Trace one training step from a wrong guess to a better one.
During training the model reads "The cat sat on the" and guesses "table". The real next word was "mat". What happens next?
Yes.That nudge is the whole of learning. It is tiny — a single example barely moves anything — and it happens billions of times.
Not quite.Nothing is stored. The only trace an example leaves is the small change it made to the numbers.
Not quite.There are no rules anywhere in it. There are only weights, and a procedure for adjusting them.
Put one training step in order.
Tap them in order — first to last.
Hide the next word→Guess→Measure the error→Nudge the weights
The entire training procedure. Repeated more times than anyone can picture.Yes.Four steps, repeated for trillions of words. Nothing in the loop understands anything; the understanding, such as it is, is what the loop leaves behind.
Learning here means one thing: make the error smaller. Everything the model can do is a side effect of getting good at that.
Slide through training and watch the error fall.
0.1%5%40%100%
How wrong it is, on average9.8
A perfect guesser0
0.1%It has learned that some words are common. Output is word soup with plausible spacing.
How wrong it is, on average4.2
A perfect guesser0
5%Grammar mostly holds. Sentences run on and contradict each other.
How wrong it is, on average2.6
A perfect guesser0
40%Fluent, and confidently wrong about facts.
How wrong it is, on average2.1
A perfect guesser0
100%Better again — and the curve has flattened. Huge extra effort for small gains.
Fluency arrived long before accuracy. Why?
Yes.Sounding right is the cheapest thing to learn, so it is learned first. That mismatch — fluent early, accurate late — is the origin of most of the trouble with these systems.
Not quite.There are later stages, and they mostly shape behaviour and tone. The gap appears within this one.
Not quite.Most of it is broadly accurate. A single fact still appears a handful of times against grammar's every sentence.
Move the control to see what changes.
The word "the" appears in nearly every document. A particular small town's population appears in, say, four. If learning is driven by repetition, roughly how many more chances did the model get to learn the first?
times more
Yes.The exact number does not matter and cannot be pinned down — the point is the scale of the gap. Anything rare gets almost no pressure from the training, whatever its importance.
Which of these will the training loop learn WELL, and which barely at all?
How an English sentence is ordered.
The opening hours of one shop in one town.
The shape of a polite refusal.
A detail mentioned in exactly one obscure document.
How a recipe is usually laid out.
Yes.The loop rewards patterns that recur. Anything that appears once or twice leaves almost no mark — and there is no signal in the output telling you which kind of thing you just asked for.
After the main training, humans rate thousands of answers and the model is trained further on those ratings. What does that stage change?
Yes.This is where helpfulness, refusals and tone are shaped. It is powerful and it works with material already there; it cannot add a fact the first stage never met.
Not quite.A few can leak in from the examples, but the volume is tiny beside the first stage. It is a stage about behaviour.
Not quite.Nothing here is fixed individually. Ratings pull the whole distribution of answers in a direction.
Lesson complete
Training is one move repeated: guess, measure the error, nudge the weights.