How a large language model actually works — from words, to numbers, to meaning, to action. Five interactive illustrations that open up the machinery, one idea at a time.
Text becomes tokens, then vectors — and self-attention lets context decide what a word means. The honest answer to "it's just autocorrect."
Where those vectors live: meaning as the angle between them, and the staggering vastness of high-dimensional space.
Hit enter and the model re-reads the whole conversation to reply — it stores nothing on its own. Meet inference, the context window, and memory.
A fabricated citation, built exactly like a real one. Why “don't hallucinate” doesn't help — and how real sources (RAG) fix it.
How tools give a language model hands and legs — letting it take actions and do real work, not just talk.