How language models work, without the maths

You do not need to understand transformers to use AI well, in the same way you do not need to understand combustion to drive. But you do need to know why the engine stalls on hills. This module covers the four pieces of machinery that explain almost every odd behaviour you will encounter.

1. It reads in tokens, not words

Text is chopped into tokens: roughly word-fragments. "Teaching" might be one token; an unusual name might be four. The model only ever sees these fragments and their statistical relationships. It never sees letters the way you do.

This immediately explains a family of embarrassing failures. Ask how many times the letter 'r' appears in a word and it may well get it wrong, not because it is stupid but because it is not looking at letters at all. It is like asking someone to count brushstrokes in a photograph of a painting.

Practical consequence: never trust it for character-level work — spelling puzzles, counting letters, precise text manipulation by position. Do trust it for meaning-level work, which is nearly everything else you want.

2. It was trained once, and then frozen

A model learns from a vast snapshot of text taken up to a cutoff date, and then stops learning. It does not remember your last conversation. It does not learn from your corrections. Every conversation starts from the same frozen state.

Two things follow. First, its knowledge of recent events is unreliable unless the tool has web search attached, and you should check whether yours does. Second, when you painstakingly explain your writing style, that effort evaporates when you open a new chat. If something matters, save it somewhere you can paste it back in. Many tools now offer a memory or custom-instructions feature that does this for you; it is worth ten minutes to set up.

3. The context window is its whole world

Everything the model can consider at once — your instructions, the document you pasted, the entire conversation so far, and its own reply — lives in a fixed-size space called the context window. Nothing outside it exists.

Modern windows are large, often hundreds of pages worth. But they are finite, and long conversations degrade as early material falls out or gets crowded. This is why a chat that was going brilliantly can drift into vagueness after an hour.

4. It picks; it does not look up

At every step the model holds a ranked list of tokens that could plausibly come next, and it chooses one of them. The choice involves deliberate randomness. Run the same prompt twice and you will often get two different answers, not because the tool is unreliable, but because picking is what it does.

The dial that controls this is called temperature. Turned down, the model almost always takes the highest-ranked token and answers predictably. Turned up, it ranges further down the list and becomes more inventive and less careful. Developer tools expose temperature directly. Most chat products hide it and choose a middle setting for you.

Three consequences are worth carrying around. Any single answer is one sample, not the model's considered position. The regenerate button is therefore a free second opinion, and two runs that disagree tell you the ground is soft and you should verify. Most importantly, a prompt that produced something excellent once has not been proven; it may simply have been lucky.

Why it sometimes gets better if you just ask

Telling a model to "think step by step" or "work through this carefully before answering" often genuinely improves the answer. It sounds like superstition. It is not. The model produces text one token at a time, and each token is influenced by everything before it, so text that lays out intermediate steps gives the final answer better ground to stand on. Reasoning it writes down is reasoning it can use. Reasoning it skips never happened.

Newer 'reasoning' models do this internally by default, which is why they are slower and better on hard problems. Either way, the principle holds: room to work produces better work.