Draw a digit and watch a neural network read it.
Every layer below runs live in your browser. The network is a 207k-parameter CNN trained with glassbox, a deep learning framework written from scratch in NumPy: no PyTorch, no TensorFlow.
Loading the model…
Inside the network
Each panel is one layer's output for your drawing. The more strongly a unit fires, the more it stands out from the panel, relative to the strongest in that layer.
Your drawing as the network sees it: 784 pixels, prepared the same way MNIST was.
16 filters, each tuned to a simple pattern such as an edge or a curve.
32 filters that combine those patterns into strokes and corners.
128 neurons, each weighing evidence from the whole image.
Raw scores. Softmax turns them into the probabilities above.
A small GPT that learned to write like Shakespeare.
It writes one character at a time, and every weight was learned from scratch with glassbox on a laptop CPU. Hover or tap any character to see which earlier characters an attention head looked at while writing it.