Interactive atlas
See how AI models work inside
Short lessons with live demos. Type your own text, move the sliders and watch tokens, vectors and attention change in front of you. No math background needed, nothing to install.
Start with tokenizationBasics
Tokenization
Watch byte pair encoding learn a vocabulary in your browser. Count pairs, merge them step by step and see why some texts cost more tokens than others.
Open the lessonEmbeddings
Words as vectors you can see. Explore a map of meanings, measure similarity by angle, do king minus man plus woman and run a tiny semantic search.
Open the lessonNeural networks
A neuron is a weighted sum and a bend. Build a small network and watch it draw a decision line.
Open the lessonTraining
How a model learns from its mistakes. Loss, gradients and learning rate, live.
Open the lessonLanguage models
Next-token prediction
A language model only guesses the next token. See the probabilities and how temperature, top-k and top-p change the choice.
Open the lessonAttention
How every token looks at the others. Queries, keys, values and softmax, computed in front of you.
Open the lessonTransformer
Follow one token through a transformer block. Attention, feed-forward layer, residuals and normalization.
Open the lessonContext window
Why a model forgets the start of a long chat and what fits into its context.
Open the lessonMixture of experts
Huge models that use only a part of themselves for each token. Watch the router choose experts.
Open the lessonImages
Convolution
Slide a small filter over a picture and find edges, corners and textures.
Open the lessonText and images
How a model learns that a photo of a cat and the words a cat are the same thing.
Open the lessonDiffusion
Image models start from pure noise and remove it step by step. Watch it happen.
Open the lessonBuilding with models
RAG
Retrieval augmented generation. Split documents, embed them, find the right pieces and hand them to the model.
Open the lessonFine-tuning and LoRA
Change a model with a little training. Full fine-tuning against small LoRA adapters.
Open the lessonQuantization
Store weights in 8 or 4 bits instead of 16 and see what is lost.
Open the lessonAgents and tools
A model that plans, calls tools, reads the results and tries again.
Open the lesson