Millions of people now use a large language model every day, and almost none of them could say what happens after they press Enter. That's fine for casual use — you don't need combustion theory to drive a car. But the moment you rely on AI for real work, the mental model matters. People who think they're talking to a digital encyclopedia get burned by confident errors. People who understand what the machine is actually doing write better prompts, catch mistakes earlier, and know exactly when to trust it. This is the tour, in plain English.
#One job: predict the next token
Strip away everything and a language model does exactly one thing: given some text, it predicts what token comes next. During training it read a colossal amount of text — books, articles, code, conversations — and adjusted billions of internal dials until its predictions got very, very good. That's it. There's no database of facts inside, no list of rules, no lookup table. Everything it 'knows' is compressed into those dials as statistical patterns: which words, ideas, and structures tend to follow which.
#What happens when you press Enter
Everything you typed, plus system instructions you never see.
The text is split into small chunks — roughly ¾ of a word each.
The model scores every possible next token and picks a strong one.
The new token is added to the text and the loop runs again.
First, your words are broken into tokens — chunks of roughly three-quarters of a word. The model never sees letters or words, only these numbered pieces. Then it does its single trick: it calculates, across its entire vocabulary, how likely each possible token is to come next, and selects one. That token is appended to the text, and the whole process repeats — one token at a time, each choice shaped by everything before it — until the answer is complete. The fluid paragraph you receive was assembled piece by piece by a prediction engine running in a loop.
#So why does it feel intelligent?
Because at sufficient scale, predicting text well requires modeling the things text is about. To continue a physics explanation, the patterns must encode physics. To finish a legal argument, they must capture how legal arguments work. Reasoning, style, structure, even humor — all of it is embedded in the statistics of human writing, and a big enough model absorbs it. The result genuinely is remarkable. It's just remarkable in a specific way: fluent pattern completion, not consultation of a mind or a database.
#The context window: the model's working memory
A model can only 'see' a fixed amount of text at once — the context window. Your conversation, your pasted documents, the instructions, and the answer being generated all have to fit inside it. Anything that falls outside might as well not exist. This is why a long chat can seem to 'forget' what you said an hour ago, and why the size of the window matters so much for real work.
#What this means for how you use it
- It can't know your business unless you tell it — your policies, prices, and context have to be in the prompt (or retrieved into it) every time.
- Confidence is not evidence — right and wrong answers are generated by the same machinery, so verify anything that matters.
- Better input, better output — the model completes the pattern you start, and a vague pattern completes vaguely.
- It doesn't learn from your conversation — close the chat and it's gone, unless the product explicitly builds a memory on top.
- It has no goals and takes no actions — until someone wires it into tools, it only ever produces text.
“The model isn't a mind and isn't a search engine. It's the world's most sophisticated pattern-completion machine — and once you treat it as exactly that, it stops surprising you and starts working for you.”
None of this diminishes what these systems can do — it locates it. A machine that completes patterns this well is a genuinely new kind of tool, capable of drafting, translating, summarizing, and reasoning through problems at extraordinary speed. Use it knowing what's under the hood — a prediction loop with a finite window and no concept of truth — and you'll get its strengths without inheriting its blind spots.
