Large language models, explained gently
Facts effective October 18, 2026 · re-checked every month
Every chatbot you’ve heard of — ChatGPT, Claude, Gemini, Grok, Copilot — is a shop window. The engine behind the glass is a Large Language Model, or LLM. Here’s what those three words actually mean, without a single equation.
Start with autocomplete
Your phone already predicts your next word when you text — type “Happy” and it offers “birthday.” Mildly useful, frequently wrong, occasionally embarrassing.
Now imagine that same idea scaled up beyond recognition: trained not on your last hundred texts but on a meaningful fraction of everything humanity has written. And instead of predicting one word, it predicts the next word, then the next, then the next — thousands of times in a row, each choice informed by everything that came before it in your conversation.
At small scale, that gives you “Happy birthday.” At vast scale, something remarkable happens: predicting language this well requires absorbing the patterns of the world the language describes — how arguments hang together, how recipes proceed, how condolence notes differ from complaint letters. Master the language and you inherit a working sketch of everything the language talks about.
The three words, one at a time
“Large Language Model” sounds like something you’d need a degree to unpack. You don’t. Take the words one at a time, the way you’d read the ingredients on a label.
Large
Two kinds of large. First, the reading: it trained on a share of nearly everything humanity has published — more than any person could read in a hundred thousand lifetimes. Second, the machinery: billions of tiny internal dials, each nudged a hair at a time during training until the whole thing predicts language well. Large is not a boast here. It is simply the scale at which the trick starts to work.
Language
This is the surprising part. It wasn’t taught facts, or grammar rules, or history. It was taught language — and because language is how we carry every idea we’ve ever had, learning language well drags along a rough working knowledge of the world the language describes. Master how sentences about medicine fit together and you absorb a good deal about medicine, sideways.
Model
A model is a working imitation, the way a model train imitates a railroad or a globe imitates the Earth. It isn’t the real thing; it’s a scaled-down stand-in that behaves enough like the real thing to be useful. An LLM is a model of how human language behaves. Useful, and — like every model — not to be mistaken for the territory itself.
Why it sounds so human
Because it learned from us. Every graceful sentence it produces is an echo of patterns in human writing. That’s also why it has our flaws: it picked up our clichés, our confident wrongness, our tendency to fill silence rather than say “I don’t know.”
It helps to know what an LLM is not, because the missing pieces explain most of its odd behavior. It is not connected to your bank, your doctor, or your memories unless you hand it those things in the conversation. Out of the box it has no clock and no calendar — ask it today’s date and it may guess. It does not learn from your chat the way a friend would remember your birthday; when the conversation ends, it forgets, and that is a mercy for your privacy. And it has no stake in being right. A person who bluffs at least feels the little pang of bluffing. The model feels nothing, which is exactly why it bluffs so smoothly.
Why it makes things up
Remember: it composes; it doesn’t look up. When a question lands in territory where its patterns are thin, it does exactly what it was built to do — produce the most plausible-sounding continuation. The result is a “hallucination”: a confident, fluent, wrong answer. A book that doesn’t exist. A statistic with no source. A court case no court ever heard.
This isn’t lying — lying requires knowing better. It’s more like a gifted after-dinner storyteller who would rather be interesting than accurate. Enjoy the company; verify the facts.
The verification habit, made specific
For anything involving money, health, law, or a decision you’d hate to redo: ask the AI for its sources, then check one yourself — or simply ask your pharmacist, doctor, or banker. Two minutes of checking beats two weeks of untangling.
Why it feels like it “remembers” — and mostly doesn’t
Within a single conversation, the model can refer back to what you said three questions ago. That feels like memory, and it’s worth understanding what it actually is, because it shapes how you use the tool. Each time you send a message, the model quietly re-reads the whole conversation so far and then predicts what comes next. It isn’t remembering you; it’s re-reading the transcript, start to finish, every single time — like a diligent assistant who has no memory but keeps flawless notes and glances at them before every reply.
