AI Core Concepts: 11 Building Blocks Explained in Plain English

Every AI chatbot you’ve used runs on the same handful of ideas. Learn them once and a lot of AI news starts to make sense.

On the Periodic Table of AI, those ideas are the green tiles in the Core Concepts group. There are 11 of them. This post walks through all 11 in the order they show up when you type a question and get an answer.

Glowing blocks forming a sentence break into small pieces that flow through a web of green light nodes and reassemble into a new line of blocks
A question breaks into tokens, passes through the network, and comes back out as an answer.

The big picture

Artificial Intelligence

Artificial Intelligence is the umbrella term. It means software that does things we’d normally expect a person to do, like understanding language, spotting patterns, or making decisions.

It’s tile number 1 for a reason. Everything else on the chart is a piece of it.

Neural Network

A neural network is the engine under most modern AI. It’s a big stack of simple math units, linked together in layers. Each link has a strength, and training is the process of adjusting those strengths until the network gets good at its job.

The name comes from the brain, but don’t take the comparison too far. It’s a loose inspiration, not a copy.

Language Model

A language model is a neural network trained on huge amounts of text. Its core skill is guessing what comes next. Give it “peanut butter and” and it will say “jelly” with high confidence.

Do that guess over and over, and you get whole sentences, essays, and code. That one trick powers every chatbot you’ve used.

How a model reads your words

Tokens

A model doesn’t read words the way you do. It chops text into small pieces called tokens. A token might be a whole short word, part of a long word, or just a bit of punctuation. On average, a token is a bit less than one English word.

Tokens matter because they’re the unit everything gets counted in, including how much a model can handle at once and how much some services charge.

Embeddings

Next, each token gets turned into a long list of numbers called an embedding. The numbers capture meaning, so words that mean similar things end up with similar numbers. “Dog” and “puppy” land close together. “Dog” and “mortgage” don’t.

That’s how a model can tell that two sentences are about the same thing even when they share no words.

Transformer

The transformer is the design that made today’s AI possible. It was introduced in a 2017 research paper from Google called “Attention Is All You Need.”

Its big idea is attention. When the model reads a sentence, it works out which words matter most to each other. In “The trophy didn’t fit in the suitcase because it was too big,” attention helps the model figure out that “it” means the trophy.

What the model knows and how much it can hold

Parameters

Parameters are the numbers inside the network that training adjusts. Think of them as the model’s learned knowledge, stored as a very large pile of settings.

When you hear that a model has “70 billion parameters,” that’s the count. More parameters can mean more capability, but it also means more cost to run. Bigger isn’t automatically better.

Context Window

The context window is how much text a model can keep in view at one time. It covers your question, the model’s answer, and anything else in the conversation, all measured in tokens.

Once a conversation grows past the window, the model can’t see the early parts anymore. That’s why a long chat sometimes seems to forget what you said at the start.

Inference

Training is when a model learns. Inference is when it works. Every time you send a message and the model writes back, that’s inference.

Training happens once, ahead of time, and costs a fortune. Inference happens every time anyone uses the model, which is why running AI for millions of people takes so much computing power.

Two ways to build smarter

Mixture of Experts

A mixture of experts model is built from many smaller specialist networks instead of one giant one. A routing step picks a few of them for each piece of input, and the rest stay idle.

The payoff is a model that has a lot of total knowledge but only uses part of it at a time, which saves computing power.

Diffusion Model

A diffusion model is a different approach, and it’s the one behind most AI image generators. It learns by practicing on pictures with noise added, then learns to remove the noise step by step.

To make a new image, it starts from pure static and cleans it up a little at a time, guided by your description, until a picture appears.

How they fit together

Here’s the whole trip. You type a question. The model chops it into tokens and turns them into embeddings. A transformer uses attention to work out what matters. The parameters, which hold everything the model learned in training, shape the response. The context window sets how much of the conversation the model can see, and inference is the act of producing the answer, one token at a time.

That’s a simplified picture, but it’s accurate enough to hold onto.

Where to go next

If you want to see how these ideas connect to how AI gets taught, the next group on the chart is Machine Learning, and it’s the next post. In the meantime, click any of the tiles above and read the full page. Start with Tokens if you only have a couple of minutes.

1 thought on “AI Core Concepts: 11 Building Blocks Explained in Plain English”

Comments are closed.