Nobody wrote
the rules.
A model’s knowledge, manner and quirks all come from how it was raised. Seven stages.
Learned, not programmed.
Nobody sat down and typed in the law of negligence. There is no file inside the model that lists the elements of a claim, no rule that says how to draft a chronology. The model learned everything the same way: by playing one game — guess the next word — across a vast amount of text.
Everything else follows from that. The knowledge, the manner, the gaps and the quirks are all consequences of how the model was trained, not decisions anyone made line by line.
Rules written by people
A programmer decides what happens in every case, and the machine does exactly that.
Patterns absorbed from text
Nobody writes the behaviour. It emerges from what the model read and how it was corrected.
Pretraining: the reading years.
Stage one is scale. The model reads a colossal corpus — books, articles, code, the public web — predicting the next word billions of times and getting marginally better each time. This is where the knowledge comes from, along with the grammar, the style, and the reasoning patterns.
It is also where the cutoff comes from. The model’s world stops on the day the corpus was assembled. Anything decided, published or amended after that date is simply not in there.
A judgment handed down the week after the cutoff is invisible — not obscure, not half-remembered. Absent.
What pretraining produces.
The raw pretrained model — the base model — is not an assistant. It has never been told to be helpful. Ask it a question and it may simply continue with more questions, because on the internet, questions are often followed by other questions. It completes text; it does not serve you.
It has the knowledge but not the manners. It is autocomplete with a doctorate. Turning that into something you can brief takes two more stages of upbringing — and you can watch them happen below.
One prompt, three upbringings.
What should I do if I’ve missed a filing deadline? What should I do if my solicitor missed a deadline? 17 answers · Legal forum · Sorted by newest —
Taught to please.
Stage three is human feedback. Humans compare pairs of answers and pick the better one, over and over; the model is tuned towards the answers people preferred. This is where the helpfulness comes from — the structure, the clear headings, the appropriate caution.
It also has a side effect: agreeableness. A model rewarded for answers people like will lean towards telling you your argument is strong.
Praise from a model is weak evidence. It was trained on what people preferred to hear, and people prefer to hear that their skeleton argument is persuasive.
The fix is to make honesty the thing it is asked for. Adversarial instruction works with the training, not against it:
The quirks, explained.
| Quirk | Where it comes from | |
|---|---|---|
| Hedges and caveats | → | Human raters preferred caution |
| Agrees with you too readily | → | Preferred answers were agreeable ones |
| Confident tone even when wrong | → | Fluent confidence reads well; raters can't check every fact |
| Knowledge stops at a date | → | The corpus was frozen at the cutoff |
| Better at common problems than rare ones | → | Prediction is strongest where examples were plentiful |
Raised, not built.
A model’s behaviour is the residue of its upbringing: the reading (pretraining), the schooling (instruction tuning), the finishing (human feedback). Nothing about how it acts was written down as a rule; all of it was absorbed.
That is why prompting works the way it does — you are steering patterns, not invoking commands. And it is why hallucination persists: a machine trained to produce plausible text will produce plausible text even when the facts have run out.
It is also why every generation of models behaves a little differently. Nobody rewrote the rules — there were never any rules to rewrite. The upbringing changed.
That’s the whole story.
From next-word prediction to a colleague you can brief: you’ve now seen every layer of the machine. Explore the full series, or put it to work.
Browse the full Explainer series →