A field guide to language models

Model Mechanics

How a language model turns text into the next token, and how to work with one: what tokens cost, how much thinking to buy, when to split work across agents, and how agent loops and skills run.

The machine, mapped

Figure 1 shows the whole lifecycle of a model. The back deck builds the model once. The front deck runs it once for every token it writes. Hover over a station to read what it does, and open its explainer from there.

The machineThe back deck builds the model once. The front deck runs it once for each token it writes.Open the explainer

Back deck: build, onceFront deck: run, once per tokenType slugs: tokens, each with its IDStation sizes and counts are illustrative. The token IDs come from the site's teaching tokenizer, which has 247 vocabulary entries.

The route

Five pieces, in order

Start with tokens if you are new. Each piece stands on its own, and each ends with a short summary.

1 · Start · 10 minTokens are the meterSee how a model cuts text into tokens. Compare the cost of two messages, then lower a sample monthly bill.
2 · The machine · 18 minInside a language modelFollow one model from its training corpus to the next token: tokenizer, embeddings, transformer blocks, pre-training, and sampling.
3 · Choice · 6 minReasoning effortMove a thinking budget and find where more tokens stop improving the answer.
4 · Choice · 7 minOrchestration and subagentsCompare one crowded context with four scoped workers, then decide which jobs to split.
5 · Choice · 7 minLoops and skillsFollow an agent's tool loop as its context grows, and see why a skill loads only when a task needs it.