The Third Umpire · Lesson 1 — The Nets ← Course

The nets — predicting the next word.

The model only ever guesses the next word — so why does that add up to something intelligent?

Everything so far has been the umpire doing one small thing, over and over: read the words up to now, place odds on what comes next, and pick one. (Each “word” here is really a token — the chunk from the last page; we'll just say “word”.) It sounds far too simple to produce a thoughtful answer. So here is the real puzzle: how does guessing one word at a time add up to something that seems to understand?

The answer: to guess the next word well, you have to know a surprising amount. Finish “the capital of France is ___” and you need a fact. “She tripped on the step and ___” needs how the world works; “2 + 2 = ___” needs arithmetic; “he was furious, so he ___” needs a sense of how people behave. Every blank quietly demands real knowledge.

A young cricketer in the nets doesn't memorise every delivery they face — by facing thousands, they slowly soak up the patterns of the game until the right response becomes instinct. A model learns the same way. In a phase called training, it reads an enormous amount of human writing and does nothing but predict the next word, nudging itself whenever it's wrong. And nobody grades it — the text is its own answer key, the real next word always sitting there to check against.

Do that across billions of sentences and memorising is hopeless — you're forced to absorb the regularities underneath: grammar, facts, cause and effect, the shape of an argument. That hard-won sense of how things fit together is where the apparent intelligence comes from.

Watch a few deliveries — each one is a “guess the next word”. See what the umpire must already know to play it.

Each ball is a “guess the next word”. Watch the call, then see what it took.
Ball 1 of 5
what predicting well demanded
· a fact about the world
· cause and effect
· arithmetic
· how people behave
· how words relate

One move, learned from reading.

The model only ever predicts the next token. Nobody wrote these answers in by hand — it picked them up from text, the way a batter learns the game by facing it.

model.predict_next("The capital of France is")
# → "Paris"   (one token)

# then it adds that token and calls predict_next again — word after word.
# never hard-coded — learned from billions of next-word guesses.

Predicting well is learning how it works.

Five balls, five different kinds of knowing — a fact, then cause and effect, then arithmetic, then how people act, then how words relate. None of it was programmed in. The only way to get good at “guess the next word”, across everything people have written, is to take on how language, the world, and reasoning actually behave.

That's where the intelligence comes from. Like a cricketer in the nets, the model never memorised every ball; by facing enough of them it learned the patterns underneath — and those patterns are what make each small bet land somewhere sensible.

Go deeper — is it really “understanding”? optional

Predicting well is a kind of compression

There's far too much text to memorise, so a model that predicts well can't just be storing answers — it has to find compact patterns that generate the text: the rules of grammar, the regularities of the world, the shapes of common arguments. Learning to predict the text becomes, in effect, learning a working model of whatever produced it. That's why prediction and understanding start to blur together.

Skills appear as the practice scales up

With more text to learn from and a larger model to learn it, these abilities sharpen — and some capabilities show up that nobody trained for directly, simply because they help predict text better. More practice in the nets, more of the game absorbed.

The honest debate

Whether this counts as genuine “understanding” or an extraordinarily good imitation of it is unsettled, and reasonable experts disagree. It's also worth knowing that fluent language nudges us to read in more intent than may be there. Hold both ideas at once: the capability is real and often genuinely useful, and we should be careful not to over-read it. That gap is exactly where a later lesson on hallucination lives — a confident, fluent sentence is not a guarantee of a true one.