Algorithmic thinking
The artificial neuron
McCulloch–Pitts Neuron
Explore the math behind the first model of an artificial neuron: binary inputs, weights and a firing threshold.
Single Neuron & Activations
Compare a McCulloch–Pitts neuron with and without activation functions like step, sigmoid, tanh and ReLU, and see how each shapes its output.
Perceptrons & Shapes
See how each perceptron draws a single line, and how combining lines with activations carves out more complex shapes and decision regions.
Multilayer Perceptron
Stack perceptrons into hidden layers and watch a multilayer network train, with live visualizations of its weights, loss and decision boundary as it learns.
Learning from data
Iris Classification
Apply classification algorithms to the classic problem of telling Iris species apart from petal and sepal measurements, and compare their decision boundaries.
Gradient Descent
Watch basic gradient descent walk downhill on a loss surface, and see how the learning rate changes its path.
Backpropagation
Follow gradient descent through a network as errors flow backward and every weight gets its update.
Backprop Under the Hood
Work through backpropagation one step at a time: apply the chain rule, compute each local gradient, and watch how every weight in the network gets adjusted during training.
Measuring fit
Loss Landscape
Compare loss functions such as mean squared error, mean absolute error and cross-entropy, and see how each one shapes the landscape a model has to descend.
Entropy, Cross-Entropy & KL
Explore Shannon entropy, cross-entropy loss, KL divergence and related measures of how well a classifier's predicted distribution fits the true one.
Softmax & Temperature
Watch softmax turn a model's raw outputs into probabilities, then turn the temperature up and down to sharpen or flatten the distribution.
Precision, Recall & Overfitting
Measure a model with precision, recall and accuracy, spot overfitting as training and validation diverge, and fight it with regularization.
Deep network architectures
Convolution Kernels
Apply different kernels — edge detectors, blurs, sharpeners — and see the feature maps a CNN layer produces.
MNIST Digit Classifier
Draw a digit and let a CNN trained on the MNIST dataset classify it, with confidence for each class.
RNNs & LSTM
See how recurrent networks carry state from one time step to the next, and how LSTM gates decide what to remember and forget when modeling time series data.
Autoencoders
Squeeze inputs through an encoder's bottleneck into a compact latent representation, then watch the decoder rebuild them from it.
Language models & agents
Attention & Transformers
See how attention lets each token weigh every other token, and how stacked attention layers form the transformer architecture behind modern language models.
Transformer Attention Lab
Teach attention to solve puzzles by looking across the full context, and watch the attention patterns that emerge to see the key ideas behind the transformer mechanism.
Embedding Space
The math behind embeddings: vectors, similarity and nearest neighbors, and how they power semantic search and retrieval-augmented generation (RAG).
Agent Patterns
A tour of agentic patterns — how they're structured, how they differ, and which use cases each one suits.
Jev by TypeSafe
An interactive look at how Jev, by TypeSafe, solves classification problems more efficiently than large language models.
Learning from rewards
Reinforcement Learning
Explore the basics of reinforcement learning — agents, actions, rewards and discount factors — and see how they shape what an agent learns to value.
Thin Ice (Frozen Lake)
Train an agent to cross the Frozen Lake grid to the goal without falling through the holes, and watch its value estimates and policy improve episode by episode.
Balancing Act (Cart Pole)
Teach an agent to keep a pole balanced on a moving cart by nudging it left or right, and see how rewards for staying upright shape what it learns.