It’s a prediction machine
At heart, generative AI does one simple thing astonishingly well: it guesses the next word.
Give it “The cat sat on the…” and it ranks every word it knows by how likely it is to come next. The brightest guess wins.
An interactive explainer
You type a question. It answers — fluently, like it understands. It doesn’t, not the way you do. Scroll to see what’s actually happening behind the cursor.
Every word, layout, and animation on this page was generated by an AI model. A human contributed only the prompt below, [a human: and a little tweak on the teaser] — no hand-written copy, no hand-coded markup.
SCROLL TO BEGIN ↓
At heart, generative AI does one simple thing astonishingly well: it guesses the next word.
Give it “The cat sat on the…” and it ranks every word it knows by how likely it is to come next. The brightest guess wins.
It doesn’t plan the sentence. There’s no outline, no idea of where it’s going.
It picks a word, adds it, then re-reads the whole thing to pick the next. A fluent paragraph is just this loop, running very fast.
Before you ever typed a word, the model was shown a staggering amount of human writing — books, websites, code.
It wasn’t memorizing. Across billions of examples it learned which words tend to follow which, and packed that into millions of tunable dials.
It has no database to look things up in. It learned the shape of language — how ideas connect, how sentences flow.
That’s why it sounds so human, and why it can write about combinations it was never explicitly taught.
The model is rewarded for sounding right, not for being right.
So when it doesn’t know, it doesn’t stop — it generates the most plausible-looking answer. Confident, fluent, and sometimes completely invented. (People call this “hallucination.”)
Swap words for pixels, audio samples, or lines of code and the recipe holds:
learn from millions of examples, then generate new ones that fit the patterns. One idea, many forms.
It’s a fast first-draft partner. Lean on it for the things it’s built for — and double-check the rest.
It’s autocomplete that read the internet.
Powerful, fluent, and confidently wrong sometimes. Now you know what’s behind the cursor.
This is a deliberate simplification. Real models break text into “tokens” (word-pieces), and image generators often work by removing noise rather than predicting a next pixel — but “learn patterns from huge data, then generate what fits” captures the shared idea.