AI Basics

Artificial Intelligence, Explained

What an AI model actually is, in plain language, and the four layers of the platform built around it. No background assumed.

Start here

AI in Under Two Minutes

A ground-up explanation of AI for people with zero technical background.

AI works like the predictive text on your phone, but on a much larger scale.

Training data

The text an AI model was built from (books, Wikipedia, Reddit, news, code, and a great deal of the rest of the web). It is read once, while the model is being built, and is never consulted again while you type.
Instead of predicting your next word from your own past messages, an AI model predicts the next piece of text using patterns it learned during training.

AI model

A specific prediction algorithm, built from a fixed set of training data. ChatGPT runs on GPT models, Claude on ones named Opus, Sonnet, and Haiku. A new version means new training, not a software update.
A model uses training data to create a prediction equation, like using individual dots on a graph to determine a line of best fit.

Parameters

The numbers that make up the equation. (So in the famous y = mx + b there are two parameters, m and b.) GPT-3 had 175 billion of them (Brown et al., 2020). Open-weight models still publish their counts, now into the hundreds of billions; the companies behind ChatGPT, Claude, and Gemini do not publish counts for their flagship models.
This explains why AI is not always 100% accurate. It doesn't save or remember its training data verbatim, only the equation it helped create.

Pretraining

Building an AI model by fitting it to a body of training data, the way a line of best fit is fitted to a set of dots. What comes out is the equation, not the original dots.
It is also why an AI model can hallucinate. It is not looking anything up. It is applying the equation.

Hallucination

Output that is fluent, confident, and wrong. Hallucination is not a malfunction, just ordinary behaviour making the wrong prediction. A system whose only task is to continue text plausibly has no separate step at which it checks whether the continuation is true.
When you use an AI model, what you type (the prompt) is fed into the equation, along with the conversation so far and anything else you have handed it (documents, photos, instructions, etc.), so the model can build a response one token at a time, each token the most likely thing to come next.

Token

A chunk of text, very roughly a short word or part of one. Models read and write in tokens, not letters or sentences.

Turn

One exchange: your message, and the answer that follows it. The model runs once per turn and re-reads everything it has been given each time, because nothing carries from one turn to the next on its own.

Going deeper

Inside an AI Platform

AI platforms have different layers at which memory functions can happen.

The weights

The parameters. Fixed at release, and the same for every user.

+ what it sees

The context window

The whole conversation so far, re-read from scratch every turn.

This is the model

+ what is kept

The memory layer

Notes kept beside the model, pasted back in.

+ what is allowed

The product

Everything wrapped around it: prompt, tools, and refusals.

This varies from platform to platform

In detail

The Four Layers of an AI Platform

01 The weights

The equation itself, fixed at the factory. Its parameters were set once, from the training data, and that text is not kept by the time you type. The model does not consult it. What is left is the equation it produced, which still carries much of what the text said.

Where you've met it: The difference between GPT-4 and GPT-5, or Sonnet and Opus. A new model means new weights, and nothing you do in a chat edits them.
What it actually is: Billions of parameters, and hundreds of billions in the largest published models, fixed at release and identical for every user in the world.
What changes it: The lab, at training time, months apart. New training, not an update.

02 The context window

The message you are currently typing, except that the entire conversation counts as the message. Predictive text sees the words in front of it and nothing else; send it, start a new message, and it has no idea what you just wrote.

Where you've met it: The conversation you are in right now. It is why a long chat begins to drift, and why a new one starts blank.
What it actually is: The text handed to the model on this turn: your message, the conversation so far, files, and instructions. The model re-reads all of it from scratch every turn.
What changes it: Every turn. It is discarded when the session ends.

03 The memory layer

Someone sitting beside you with a notebook, who writes down things you have said and, before you type, slips the pages they judge relevant onto your screen, where the keyboard reads them as though you had typed them yourself.

Where you've met it: ChatGPT's "Memory updated" notice, Claude's memory across chats, the setting that turns it off.
What it actually is: An ordinary database beside the model, a rule for what to write down, and a similarity search for what to fetch back. Whatever it fetches is pasted into the context window, because that is the only way anything reaches the model at all.
What changes it: A write policy someone at the company chose, running at the moment you speak.

04 The product

The app the keyboard lives in, which decides whether you get a notebook at all, who is allowed to read it, and what the keyboard will refuse to type.

Where you've met it: The name, the personality, the refusals, and the settings. It is also why ChatGPT and Claude answer the same question differently.
What it actually is: Everything wrapped around the model: interface, system prompt, tools, safety layers, and the defaults you did not pick.
What changes it: A product decision, shipped whenever.