Learn
Working notes on the machinery underneath — how these systems actually compute, not how they are usually described. Interactive where interaction earns its keep.
What Attention Actually Does
Against the MLP trained to a perfect score in part one, take apart the one step a transformer adds: “where to read” is computed live — three on-the-spot proofs; multi-head ＝ reading quota, with 1/2/4 heads trained in front of you to watch the division of labor grow; and why the FFN widens then shrinks. The code panes show real source — the lines shown are the lines running.
Know arithmetic? Then you can train a neural network
Build a 2,554-parameter MLP on all 100 products from 0×0 to 9×9. Follow one causal chain and one shared parameter to derive loss, gradients, batch training and backpropagation — with the boundary between fitting 100/100 and generalisation stated explicitly.
The extended Kalman filter, starting from intuition
Where the Jacobians actually come from, why the covariance update is written the way it is, and what breaks when the linearisation is bad. Derived rather than quoted — then fused live from GPS, a wheel encoder and an IMU.
A transformer small enough to read
The unchanged pure-web lab: an 8D, two-layer, two-head decoder-only Transformer with 3,136 parameters. TensorFlow.js and its WASM backend are vendored locally; training, inference, parameter tables, traces and the 3D walkthrough all run inside the browser.