Lesson 2 of 6
What "training" really is
11 min read
"Training" sounds like loading a program full of facts. It's stranger — and simpler — than that. At its heart, training is a machine turning dials until it guesses better.
A guess, then a nudge
Here's the whole idea in one move. The model reads a scrap of text — "grass is" — and guesses the next word. Inside it are a huge number of tiny dials that shape that guess. At first the dials are random, so the guess is nonsense. Then it peeks at the real next word in the text — "green" — and nudges the dials a hair so that next time, "green" scores a little higher. Guess, check against reality, nudge. That's the entire trick.
Learning to shoot a basketball is the same loop: shoot, see the miss, adjust. The model shoots at the next word and adjusts its dials.
Millions of these dials
The scene gives you three dials so you can feel it. A real model has billions — far too many for any person to set by hand. That's the point of "Practice": instead of a human tuning dials, the machine measures how wrong each guess was and nudges every dial a tiny bit in the direction that would have helped. One nudge barely matters. Trillions of nudges, across an ocean of text, add up to a model that predicts remarkably well.
Those dials have a name: parameters. When you hear that a model has "70 billion parameters," that's 70 billion dials that got tuned during training.
Training is automatic dial-tuning: guess, measure the miss, nudge every dial toward a better guess — repeated at an enormous scale.
Nothing gets stored as a fact
Notice what never happened: no fact was filed away. The model didn't save "grass is green" in a box somewhere. It only adjusted dials so that the pattern comes out right. All of its "knowledge" lives in the exact settings of those dials — a giant web of nudges, not a shelf of facts.
That's why you can't open a model and find a sentence inside it, or delete one fact cleanly. Everything is smeared across the dials at once. A trained model is the frozen result of all that practice — a single enormous setting that happens to predict language well.
A model's knowledge isn't a stored list. It's the settings of billions of dials, tuned so the patterns come out right.
What to carry forward
- —Training = guess the next word, check the real one, nudge the dials — over and over.
- —The dials are called parameters; real models have billions, tuned automatically.
- —Knowledge lives in the dial settings, not as stored facts you can look up or delete.
In plain terms, what is "training" a model?
Continue in the app
Take the whole How AI Learns course — tracked
Get your personalized path, progress and streaks in the app — this lesson and every next one, in order.