TL;DR: A large language model does exactly one thing: given some text, it predicts the next chunk of text, over and over. Everything else, the fluent answers, the code, the eerily good advice, and the confident nonsense, falls out of that single mechanism running at enormous scale. This explainer builds the mental model with no math: what a token is, what training actually changes, why the model has no database to look things up in, and why “it made that up” is not a bug but the same process that produces its best work, pointed at a question it cannot answer. Understand the mechanism and you stop being surprised by the model and start directing it. Learn like a student who wants to know how the engine works, use it like a CEO who knows what the tool can and cannot be trusted to do.
Hundreds of millions of people now type into an AI model every day and could not tell you, even roughly, what happens after they hit enter. That is not a failure of intelligence. Nobody explained it. The result is a workforce that treats a next-word predictor like a search engine, an oracle, or a person, and then feels betrayed when it behaves like none of those things.
Understanding the mechanism is quickly becoming a core literacy, not a specialist skill. As more work runs through these tools, the people who get the most out of them are not the ones who know the cleverest prompts. They are the ones with an accurate mental model of what is happening under the hood, because that model tells them what to trust, what to check, and how to ask. We made the broader case in Prompt Engineering Is Not Enough. This piece builds the foundation that sits underneath it.
The one thing an LLM does
Here is the whole engine in one sentence: a large language model takes the text you give it and repeatedly predicts the most plausible next piece of text.
That is not a simplification of what it does. It is what it does. You type a question; the model reads it and predicts a likely first word of an answer. Then it reads your question plus that first word and predicts the second. Then the third. It writes the way you might finish someone else’s sentence, except it never stops finishing, one small step at a time, until the answer is complete.
The unit it predicts is not exactly a word. It is a token, a chunk of text that is often a whole word but can be part of one, a piece of punctuation, or a common word-fragment. “Understanding” might be one token or split into “Understand” and “ing.” This is why models sometimes miscount the letters in a word or stumble on wordplay: they do not see letters the way you do. They see tokens.
Everything impressive about these systems is this one loop, running at a scale that is hard to picture. And everything frustrating about them is the same loop, running past the edge of what it reliably knows.
What “training” actually changed
If the model just predicts the next token, where does its apparent knowledge come from? From training, and this is the part worth getting right.
A model is, mechanically, a vast set of numbers called parameters, often billions of them. Before training, those numbers are essentially random, and the model’s predictions are gibberish. Training is the process of showing the model an enormous amount of text and, every time it predicts a next token, nudging those billions of numbers slightly so that next time its prediction is a little closer to what actually came next in the real text.
Do that across a huge slice of the written internet, books, and code, trillions of times, and something remarkable happens. To get good at predicting the next token in any text, the model is forced to encode patterns: grammar, facts, styles, reasoning steps, the structure of a legal clause, the rhythm of a poem. Nobody programs these in. They are the statistical residue of prediction done at scale. The architecture that made this practical is called a transformer, introduced in the 2017 research paper “Attention Is All You Need,” whose key idea, attention, lets the model weigh which earlier words matter most for predicting the next one.
Crucially, when training ends, the numbers freeze. The model does not keep learning from your conversations, and it has no memory of yesterday unless the system around it feeds that back in. What it knows is baked into those frozen parameters, which is why every model has a knowledge cutoff and why it cannot tell you what happened this morning.
The scale of this process has been climbing at a pace worth seeing in one place.
Table 1. The trajectory of the technology (verified data)
| Measure | Finding | Source |
|---|---|---|
| Training compute | Doubling roughly every 5 months | Stanford HAI AI Index 2025 |
| Training dataset size | Doubling roughly every 8 months | Stanford HAI AI Index 2025 |
| Cost to query a GPT-3.5-level model | Fell from $20 to $0.07 per million tokens (Nov 2022 to Oct 2024) | Stanford HAI AI Index 2025 |
| Share of notable models from industry, 2024 | ~90% (up from ~60% in 2023) | Stanford HAI AI Index 2025 |
The middle row is the one that quietly changed the world: a more than 280-fold drop in the cost of a capable model’s output in under two years is why the technology went from a demo to a utility.
Why it does not “look things up”
Now the single most useful thing to understand, because it explains the behavior that trips everyone up.
