What AI Actually Is

You've used ChatGPT or Gemini or Claude, but what actually happens when you hit enter?

Text Completion

Think of an LLM as a function

LLMs are a kind of AI technology that we know about because it is the underlying technology powering the Chatbots we all know: ChatGPT, Gemini, Claude etc. At a very high level you can think of an LLM large language model — as a function, the same kind you met in high school algebra. f(x) = x + 2: pass in 1, get 3 — the same thing every time.

The internals of an LLM is astronomically larger than an algebraic function, but it’s still text in, text out. Pass in "The cat sat on the" and it returns the most likely next word: "mat." That's the whole trick — predict the next word — done by billions of numbers doing simple math, very fast. Do that well enough, at scale, and you get something that can write, summarize, and answer questions.

ChatGPT is not an LLM

ChatGPT is an app wrapped around an LLM. Before your message ever reaches the model, the app loads it up with extras — your conversation history, your account details, the date, tone instructions.

That's why the same question gets different answers on different days: the wrapper (plus a randomness dial called temperature) changes what actually hits the model. The model itself is just math.

Experiment: The Round Trip

You know the theory, now watch it happen. Send a message and follow it through every box it touches. Run both rounds, and keep your eye on the middle box the second time:

Where did the billions of numbers come from?

Nobody typed them in. They were trained: take a mountain of text: books, websites, forums, code; hide a word, make the model guess it, and nudge the numbers a tiny bit whenever it guesses wrong. Repeat trillions of times on warehouse-scale computers. That whole process is what "training" means, and it happened once, months or years ago, before you ever opened the app.

When training ends, the numbers freeze. Two things you've probably noticed follow directly from that: the model has a knowledge cutoff. Ask about last week and it has to guess, or have the wrapper hand it search results — and it never learns from talking to you. Your chats don't change the numbers. When a headline says an AI is "learning," it's describing that one-time nudging process — not something happening while you type.

Catch it in the wild (5 min)

The simulations above are rigged — we wrote them. Time to catch the real thing doing exactly what they showed. Open a brand-new chat in whatever chatbot you use — ChatGPT, Claude, Gemini, doesn't matter.

1. Catch the wrapper red-handed. Ask: "What's today's date?" It answers — but the model's numbers were frozen long before today. It can only know because the app pasted the date into the bundle, exactly like you watched in Experiment 2.
2. Catch the missing memory. Open another brand-new chat and ask: "What did we just talk about?" Nothing. No bundle, no history: every chat starts from the same frozen function.
3. Catch the "memory" feature. If your chatbot does know your name or what you do for work, you caught the wrapper again: apps like ChatGPT keep saved notes about you and quietly paste those into the bundle, same as chat history. The model still remembers nothing.

It isn't conscious, it isn't moody, and it doesn't know you.

Decode the buzzwords

Now headlines read differently. When someone says the AI "hallucinated" or is "learning," there's a plainer truth underneath. Keep the Buzzword Decoder handy; it's the hype-to-reality translation table for eight terms you'll hear constantly.

Review in video

Go deeper (optional homework)

• Stop Talking About AI Like It's Human — why the words we use muddy our understanding (4 min)
• LLMs Don't Remember You! — the full story behind the no-memory experiment (5 min)
• AI Isn't Magic: Explain It to Your Friends — the deep dive, good for sharing (members) (8 min)

Next up

Lesson 2: Your first real automation. You'll stop chatting with AI and start making it do work for you — something real running on your computer in 20 minutes. Want a head start? Try Build Your Own Baby Chatbot!, where you make a tiny app that shows next-word prediction from the inside.