If your eyes glaze over every time someone says "large language model" — welcome. You are in very good company. The phrase has become shorthand for "the thing behind ChatGPT", and most of us nod along without ever being handed a clean, short definition we can actually use.
I avoided learning how these tools actually worked for longer than I'd admit. It felt like it belonged to people with computer science degrees, and every explainer I read turned into a maths lecture by the third paragraph. The mental model below is what finally made it click — and once it clicked, my prompts got better almost immediately.
This is the first piece in our Terminology Tamer series. The aim is simple: by the end, you can explain what an LLM is in one sentence, picture what it's actually doing when you talk to it, and make sharper decisions about when it's the right tool for the job.
The one-sentence answer
A large language model is a text-prediction system trained on huge amounts of text, so it can produce the most statistically plausible next words for whatever prompt you give it.
That's it. That's the definition we're going to unpack.
It is not a search engine. It is not a database. It is not a reasoning mind behind a chat window. It is a very capable pattern generator. Keep that sentence nearby while you read the rest of this — if anything here feels wobbly, it's usually because we've quietly drifted back into treating an LLM as something it isn't.
The "large" bit is doing real work in that definition. It refers to the scale of the training text and the number of internal parameters (the dials the model tunes during training). The specific numbers reported for named models tend to be in the billions of parameters and trillions of words of training text, but we won't pin those down precisely — vendors rarely publish exact figures, and the number changes with every new release. What matters for a working professional is the shape of the thing, not the precise figures.
How to actually picture what an LLM does
Here is the mental model I wish someone had handed me on day one.
Imagine the autocomplete on your phone. You type "I'll be there in ten" and it offers "minutes". That suggestion comes from a small model that has been trained on patterns in text messages. You accept it, you carry on.
Now imagine that same idea, but scaled up astronomically. Instead of the last few words of a text message, the model has been trained on a huge and varied body of text — books, articles, transcripts, code, forum posts, documentation. Instead of suggesting the next word from a handful of options, it ranks every possible next token (a token is roughly three-quarters of a word for English on average — other languages can be more or less efficient) and picks one based on how probable it is given everything you've written so far.
Then it does that again. And again. And again, token by token, until it produces something that looks like a coherent answer.
That is the whole trick.
There is no tiny person behind the screen consulting a library. There is no lookup in a fact table. There is a single, astonishingly large pattern generator that takes your prompt, runs it through billions of learned connections, and produces "the next word most consistent with everything I've been trained on."
When we talk about prompt quality — and we talk about it a lot on this site — this is why it matters. The model isn't guessing your intent. It's predicting plausible continuations of the text in front of it. Give it vague text in, you get vague text out. Give it a crisp brief with context, task, format, and constraints, and the statistical terrain changes entirely. The four-part prompt formula exists for exactly this reason.
🧠 Quick Myth Buster: True or false — the underlying language model itself remembers your past conversations the way a colleague would.
Answer: False. The model itself has no memory beyond the text inside the current conversation window. Anything that feels like it "remembering you across sessions" is a product-level memory feature — a separate layer that quietly re-injects notes from past chats into each new one. ChatGPT, Claude and Gemini all now ship some form of this, and in most tiers it is on by default, which is why it often feels like the model remembers you. It doesn't. The product does. Knowing the difference changes how much you trust it.
Five things an LLM is not — and why each one matters for how you use it
This is the section I'd have taped above my monitor when I was starting out. Each of these mistakes costs real time.
1. It is not a search engine. Google indexes the live web and returns links. An LLM has been trained on a snapshot of text and generates new words each time you ask. Ask it for "the latest figures on X" and, unless it's been connected to a live search tool, it will happily produce confident-sounding numbers that were plausible during training but may not reflect the current state of the world. For anything time-sensitive, either use a tool with real web search built in, or bring the current source material into the prompt yourself.
2. It is not a database. Databases retrieve exact records. LLMs reconstruct plausible text. That's why two runs of the same prompt can give you two slightly different answers, and why it can occasionally invent a statistic, a citation, or a court case that sounds exactly right and does not exist. If the output has to be exact — a phone number, a clause from a contract, a specific figure — pull it from the source, not from an LLM's memory.
