Learn how LLMs actually work, by playing with them.
Six guided modules with browser-native simulators. Tokenize text, build vocabularies, train a tiny model, retrieve documents and wire agents together. No GPU. No setup.
/ curriculum
The full stack of an LLM
Follow them in order or jump to what you're curious about.
Tokens
See how text becomes numbers with a live BPE-style tokenizer.
Vocabulary
Build a vocab from a corpus and watch coverage grow.
Pre-training
Train a tiny next-token predictor in your browser.
Fine-tuning
Watch loss curves move as you teach a model new behavior.
Retrieval (RAG)
Index documents, embed queries, and rank chunks live.
Orchestration
Compose agents, tools, and routers into a flow.
/ key ideas
A model in one paragraph
A language model takes text, breaks it into tokens from a fixed vocabulary, and predicts the next token. It learns this skill during pre-training on huge corpora, then is steered with fine-tuning. To answer about things it never saw, we attach retrieval. To make it act, we wrap it in orchestration, with tools, routers, and memory.