How a language model gets made
Five stages from a pile of text to something you can talk to, and where the safety work actually attaches. Click through the stages.
Stages one to four change the model itself. Stage five wraps around it without touching it. Almost every confusing claim about what a model "knows" or "learns" comes from mixing those two up.
0. one function
A language model is a function that takes a sequence of word fragments and returns a probability distribution over what comes next.
That it stores facts in a database it looks things up in.
1. pretraining
The model is fitted to an enormous pile of text by repeatedly predicting the next fragment and adjusting when it is wrong.
That the model is being taught facts. It is being fitted to a distribution.
2. supervised fine-tuning
The base model is shown many examples of a request followed by a good response, and its weights shift toward producing that shape.
That this stage adds new knowledge. Mostly it selects for a format and a manner already latent in the base model.
3. preference training
People, or a model following written principles, compare pairs of responses. Those comparisons train a second model that scores responses, and that scorer is used to update the first model.
That this teaches the model human values. It optimises against a learned proxy for the judgement of one particular pool of labellers under one particular set of instructions.
4. evaluation and the gate
Before release, the model is deliberately attacked, measured against dangerous capability thresholds, and checked against a published framework that is supposed to decide whether it ships.
That an evaluation certifies a model is safe. The UK safety institute explicitly rejects that framing.
5. deployment safeguards
System prompts, filters, classifiers, monitoring and usage policies sit around the finished model while it runs.
That the model learns from your conversations as you have them. Deployed weights are static.
What is honestly not public
The pipeline above is a teaching sequence and it is broadly right. But no frontier lab has published a full end to end recipe for a current production model. What exists publicly is older method papers, one genuinely detailed technical report for an open weight model, and model cards that describe evaluation far more thoroughly than they describe training.
So the accurate picture is the documented pipeline plus the fog over the current instances of it. Anyone who draws this as a clean, fully known assembly line is drawing something nobody outside those labs can actually see.
Two more corrections worth carrying: real pipelines loop rather than run once, and the published figures people quote for model size and training compute are almost always taken from open weight models or regulatory thresholds, not from the closed systems being discussed.
Sources
Everything above was checked against these during the research pass. Where a claim is contested, the sources that disagree are both listed.