Ava was hungry again, and the fridge looked sad.

There was half a roast chicken left from Sunday, a bag of spinach that was one day away from being compost, a lemon, some pasta in the cupboard, and not much else. She stared into the cold light for a full minute before giving up and pulling out her phone.

“What should I have for dinner?” she typed into the AI app her mom had downloaded.

The reply came back in about two seconds:

“You could make a quick lemon-chicken pasta with the leftover roast chicken and the spinach before it wilts. Boil pasta, warm the chicken in a pan with olive oil and lemon, toss in the spinach at the end so it just wilts, finish with salt and pepper. Twelve minutes, one pot.”

Ava blinked.

She walked into the living room, holding her phone like it had just done a magic trick. “Mom. How did it know what was in our fridge?”

“Oh,” her mom said, not looking up from her book. “I told it this morning. I gave it a list.”

“OK, but… you told it ‘leftover chicken, spinach, lemon, pasta.’ You didn’t tell it how to cook. You didn’t tell it about me. How did it turn a grocery list into a dinner plan?”

Her mom shrugged. “It’s smart, I guess.”

But “smart” isn’t really an answer. Ava wanted to know what was actually happening in there.

Every Agent Has a Brain

In Chapter 2, we figured out that an agent doesn’t just talk — it acts. It plans, it reaches into the world, it does things.

But before any of that acting can happen, something has to do the thinking. Something has to look at “leftover chicken, spinach, lemon, pasta” and decide that yes, those ingredients can become a meal, and here’s a sensible order to cook them in.

That thinking part has a name: a Large Language Model, or LLM for short.

The LLM is the brain inside almost every AI agent you’ll ever meet. ChatGPT, Gemini, Claude, the assistant on your phone, the chatbot that pops up when you visit a help page — there’s an LLM sitting in the middle of every one of them, doing the heavy thinking.

So what is an LLM? Here’s the simplest way to put it:

A Large Language Model is a pattern-matching engine that has read so much text it can predict what words come next in almost any sentence — and that simple trick turns out to be powerful enough to write essays, answer questions, and plan dinners.

That sentence sounds wild. Predicting the next word can plan dinner? Let’s break it down.

The World’s Most Well-Read Librarian

Imagine a librarian. Not a normal one. The librarian.

This librarian has read every book in the library. Every cookbook, every novel, every science textbook, every history book, every poetry collection, every instruction manual. She’s also read every news article ever written, every Wikipedia page, every recipe blog, every forum post about how to cook chicken, every email signature, every Shakespeare play.

She has not just read these things — she has read them so many times that she remembers the patterns in them. She knows that the words “Once upon a” are almost always followed by “time.” She knows that recipes for pasta tend to include the words “boil,” “salt,” and “drain.” She knows that when someone lists “chicken, spinach, lemon, pasta,” cooks across thousands of recipes have combined those exact ingredients in a few classic ways.

Now here’s the strange part: this librarian can’t actually go look anything up. She doesn’t have the books in front of her anymore. All she has is the patterns soaked into her memory from years of reading.

But the patterns are enough.

Ask her, “What rhymes with ‘orange’?” and she’ll search her memory for words that have shown up near “orange” in rhyming poems. Ask her, “How do I cook chicken with spinach?” and she’ll pull from the patterns of thousands of recipes she’s seen.

That’s an LLM. It’s the world’s most well-read librarian, except instead of reading for sixty years, it read for a few months — because it was reading at the speed of a supercomputer.

How Much Text Are We Talking About?

The “Large” in Large Language Model is not a marketing word. It is doing real work.

When the company OpenAI built GPT-3 in 2020 — one of the models that kicked off the modern AI wave — they fed it about 45 terabytes of raw text scraped from the internet, which they then filtered down to roughly 570 gigabytes of high-quality text for training (according to the paper “Language Models are Few-Shot Learners” by Brown and his team, 2020). That filtered pile of text worked out to around 400 billion little chunks of language called tokens — roughly the size of a stack of words that, if you printed them out, would tower past the clouds.

To put that in human terms: if you digitized every printed book in the Library of Congress — and the Library of Congress holds tens of millions of books — you would get about 10 terabytes of text. (That estimate comes from a study by information scientists Peter Lyman and Hal Varian back in 2000, and it’s still the number people quote.)

So GPT-3’s raw training pile was roughly four to five times the size of every book in the largest library on Earth.

And GPT-3 was 2020. That’s ancient history in AI years.

By 2024, Meta released Llama 3, a free-to-download LLM trained on more than 15 trillion tokens — nearly 40 times the size of GPT-3’s filtered pile. The biggest version, Llama 3.1, has 405 billion parameters (parameters are like the tiny dials inside the model’s brain — the more dials, the more nuance it can capture). Models being built today are bigger still.

