What Is a Word-Guessing Machine?
The single core idea behind every large language model: it isn't thinking in the way a person does, it's guessing the next word, extremely well.
A game you already know how to play
Try finishing this sentence: "The sky is..." You probably thought "blue" before you even got to the end of reading it. Nobody taught you a rule for that. You've just seen the word "sky" followed by "blue" so many times that your brain filled in the gap automatically.
That's most of what's happening inside a large language model, the kind of AI behind a chat assistant. At its core, it's a machine that has read an enormous amount of text and learned, extremely precisely, what word (or piece of a word) tends to come next after any given stretch of words. Give it the start of something, and it guesses the next piece. Then it adds that guess to what it already had, and guesses the next piece after that. One piece at a time, over and over, until it has produced a whole answer.
It isn't "thinking" the way you do
This is the part that trips people up, because the output can feel like it's reasoning things through. It isn't reasoning the way a person sits and thinks about a problem. It's predicting, one small step at a time, what text would plausibly come next, based on patterns it picked up from reading a huge amount of human writing. It's less like a person having a thought and more like an incredibly well-practiced guesser who has seen almost every kind of sentence there is and has a very good instinct for what usually follows.
That's not a small trick. Guess well enough, over and over, at the scale of a whole paragraph, and the result can look a lot like understanding, even though the mechanism underneath is just repeated, very good guessing.
The pieces it guesses with: tokens
It doesn't guess whole words one at a time, exactly. It works with pieces called tokens, which are sometimes a whole word ("cat") and sometimes a chunk of one ("un" + "happy" + "ness"). Splitting text into tokens lets the same machine handle rare words, made-up words, and even other languages, by breaking anything down into pieces it has seen before, without needing a fixed list of every possible word in advance.
Cover the end of this sentence with your hand and guess what word comes next: "Once upon a..." Now try a harder one: "The capital of France is..." Notice how confident you felt about both. That confidence, built from having seen the pattern before, is the same thing a language model is doing, just at a scale of essentially everything ever written that it was trained on.