How the model reads

Workshop · 1 of 7 · One mental model that explains almost everything

Every conversation with an AI happens inside a context window — the model's working desk. Everything you type, every document you paste, every reply it gives you: all of it lands on that desk, and the desk is finite. Claude's holds about 200,000 tokens — roughly 500 pages. That sounds enormous. It behaves smaller than you think, and knowing why is the single highest-leverage thing you can learn about AI.

Tokens, not words

Text is chopped into small pieces — a token is a word, part of a word, a punctuation mark. “The architect walked into the meeting” is seven-ish tokens. Everything the model does, it does over tokens: your 40-page PDF is not “a document” to the model; it is thirty thousand tokens sitting on the desk.

Meaning lives in a space

Under the hood, each token becomes a point in a vast meaning-space, where “architect” sits near “engineer” and far from “banana.” The model's skill is navigating that space: given everything on the desk so far, it predicts what belongs next, one token at a time. That is the whole trick. There is no database of facts being looked up, no little person reasoning inside — there is an extraordinarily good prediction of what comes next, shaped entirely by what is on the desk.

The desk is everything

The model cannot use what isn't in the window, and it cannot ignore what is. Put the wrong things there — outdated notes, three contradictory drafts, an angry aside — and they shape every prediction that follows. Put the right things there, in the right form, and the same model becomes dramatically better. People say “the AI is bad at my task” when the honest sentence is usually “my desk is a mess.”

So arrange the desk on purpose: load what matters, leave out what doesn't, and put your actual question last — the window remembers its edges best. Next: what happens when the model gets hands.