3. It is not reasoning like a person. This is the hardest one to accept, because the outputs often look like reasoning. Ask an LLM to walk through a problem step by step and it will produce something that reads like thinking. What it's doing is generating text that matches the patterns of reasoning it has been trained on. Sometimes that's close enough to be useful. Sometimes it's a confident-sounding answer to a question it didn't actually work through. Always sanity-check the logic, especially on anything consequential.
4. It is not always up to date. Every model has a training cut-off date, beyond which it has no information. Even with a live search tool attached, the model's internal picture of the world has a boundary. If you're asking about something that happened last month, last year, or in a fast-moving field, assume the model might be out of date and check.
5. The model itself has no memory across sessions — even when the product does. We covered this in the pivot above — worth repeating because it's the assumption that catches people out most. Inside a single conversation, the model retains what you've said. Outside of that, anything that feels like memory is a product feature stitched on top, not the model remembering you. Most major tools now ship one (and in many tiers it is on by default), but the notes it keeps are short, lossy, and not always accurate. If you have a favourite context or style you use often, keep the authoritative version somewhere you can paste back in — or save it as a reusable starting point in the Prompt Library.
Each of these "is nots" maps to a habit. Treat an LLM as a search engine and you'll get stale or invented facts. Treat it as a database and you'll trust a number you shouldn't. Treat it as a colleague with perfect recall across weeks and you'll be bewildered when it "forgets" your project. Treat it for what it is — a text generator that is extraordinary at patterns and honest about nothing on its own — and you'll brief it properly.
💬 "The most useful shift I made was to stop asking an LLM what it knows, and start asking it to work with what I give it."
Why this matters for your actual work
Once this mental model lands, a few things change about the way we brief these tools.
We stop asking narrow factual questions in isolation and start pasting in the source material. Summarising a report you've uploaded is very different from asking the model to recall it — the first plays to the model's strength, the second invites a hallucination. Accurate summaries beat confident recall every time.
We stop treating the first output as a finished product. Because the model is predicting plausible next words, the first draft is almost always a generic version of plausible. Iteration — "make paragraph two more concrete", "cut the hedging", "rewrite in the voice of our handbook" — is where the quality actually lives. Writers wouldn't publish a first draft. This is the same idea.
We become realistic about which tasks are genuinely well-suited to an LLM and which aren't. Drafting, summarising, restructuring, brainstorming, translating, and explaining — strong suits. Retrieving precise real-world facts, doing fresh arithmetic on live data, or producing something that must be exactly right the first time — weaker suits without extra scaffolding. This is the same instinct behind the Educate pillar of our ICE Method: understand the shape of the tool before you delegate serious work to it.
And finally, we get kinder to ourselves. If you've had a bad experience with an AI assistant recently, it's almost certainly not because you're bad at AI. It's usually because, at some quiet level, the brief assumed the model was something it isn't.
Where to go next in the glossary
If you want the full walkthrough of what happens inside the model — training, tokens, context windows, the whole end-to-end pipeline — our companion piece How AI Works: Understanding Large Language Models is the natural next stop. This article is the "what is it" version. That one is the "how does it actually run" version.
For quick-reference definitions of the other terms that show up alongside LLM — tokens, context window, temperature, hallucination, system message — the AI Glossary: Essential Terms keeps each one under a paragraph so you can look up whatever comes up in your next meeting without a rabbit hole.
And if you'd rather stop reading and start using one of these tools with a clearer head, all three of the main consumer models have free tiers available right now: ChatGPT, Claude, and Gemini. The mental model is the same for all of them.
Your next step
You now have a one-sentence definition, a mental model you can sketch on a napkin, and five clear "is nots" to hang your prompting habits on. That is a genuinely useful amount of ground to cover in under ten minutes of reading.
If you'd like to turn this understanding into sharper day-to-day results, the AI for Beginners learning path walks through your first real workflows step by step, and the Prompt Library has ready-made briefs you can borrow for the kinds of tasks LLMs genuinely shine at.
Take what's useful from this piece and leave the rest. You don't need to become an engineer to work confidently with these tools — you just need an honest picture of what they are. You've got that now. Build from here.