Year	Model	Training text (tokens)	Parameters
2020	GPT-3	~400 billion	175 billion
2024	Llama 3.1	15+ trillion	up to 405 billion

No human will ever read this much. You couldn’t do it in a thousand lifetimes. The LLM read it in months.

The Magic Trick: Predict the Next Word

Here is the part that surprises almost everyone the first time they hear it.

An LLM, deep down, is doing one very simple job: looking at some words and guessing what word comes next.

That’s it. That’s the whole trick.

When Ava typed “What should I have for dinner?” the LLM didn’t understand her question the way a friend would. It looked at her words and started predicting, one word at a time, what a helpful response would probably look like.

First word: “You.” Why “You”? Because in the patterns the LLM has seen, when someone asks “what should I have for dinner,” the answer often starts with “You could…” or “You might…” or “Try…”

Second word: “could.” Because the patterns say so.

Third word: “make.” Then “a.” Then “quick.” Then “lemon-chicken.” Then “pasta.”

It builds the answer one word at a time, each time asking itself the same little question: given everything I’ve written so far, what word is most likely to come next?

It’s a bit like the autocomplete on your phone, but with a memory that has read a sizable chunk of the internet. Your phone’s autocomplete guesses the next word based on what you type. An LLM guesses the next word based on what humanity has typed.

And it turns out — and this is genuinely strange and beautiful — that if you get really, really good at “guess the next word,” you accidentally get good at almost everything else. Writing essays is just guessing the next word, over and over. Explaining science is guessing the next word. Suggesting a dinner is guessing the next word. Even reasoning step by step (“first I’ll do this, then I’ll do that…”) is just guessing the next word in a thoughtful pattern the model has seen humans use.

What LLMs Are NOT

This is important, because a lot of grown-ups get it wrong.

An LLM is not looking things up. It can’t open a book. It can’t check Wikipedia. It only has the patterns soaked into its memory from training. If the training ended in 2024, it doesn’t know what happened in 2025 — unless an agent fetches that info for it (we’ll get to that).

An LLM is not “understanding” the way you understand. When you read “Ava is hungry,” you picture a kid with a rumbling stomach. An LLM doesn’t picture anything. It just knows that “Ava is hungry” tends to be followed by words about food, kitchens, and snacks. Its “knowing” is pattern, not feeling.

An LLM does not know what’s true. It knows what’s common. If a million web pages say something wrong, the LLM may confidently repeat it. This is why AI sometimes “hallucinates” — invents facts that sound right but aren’t. It’s not lying. It’s just predicting the next plausible-sounding word, and sometimes plausible isn’t the same as correct.

Knowing what an LLM isn’t is just as important as knowing what it is. It’s a brilliant pattern-matcher. It is not an oracle.

A Brain Needs a Body

Here’s the last big idea of this chapter — and it’s the bridge to everything that comes next.

The LLM is the brain of an agent. But a brain alone can’t do anything. A brain in a jar can think, but it can’t open the fridge.

Ava’s dinner-suggesting AI is just an LLM by itself — it gave her words, not action. It didn’t actually cook. It didn’t add anything to her grocery list. It didn’t set a timer for the pasta. If she wanted those things, she’d need an agent — an LLM with a body.

What’s a “body” for an LLM? Things like:

• Tools it can use — a calculator, a calendar, a web browser, a smart-home controller
• Memory of what’s happened recently — what Ava asked yesterday, what’s in the fridge
• The ability to take actions in the real world — sending messages, updating files, controlling devices

In the next few chapters, we’ll see how all those pieces snap onto the LLM-brain and start working together. That’s the moment a “chatbot” becomes an “agent.” That’s the moment talking becomes doing.

Key Takeaways

• Every AI agent has an LLM at its core — a Large Language Model is the “brain” that handles all the thinking, planning, and language
• LLMs are pattern-matchers, not lookup machines — they’ve absorbed huge amounts of text and learned what words tend to follow what
• GPT-3 (2020) was trained on roughly 570 GB of filtered text — about 400 billion tokens — filtered down from 45 TB of scraped internet text. Today’s models like Llama 3.1 (2024) train on 15+ trillion tokens
• The whole trick is predicting the next word, one at a time — which turns out to be enough to write, explain, and reason
• LLMs don’t truly “understand” — they predict. That’s why they sometimes confidently get things wrong
• An LLM alone is just a brain in a jar. To become an agent, it needs a body — tools, memory, and the ability to act

---

So the brain can think. It can take “leftover chicken and spinach” and spin out a recipe. But thinking once isn’t enough — a real agent has to keep going. It has to start a plan, check whether the plan is working, notice when something goes wrong, and adjust. How does it do that, again and again, without a human pressing a button each time? The answer is one of the simplest and most powerful ideas in all of AI: the loop. In the next chapter, we’ll watch one spin.