Latent

Curriculum atlas

Every lesson and concept on one map: 95 lessons and 368 concepts across 11 tracks. Trails follow the order lessons build on each other, and the lessons you complete become explored terrain.

Show the atlas as

Complete a lesson and it shows here as explored terrain.

AI from zero 13 lessons1 What is AI, really?2 Turning things into numbers3 Getting less wrong4 The artificial neuron5 Networks of neurons6 How a chatbot writes7 What AI gets wrong8 AI that takes actions9 Where your AI runs10 Safety in AI products11 Open and closed models12 How AI sees, hears, and makes pictures13 When someone attacks an AI appHow machines learn 11 lessons1 The learning loop2 Linear regression and loss3 Gradient descent4 Better optimizers5 Probability and Bayes' rule6 Classification and probability7 Overfitting and generalization8 Decision trees9 Ensembles: forests and boosting10 Measuring a classifier11 Finding groups: k-meansNeural networks 7 lessons1 Layers are matrix multiplications2 Why nonlinearity matters3 Backpropagation4 Training a network5 Convolutions6 Reading handwriting7 Autoencoders and latent spacesLanguage as vectors 8 lessons1 Tokens2 Vectors and similarity3 Word embeddings4 Seeing high dimensions5 Search by meaning6 Retrieval-augmented generation7 Reading an embedding8 How your feed decides: recommendation systemsTransformers and LLMs 13 lessons1 Predicting the next token2 Before transformers: recurrent networks3 Attention, step by step4 Many heads and the causal mask5 Where words are6 The transformer block7 Inside a real language model8 How LLMs are trained9 Where training data comes from10 Fine-tuning and LoRA11 Generating fast: the KV cache12 Quantization: smaller numbers13 How a chat assistant worksLearning by trial and error 5 lessons1 Explore or exploit: bandits2 States, actions, and rewards3 Learning from experience: Q-learning4 Policy gradients5 Rewarding LLMs, and reward hackingInside the model 6 lessons1 Reading a model layer by layer2 How circuits form: induction heads3 Superposition: more features than neurons4 Switching parts off and steering5 Alignment: getting the goal right6 How alignment is attemptedModels that see 9 lessons1 How computers see2 Recognizing objects with your camera3 Images and words together4 Depth from a single image5 Where the classifier looks6 Finding anything you can name7 Cutting objects out8 How a model reads a picture9 Finding a body's poseModels that hear and speak 5 lessons1 How computers hear2 What sound is this?3 Speech to text4 Text to speech5 A voice assistant taken apartFrontiers 12 lessons1 Diffusion: from noise to data2 Diffusion on images3 Pictures, one token at a time4 One space for images and text5 Mixture of experts6 Thinking longer7 Models that use tools8 Evaluation: what a score means9 Inside an agent harness10 Compute and energy11 Long context in practice12 Where multimodal models are headingSecuring AI 6 lessons1 Attacks on AI systems2 Prompt injection3 Poisoned retrieval and memory4 Agents with too much power5 Poisoned data and the supply chain6 Adversarial examples

Drag to pan, and pinch or use ctrl + scroll (cmd + scroll on a Mac) to zoom. On a touch screen, move and zoom the map with two fingers. Tap a place for details. With a keyboard, Tab moves between lessons, arrow keys pan, plus and minus zoom, 0 shows the whole map, Space shows details, and Enter opens the lesson.

How to read this map

Each region is a track. Tracks sit left to right in the order they build on each other: a track appears to the right of every track it assumes. Trails inside a region follow the lesson order; dashed crossings run from the last lesson of one track to the first lesson of a track that builds on it.

Concepts are the glossary terms, placed beside the lesson that teaches them. Contour lines rise where concepts crowd together, so a tall hill is a lesson that introduces many new terms. Distance on the map shows grouping, not how similar two ideas are.

Your progress lives in this browser only. Mark a lesson complete at the end of its page and it lights up here, along with the concepts it teaches. See all of it on your progress page.

Try "embedding", "softmax", "overfitting", or "backpropagation".