Prediction, Not Understanding
The One Sentence That Explains Most of It
A large language model is a system that predicts what text should come next. That is the whole trick. You give it some words, it calculates which word is most likely to follow, adds that word, then does the same thing again with the slightly longer text. Repeat a few hundred times and you get an essay. It feels like conversation because human writing is full of patterns, and the model has absorbed an enormous number of them. But there is no moment where the model steps back and thinks "do I actually know this?" It is producing plausible continuations, and plausible is not the same as true.
- The model works one small chunk of text at a time, left to right, never planning the whole answer in advance the way you would
- "Most likely next word" is calculated from patterns in text, not from checking reality
- Fluent and correct are two separate properties — the model is optimised hard for the first one
- This single fact explains almost every strange thing AI does
Does It Understand Anything?
This is a genuinely contested question, and anyone who gives you a confident one-word answer is oversimplifying. What is not contested: the model has no body, no memory of yesterday unless the app gives it one, no goals of its own, and no way to check a claim against the world. What is also true: to predict text well across millions of topics, it has clearly built internal representations that behave a lot like concepts. It can apply an idea to a situation it has never seen. Somewhere between "just autocomplete" and "it thinks like you" is the honest answer, and researchers are still arguing about where.
- It has no senses, no continuous existence, and no stake in whether it is right
- It can still generalise — apply a pattern to a new case — which pure lookup could never do
- Useful working stance: treat it as an extremely well-read pattern machine, not a mind and not a database
- Be suspicious of both hype ("it is basically conscious") and dismissal ("it is just autocomplete")
Why This Matters for You Right Now
If you think of AI as a search engine, you will trust it in exactly the wrong places. Search returns documents that exist; a model generates text that did not exist until you asked. If you think of it as a person, you will assume it remembers your last conversation, cares about your grade, or would tell you when it is unsure. It does none of those by default. Getting the mental model right changes your behaviour: you start checking the things worth checking, and you stop wasting energy checking the things it is genuinely reliable at, like rephrasing a paragraph you wrote yourself.
- Not a search engine: it writes an answer rather than retrieving one, unless the tool explicitly searches
- Not a person: no memory across chats by default, no stake in your outcome
- Very reliable at transforming text you supply — summarising, restructuring, translating tone
- Much less reliable at facts it has to produce from nowhere, especially specific ones
Watch the model choose the next word, one at a time.
Prefer slides, quizzes, and saved progress? Read this lesson in the library — free, no sign-up.