The AI Words You'll Hear: A Plain-English Glossary

Somewhere soon, someone's going to say "oh, it's just a neural network with billions of parameters" — in a podcast, a meeting, a group chat — and you'll catch that little flicker of should I know what that means? go around the room. Here's the quiet secret of this whole chapter: you already do.

Every intimidating word in AI is just a label for something you've already met in plain language. So this is the optional pop-the-hood tour — and remember, way back in the second lesson, popping the hood was always a bonus, never a requirement. You don't need a single word in here to use AI well. But they're quick, and honestly kind of fun to defang. So let's go.

Start with the one that keeps following you around: tokens

You've bumped into tokens twice already — as the size of the desk back when we talked about what the AI can hold, and as the taxi meter when we talked about paying for the API. Both times with a quiet "we'll get to it." This is getting to it.

A token is just a chunk of text the AI reads and writes in — usually a whole word, sometimes a piece of one (walking might be handled as walk plus ing). That's the unit behind the things you've already seen: the desk holds so many tokens, and the API meter ticks up token by token. That's the entire idea. You don't need to know how the chopping-up works — just that the AI counts in chunks, not letters.

How it got built: training, neural network, parameters

These three travel together, and they're all about how the AI was made — long before you ever typed anything into it.

Training is the big one. It's the stretch where the model learned its patterns by reading that staggering pile of text — the "Large" in "Large Language Model" from the very first lesson. The part worth holding onto: it happened once, done by the maker, and then it stopped. The model is baked and set before you show up. That's exactly why, earlier, correcting it mid-chat never stuck — it was trained once, then frozen.

A neural network is the kind of structure that does that learning. The name sounds like brain surgery; the idea is gentler than that. It's loosely inspired by the way brain cells link up and pass signals along, and it's simply very, very good at picking up patterns from data. That really is all you need to carry. The actual math is a whole career — and squarely the "mechanic" path you already decided you were happy to skip.

Parameters are what the training actually tunes. Picture a huge bank of tiny dials inside the network, each nudged a hair during training until the whole thing hums. So when someone says a model has "seventy billion parameters," they're really just quoting its rough size — more dials, loosely speaking, more capacity. Hear the number as a ballpark, not a score.

The ones you've already got

Some of these you can practically wave through — you met them chapters ago. Four in particular sit inside each other, exactly the way we drew them in the very first lesson:

And a few more that have already come up by name:

• Model — the industry's word for the trained system itself. The first lesson's tip still works: mentally swap in "the AI brain that does the work."
• Prompt — simply what you type in. A whole lesson went into giving it a good one: role, task, target.
• Hallucination — from the last lesson: a confident, made-up answer, delivered in the very same calm voice it uses when it's right.

A few more you'll bump into

Last handful — the ones that turn up in headlines.

Multimodal describes an AI that handles more than one kind of content — text and images and audio, say. Remember the one trick wearing different costumes, from the lesson on what AI can make beyond words? A multimodal model is one that wears several of those costumes at once.

Open-weight (you'll also hear "open-source") means the maker shares the model itself, so anyone can download it and run or tweak it on their own machine — rather than only reaching it through the maker's app. Most big-name tools are closed; a few makers, like Mistral, choose to share theirs.

API you've technically already met — the "builder's door" from the lesson on paying. It's the pay-per-use pipe developers wire a model into their own software through, billed by the token.

Fine-tuning is the one genuinely new idea here: taking a model that's already been trained and nudging it with a batch of specific examples, so it gets sharper at one particular job. The giant training is the foundation; fine-tuning is a light coat of paint on top.

So where that leaves you

Look back at what just happened. Every one of those words turned out to be a label for something you already understood — a chunk of text, a learning phase, a bank of dials, a thing tucked inside a bigger thing. Not one of them was the wall it pretended to be.

So the next time one of these lands in a conversation, you won't flinch. You'll nod. And if poking around under the hood turned out to be the fun part for you — that's worth noticing. There's a whole path for the people who want to know how the engine works, or build one of their own; the second lesson called them the mechanic and the engine designer. You never have to walk through that door. But you've more than earned the right to, any time the curiosity strikes.