Volume I · LLMs

How a language model thinks — and how engineers wire it.

A language model is a next-token engine. Everything useful around it — tools, memory, loops, graphs, evals — is engineering. The KV cache is why the next token is cheap.

Start with tokens Skip to the harness

In English

A language model only does one job: guess the next piece of text. Agents, tools, memory, and loops are software wrapped around that guess. This volume walks the wrap, from the first token to the product.

Try this: Click Tokens if you have never seen a tokenizer. Skip to The harness if you already know attention and want the engineering.
Keep this: The model predicts tokens. The harness is the product.

The stack · click a layer

What this is

Interactive diagrams, not a blog. Tokenize a sentence, sample at different temperatures, watch attention move, reuse a KV cache, overflow a context window, step a tool call, run a loop, take a branch in a graph, and explode the harness around the model.

What this is not

Not a live model. Distributions and attention maps are teaching fixtures with the right shape. The claims about architecture are the real ones: next-token prediction, the loop, and the harness as the product.