A language model is built in two stages, and they do different jobs.

Pre-training is the expensive one. Predict the next word, across an enormous pile of text somebody chose and filtered, for weeks or months. Make it bigger and the predictions improve along a curve smooth enough to forecast, which is a strange thing to be true and still not fully explained. Particular skills are less orderly. Some appear to switch on at a certain size, and whether that is a real threshold or an artefact of pass-or-fail scoring is still argued about.

What that stage produces is a base model. It knows a great deal and will not answer you, because answering was never what the loop rewarded.

Post-training is the cheap one, and it is where the thing you talk to comes from. Show it written examples of good answers. Then let people compare its attempts and steer it toward what they picked. A model treated this way with 1.3 billion parameters was preferred to a base model with 175 billion.

Both stages leave their fingerprints. Fluency without a truth check means an invented fact arrives in the same confident voice as a real one. Training on approval means a tilt toward telling you what you hoped to hear, which is least useful exactly when you most want an honest answer.

And the cost shape means very few organisations do the first stage at all, so a small number of rooms decide what these models know, refuse, and cost.

Which leaves one question underneath all of it. Moment to moment, as a model writes a reply, what is actually happening?