Silicon designed with intelligence built in.

Browser sandbox for a small neural net. Add layers, swap activations, move the learning rate, and watch the decision boundary redraw as it trains.

A machine learning textbook where the algorithms animate — gradient descent, backpropagation, and diffusion models with interactive controls.

Watch signals propagate through a network architecture, with interactive linear regression, classification, and CNN views.

The same sandbox preloaded with the hardest built-in classification problem. Trying to solve it is the quickest way to feel why depth helps.

A GPT model rendered in 3D. Follow a single token through embeddings, attention layers, and MLP blocks at your own pace.

Type a prompt into a live GPT-2 and see tokenization, attention weights, and next-token prediction. Hover a matrix to see the arithmetic.

Jay Alammar's diagram-led walkthrough of the architecture — the standard companion piece to the original attention paper.

The follow-up post: how a decoder-only model actually generates text, one step at a time. Read it after the transformer piece.

Attention patterns from real transformer models across many inputs, showing how individual heads specialise on syntax or meaning.

A notebook-based attention viewer for BERT, GPT-2, and friends. Paste any sentence and inspect the heads that fire on it.

A live image classifier you can click into layer by layer to inspect feature maps, filters, and activations.

Compare how VGG, ResNet, and MobileNet extract features from the same images, and see which neurons respond to which patterns.

Convolution made concrete: drag kernels over an image and watch edge detection, blurring, and sharpening happen in the browser.

Pretrained models such as LeNet, AlexNet, and YOLO drawn in 3D, so you can rotate the stack and follow an input through every layer.

Brown University's animated course in probability and statistics — Bayesian inference, regression, and distributions, all interactive.

Small, self-contained demos for PCA, eigenvectors, Markov chains, and conditional probability.

Research articles built around interactive figures, covering feature visualisation, what networks see, and AutoML.

Gradient descent traced across a loss landscape, with SGD, momentum, and Adam side by side so the differences are visible.

Place your own training points, then watch KNN, SVM, decision trees, perceptrons, and neural nets each draw a decision boundary over them.

R2D3's scrolling story of a decision tree learning to tell San Francisco homes from New York ones. The best first thing to read.

Part two of the same series: why a deeper tree overfits, shown happening rather than asserted.

Drop your own centroids and step the algorithm to convergence. Overlapping clusters make for instructive failures.

Density-based clustering, stepped through — and a clear look at the odd cluster shapes K-means cannot handle.

Train a generative adversarial network in the browser and watch the generator and discriminator compete, with live loss curves.

Stable Diffusion turning noise into an image, one denoising step at a time, with the prompt and guidance scale under your control.

High-dimensional embeddings projected into 3D. Load Word2Vec or GloVe vectors and search, rotate, and zoom through the clusters.

Andrej Karpathy's reinforcement learning demos: an agent learning to cross a gridworld under Q-learning, SARSA, or policy gradients.

Open an .onnx, .h5, .pth, or .tflite file and read the whole architecture as an interactive graph. Useful for any model you did not write.