In Watch a model learn you rolled a ball down a loss curve. Here you get to feed the model. Click to drop points of two classes; a tiny neural network trains in real time and paints the boundary it has discovered.
Things to try:
- Drop two clean clusters, the boundary snaps between them almost instantly.
- Now interleave the classes, or draw a spiral or a ring. Watch the boundary bend into curves a straight line never could, that’s the hidden layer doing its job.
- Add a single outlier deep in enemy territory and see the network agonise over it.
Under the hood it’s exactly the loop from the other explainers: predict, measure error, nudge the weights downhill with gradient descent, just many times per second, on the points you drew. Scale this up to billions of parameters and trillions of examples and you get a large language model.