The bottom line
A large model (like the GPT behind ChatGPT) is essentially a machine extremely good at playing “word chain”: give it an opening, it predicts the most likely next word based on “experience,” then the next… until a full answer forms. Its “experience” comes from the massive human text it has read.
Understanding it through everyday life
Imagine a child who read a hundred thousand books. You give them an opener: “The weather is nice today, let’s go to the park,” and they’ll most likely continue with “for a walk” or “to play” — because that’s the most common pairing in the books they’ve read.
A large model does exactly the same — just at staggering scale:
- The books it read: nearly the entire public internet (Wikipedia, papers, books, code…).
- Its “word-chain” skill: not rote memorization, but learning the statistical patterns and deep structures between words.
- Parameters: hundreds of billions of tunable knobs that together decide how the “next word” is chosen.
Why it seems to “understand”
Play word-chain well enough and the illusion of “understanding” emerges — no, more precisely, genuine capability emerges: it can translate, write code, reason. Like a Go AI that only learned “which move wins” but ended up “knowing” Go.
But it has three essential limitations:
- It confidently makes things up (hallucination): it is never responsible for “facts,” only for “plausibility.”
- Its knowledge has a cutoff date: it only knows what it read, not what happened after.
- It doesn’t truly “think”: no goals, no intent — it just computes the most reasonable continuation given an opener.
In one line
A large model = a machine that read the entire web, pushed word-guessing to the extreme, and thereby gave rise to understanding and creativity. Used well, it’s productivity; but know this: it never guarantees what it says is true.