude Code: forward pass, backprop and the Adam optimizer in 169 lines of JavaScript, no ML library. Training takes 10 seconds, the 1080p render about 2.5 minutes.
This post quotes another post: “Opus 5.5 writes 105 lines of JS to render a starling murmuration”