Two useful consequences fall out of that. First, very long conversations can get hazy at the edges, because there’s a limit to how much transcript it can hold in view at once — the oldest parts scroll out of sight. If a chat has wandered for an hour, starting fresh often gives you a sharper answer. Second, when you close the window, those notes are gone. The next conversation begins with a clean slate and no recollection of you — which is precisely why you never need to worry that today’s question will resurface embarrassingly tomorrow.
“I spent twenty-five years reading witnesses. The confident ones weren’t always the accurate ones — and a jury can’t tell the two apart by tone alone. Neither can you tell a chatbot’s good answer from its bad one by how sure it sounds. So I do with it exactly what I did in a deposition: I ask for the source, and I don’t move until I’ve seen it.” — Gay
Ten words you can use at dinner
You don’t need the vocabulary of an engineer. You need ten words, each with a picture attached — enough to follow the conversation, ask a sharp question, and, if the mood strikes, gently correct the person who’s bluffing. Each one below comes with a line you could actually say out loud. (The full list lives in the glossary.)
LLM
Large Language Model — the engine humming behind every chatbot. When people say “the AI,” this is usually the thing they mean.
Say it like this: “ChatGPT is just the storefront. The LLM is the engine in the back doing the work.”
Training
The long, one-time process of tuning the model on oceans of text. It finished before you ever typed a word — the model isn’t reading over your shoulder now.
Say it like this: “It was trained before it ever met me, so it doesn’t know a thing I haven’t told it.”
Prompt
Whatever you say to it — your question, your request, your instructions. Better prompt in, better answer out. That’s the whole of Lesson 4.
Say it like this: “My first prompt was lazy, so I gave it more detail and the second answer was far better.”
Hallucination
A confident, fluent, entirely wrong answer — a fact it invented because inventing was easier than admitting a blank. Now you know why it happens.
Say it like this: “It gave me a book title that doesn’t exist. Classic hallucination — I checked before I ordered.”
Model
The trained AI itself — the bundle of learned patterns. New versions ship regularly, like car models, usually a bit smarter than the last.
Say it like this: “They released a new model this spring. Still no substitute for checking, mind you.”
Token
The word-pieces the model actually reads and writes — ‘grandmother’ might be two or three of them. It counts these, not words, which is why its limits are phrased so oddly.
Say it like this: “It doesn’t count words, it counts tokens — like a butcher pricing by the ounce.”
Context window
How much of your conversation the model can hold in view at once. Fill it up and the earliest bits scroll out of sight, which is why long chats get hazy.
Say it like this: “The chat got muddled, so I started fresh — the old one had run past its context window.”
Chatbot
The program you actually converse with in plain language — the storefront over the engine. ChatGPT, Claude, Gemini, Grok, and Copilot are the famous ones.
Say it like this: “I asked the chatbot to explain the letter as if to an intelligent friend. It obliged nicely.”
Generative AI
AI that makes something new — a paragraph, an image, a tune — rather than just sorting or flagging what already exists. The “write me a poem” kind.
Say it like this: “That’s generative AI — it didn’t find the birthday poem, it wrote one on the spot.”
Prompt engineering
A grand name for a simple habit: getting good at asking. Give it a role, some context, and clear instructions, and the answers improve dramatically.
Say it like this: “Prompt engineering is just a fancy way of saying I learned to ask a better question.”
The whole lesson, on one card
If you remember nothing else, remember this much — it’s enough to use these tools wisely for the rest of your life.
- An LLM predicts the next word, over and over, from patterns it learned in an ocean of human writing. That’s the engine.
- Predicting language that well means it picked up a rough working knowledge of the world the language describes. That’s why it’s useful.
- It composes answers rather than looking them up, so when its patterns run thin it invents — fluently and confidently. That’s the hallucination.
- It has no memory of you between conversations and no stake in being right. So enjoy the company, and verify the facts.
Next: meet the big five
Same engine idea, five different showrooms. Which chatbot should be your first?