an interactive intro to mechanistic interpretability

train a neural network.
then look inside it.

most tools show you a network learning. peek shows you what it learned. train a tiny net in your browser, then trace a data point through it, hover any neuron to see what it detects, and watch the decision boundary form.

network+
01

watch it think

pick any data point and trace it through the network. watch which neurons light up, layer by layer.

02

x-ray a neuron

hover any neuron to see exactly what region of space it responds to. early ones make simple cuts, deep ones bend into shapes.

03

see the boundary

a heatmap shows where the network draws the line, redrawing live as it trains. add input features and watch it shift.