AI, Explained

How does a language model remember - and what does it cost? This site walks the whole story: text becomes vectors, vectors carry meaning, attention moves it around, the memory bill arrives, and a decade of architecture is the story of paying it.

How to read this site. The left sidebar is the full curriculum - six domains (the way the wider field organizes itself), each holding one or more tracks in teaching order. Every page ends with prev/next links that walk the whole path. Technique pages carry a verdict: what it fixed and what it still costs - each residual is the next page’s reason to exist.

The first three domains walk one architecture’s story: tokens become vectors, attention moves them, training fills the weights. Alignment & Safety is the fourth - how a raw predictor gets steered, and by whom: LoRA/QLoRA for cheap adaptation, DPO/ORPO/Constitutional AI/Self-Instruct for what to steer it toward. Inference & serving pays the bill after all of that. World models is a sixth domain because it is a different objective - predict the next state of the world, not the next token. The broader model ecosystem is still not a track here; What’s coming points at it.

Many pages carry live widgets - every number in them is computed from the picture you are looking at, never typed in.

Stuck, or wondering “why does this work that way?” - open the tutor chat button (bottom-right) and ask. It answers only from these pages.

Start with the map to see the whole territory, or jump straight to Tokens and follow the path.

Companion to the video series “How Models Learned to Forget”.