An LLM has no database inside it. When you ask a factual question, it is not searching records and returning a match. It is doing the only thing it does: predicting the most plausible continuation of your question, based on the patterns baked into its parameters. When the answer is common in its training text, prediction and truth line up beautifully, and it appears to “know” the fact. When the answer is rare, ambiguous, or absent, the model does not stop and say “no result.” It cannot. It predicts a plausible-sounding continuation anyway, because that is its entire job.
That is what a hallucination is. It is not the model malfunctioning. It is the model working exactly as designed, generating fluent, confident text, pointed at a gap where the confident text happens to be wrong. The fluency is identical whether the content is true or invented, which is precisely why you cannot tell from tone. A useful mental model: the LLM is not a librarian retrieving a book. It is an extraordinarily well-read improviser who will always give you a smooth answer and has no built-in signal for the difference between remembering and making it up.
This also explains two more everyday behaviors. Sampling is why you get a different answer to the same prompt twice: the model does not always pick the single most likely next token; it samples from among the likely ones, which is what makes it feel creative rather than robotic. And the context window is the amount of text, your prompt plus the conversation so far, that the model can attend to at once. Go past it and the earliest parts simply fall out of view, which is why very long chats seem to “forget” what you said at the start.
The working mental model
Put the pieces together and you get a model that is easy to carry and changes how you use the tool.
Table 2. A working mental model of an LLM (CEOtudent editorial framework)
| If you think of it as… | You will expect… | And be wrong, because… |
|---|---|---|
| A search engine | Retrieved, sourced facts | It generates plausible text, it does not retrieve records |
| A person who understands you | Memory, intent, accountability | It has no memory between sessions and no model of truth |
| A calculator for words | Deterministic, repeatable output | It samples, so the same prompt can give different answers |
| A well-read improviser | Fluent help that needs checking | This one is roughly right |
The improviser model earns its keep in three practical moves. First, give it the material. Because it predicts from what is in front of it, pasting the relevant document, data, or context into the prompt beats relying on its frozen memory, and it is why briefing an AI well is a skill in itself, one we broke down in How to Delegate to an AI Agent. Second, verify anything that matters. Fluency is not evidence. For any factual claim, name, number, or citation you will act on, the model is a first draft to check, not a source to trust. Third, use it where being plausible is enough. Drafting, brainstorming, restructuring, explaining, and translating are tasks where a strong plausible answer is genuinely valuable; high-stakes facts and irreversible decisions are not.
None of this requires knowing the math. It requires knowing the mechanism: one loop, predicting the next token, from patterns frozen at training time, with no lookup and no built-in sense of true. Hold that, and the model stops being magic or menace and becomes what it is, a powerful tool whose limits are as knowable as its strengths. That knowing is itself the durable skill, because the tools will keep changing and the mechanism will not, a point we develop in The Manager-of-AI Playbook.
FAQ
Is the model actually thinking or reasoning?
It produces text that often looks like reasoning, and “reasoning” models generate intermediate steps that genuinely improve their answers on hard problems. But mechanically it is still predicting tokens; the reasoning-shaped output is a very sophisticated version of the same process, not a separate faculty. Treat visible reasoning as helpful and still checkable, not as proof.
Does it learn from my conversations?
Not the core model, whose parameters are frozen after training. Some products feed your recent messages or saved preferences back into the prompt to simulate memory, but that is the surrounding system, not the model updating itself. Your chat does not retrain it.
Why is it so confident when it is wrong?
Because confidence and correctness are unrelated in its output. The model generates the most fluent continuation regardless of whether the underlying content is accurate, and fluent text reads as confident text. There is no internal “I am unsure” signal attached to a hallucinated fact, which is exactly why external verification is on you.
If it is just predicting text, why is it so good?
Because predicting the next token well, across essentially all of written human knowledge, turns out to require encoding a staggering amount of structure about language, facts, and reasoning. The simplicity of the task and the richness of the result is the whole surprise of the field. Scale did the rest, as Table 1 shows.
Do I need to understand this to use AI well?
You do not need the engineering. You do need the mental model, because it is what tells you when to trust the output and when to check it. People with an accurate model of the mechanism consistently get more value and make fewer costly mistakes than people who treat the tool as a mind or a search engine.
Sources: Vaswani et al., “Attention Is All You Need,” 2017; Stanford Institute for Human-Centered Artificial Intelligence, AI Index Report 2025; foundational machine-learning literature on transformer language models and neural network training.
This content was compiled with the support of AI following in-depth research, then written and prepared for publication by the CEOtudent editorial team.
This post is also available in:
