Slide a 3×3 kernel over a picture and watch every multiply-add. The same nine weights are used at every position, which is what makes a convolution cheap and an edge detector work anywhere.

Max and average pooling shrink a feature map, kernels stack into channels, and each layer down the network sees a wider patch of the image than the one before.

Draw a digit and a real network trained on MNIST reads it: every feature map of every layer, then the ten probabilities. Edges first, then strokes, then digits.

Unroll a network over a sequence and carry its memory as a vector. Train one live to count, then see exactly where its memory runs out.

Gates are numbers between 0 and 1 that decide what to keep, forget and output. Watch trained gates hold one number while noise streams past, where a plain RNN drifts.

An encoder squeezes a whole sentence into one vector and a decoder writes the translation from it. The longer the sentence, the more gets lost — the problem attention was invented to fix.

Twenty plain layers fit a spiral worse than four. Train deep stacks with and without skip connections side by side, and watch adding the input back keep the gradient alive through every layer — the backbone of every transformer.

Squeeze 144 pixels through two numbers and back. Train it live, then drag a point around the two-number space and watch the decoder paint the picture it stands for.