UNDERSTAND AI
What is an LLM? (and what “AI” even means)
“AI” and “LLM” are everywhere, rarely explained. Here’s the short version, in everyday words: the things you chat with — ChatGPT, Claude, Gemini — are large language models, or LLMs. An LLM is a computer program that was trained on an enormous amount of human writing to do one core thing: predict language. That’s it. It’s a remarkable tool, and it is also far simpler at heart than it can feel. This page is the front door — what the machine is, and, just as important, what it isn’t.
First, “AI” vs. “LLM”
“AI” — artificial intelligence — is the big umbrella word. People use it for all sorts of software that does things we used to think only people could do: recommending a movie, recognizing a face in a photo, flagging spam, helping steer a car. It’s a broad, loose term, and it’s been around for decades.
An LLM is one specific kind of AI — the kind built for language. When you type to a chatbot and it writes back in fluent sentences, that’s an LLM doing the work. So every LLM is “AI,” but not everything called “AI” is an LLM. When this site says “AI,” we almost always mean the chat kind — the LLM.
What “model” actually means
The word that trips people up is model. It sounds like a thing full of stored answers — a giant encyclopedia, or a search engine with everything filed away. It isn’t that.
A model is a pattern-machine. During training, the program was shown a vast amount of human-written text — books, articles, websites, conversations — and it gradually adjusted itself to capture the patterns in how language fits together. It isn’t, at heart, a stored copy of the text, filed away to be looked up later. It learned, in effect, how words tend to follow other words: which ones go together, in what order, in what tone. What gets saved at the end isn’t a library of facts. It’s a huge web of those patterns. That’s the “model.”
A useful comparison: think of someone who has read so much that they’ve developed a deep feel for how sentences go — without keeping a perfect copy of every book in their head. The model has the feel, much more than the copy. That’s why it can sound authoritative and still get a fact wrong: it’s producing language that fits, not looking an answer up.
How it works, in one idea
Underneath all of it is a single, surprisingly simple job: predict the next word.
You give the model some text — your question. It looks at everything so far and predicts what word is most likely to come next, based on the patterns it learned. Then it adds that word and predicts the next one. Then again, and again, one piece at a time, until it has written a whole reply. The fluent paragraph you read was built like that — word by word, each one a prediction of what fits.
That’s the whole engine. It feels like a conversation, but mechanically it’s a very, very good guess about what comes next, repeated over and over. (We have a separate page that walks through this slowly, with examples — it’s linked at the bottom.)
Go deeper: the technical version
To be more precise: an LLM doesn’t work in whole words but in tokens — chunks of text, often a word or part of a word. At each step it produces a probability distribution over all possible next tokens (a ranked list of how likely each one is), then picks from that list — sometimes the top choice, sometimes a slightly less likely one, which is part of why answers vary. The patterns it uses live in billions of numbers called parameters, tuned during training by a process that repeatedly nudged them to make its predictions match real human text more closely (the architecture behind modern LLMs is called a transformer). The knowledge and apparent personality aren’t a separate database or a bolted-on mind — they emerge from that same next-token prediction running at scale. Some newer products add a “reasoning” mode, where the model is prompted to work through steps before it answers, or wire it to outside tools like a search engine — but those are extra steps built on top of the same prediction engine, not a different kind of thinking underneath. That’s also why a confident-sounding wrong answer — often called a hallucination — isn’t really a malfunction so much as the system doing exactly what it does: producing fluent, plausible text, whether or not it happens to be true.
What an LLM is not
This is the part that matters most, because almost every way AI gets people into trouble starts with a wrong picture of what it is. So, plainly:
- Not a search engine. On its own it isn’t looking things up on the live internet and reporting back — it’s generating likely-sounding text. (Some products bolt a real search onto the LLM — but the language part itself isn’t searching.)
- Not a database of facts. There’s no filing cabinet of verified truths inside. It can state something false as confidently as something true, because to the machine both are just well-formed language.
- Not a mind. As far as anyone can tell from how these systems work, there is no inner experience behind the words — no understanding the way a person understands, no awareness of what it’s saying.
- Not a person, and not aware. It doesn’t want, feel, or remember you the way a friend does, and it doesn’t know that you exist. When it says “I” and sounds like it does, that’s the pattern of human speech it learned — not a someone on the other end.
Here’s the thing to hold onto: an LLM can be genuinely useful and have none of those properties. Both are true at once. It can help you write, think, learn, plan, and explain — really — while being, underneath, a pattern-machine predicting words. Useful does not require a mind.
Why it feels like more than a tool
The reason an LLM can feel like a someone is simple: fluent language coming out makes us assume a thinking being is inside. That’s a deep human instinct — for our whole history, anything that spoke in full, warm, intelligent sentences was a person. LLMs are the first thing that can do that without being one.
So the fluency isn’t evidence of a mind. It’s evidence that the machine learned our language well. That gap — between how alive it sounds and what it actually is — is where the misunderstandings that hurt people tend to live. Knowing the machine is the machine doesn’t make it less useful. It makes you harder to mislead.
In one line: an LLM is a computer program that learned the patterns of human language from a huge amount of text, and produces answers by predicting what word comes next — over and over. It’s a real, powerful tool, and it is not a search engine, a fact database, a mind, or a person. Fluent language out doesn’t mean a thinking being in.
Where to go next
It talks like a person — is it one?
The most natural question, answered plainly: why fluent speech feels like a mind, and what’s really there.
Read this next →How it actually produces words
The next-word prediction idea, walked through slowly with examples — the engine in plain sight.
See how it works →Where the gap starts to matter
When “fluent” and “true” come apart, it can affect people — one documented pattern, and the evidence, stated so you can check it.
The research →Spotted something here that’s unclear, or a place this explanation falls short? Tell us.