Train small models on 2D data you can edit point by point, step through the training loop, and inspect every formula with its real, live values bound in. Nothing on screen is decorative — if the code didn't compute it, the page doesn't show it.
Hover any highlighted term in a formula and the matching visual element lights up with the term's current value. Every run is seeded and deterministic.
Gradient descent rolling down a real loss surface, next to the closed-form answer.
A probability heatmap, the sigmoid, and the weight vector that draws the boundary.
Watch splits appear one at a time, with the exhaustive gain scan that chose them.
Bootstrap samples, a strip of weak learners, and the vote that averages them.
Bagging plus random feature choice per split, with out-of-bag error for free.
Trees fitting residuals round by round, with the regularized gain formula live.
Histogram binning and leaf-wise growth, side by side against level-wise.
A playground-style network with per-neuron activation maps and visible backprop.
Toy attention with real Q/K/V math and a hand-wired induction circuit.