An interactive explainer

What is
AI?

From the chip that does the thinking to the number that holds the knowledge — explained plainly.

Scroll to explore
01

The big idea

Artificial Intelligence is software that learns patterns from examples rather than following rules someone wrote down. Instead of a programmer saying "if the photo is dark and furry it's a cat," the AI looks at millions of cat photos and figures out the pattern itself.

AI is a program that improves at a task through experience, rather than being explicitly told how to do it.

// Core definition

Modern AI — particularly the kind that powers chatbots, image generators, and language models — relies on four building blocks you'll see everywhere: CPUs, GPUs, Tokens, and Parameters. Here's what each one does.

🔲

Hardware

CPU

The general-purpose brain of a computer. Fast at sequential tasks, great at logic and decisions.

Hardware

GPU

Thousands of smaller processors working in parallel — perfect for the matrix maths AI needs.

🔤

Data

Token

The unit of text an AI reads and writes — roughly a word or part of a word.

🔢

Knowledge

Parameter

A number stored inside the AI model that encodes everything it learned during training.

Where to Start Learning AI

Free, world-class AI fluency training is now publicly available

MIT Open Learning launched Universal AI on 12 May 2026, a self-paced programme that takes learners from "AI novice to authority" with a free introductory course (Fundamentals of Programming and Machine Learning) and five further core courses plus six industry-specific courses — including Holistic AI in Medicine, AI and Entrepreneurship, and AI and Sustainability: Energy — on the MIT Learn platform. COAD's design assumes that the training infrastructure exists at this kind of quality globally and is no longer the binding constraint; what is missing is the income-bridge mechanism during the workforce-transition window, which COAD is designed to deliver. Universal AI is the kind of programme a COAD recipient cohort would access during the transition.

Source: Universal AI is "a pathway to AI fluency that's accessible and approachable to anyone, anywhere" — MIT News, 12 May 2026. MIT Learn — Universal AI →
Hardware · 01

CPU

A Central Processing Unit is the main processor inside every computer. It's built around a small number of very powerful cores — modern CPUs typically have 8 to 32. Each core can handle complex logic, branch decisions, and general-purpose tasks at extreme speed.

In AI, the CPU handles everything around the model: loading files, managing memory, running the operating system, and pre-processing data. But it isn't ideal for the actual AI calculations — those require doing millions of simple maths operations at the same time, which is where GPUs shine.

  • 8–32 powerful cores (typical desktop/server)
  • Sequential task specialist — one thing at a time, very fast
  • Runs the OS, loads data, manages AI infrastructure
  • Clock speeds around 3–5 GHz
C1
C2
C3
C4
ACTIVE
C6
C7
C8
C9

9-core CPU — each core handles complex tasks

4.2 GHz speed
9 cores
parallel tasks
Hardware · 02

GPU

A Graphics Processing Unit was originally designed to render video games — a task that requires updating millions of pixels at exactly the same time. Researchers discovered those same parallel processing abilities are perfect for AI training, where you need to multiply enormous matrices of numbers simultaneously.

Modern AI GPUs (like NVIDIA's H100) contain thousands of smaller, simpler cores. Each core is far less powerful than a CPU core, but together they can process AI calculations hundreds of times faster than a CPU alone.

  • Thousands of smaller cores (H100 has ~16,896)
  • Parallel processing — does thousands of calculations at once
  • Essential for training and running large AI models
  • AI models that take months on a CPU train in days on a GPU

400 cores — all firing in parallel

CPU (9 cores)

GPU (thousands of cores)

Data · 03

Token

AI language models don't read text the way humans do. They break everything into chunks called tokens. A token is roughly a word, a part of a word, or a punctuation mark — whatever the model's vocabulary system decides is useful.

Every time you type a message to an AI, it's converted into a list of tokens before the model sees it. Every word the AI writes back is also generated token by token. Token count is how AI companies measure usage — it covers both what you sent and what the AI replied.

  • ~1 token ≈ ¾ of a word in English
  • 100 tokens ≈ 75 words
  • "unbelievably" might be 3 tokens: "un" + "believ" + "ably"
  • Context windows are measured in tokens (e.g. 200,000)

Try this sentence broken into tokens:

"The neural network learns from data"

0 tokens
7 words
Knowledge · 04

Parameters

Parameters are the numbers that define what an AI model knows. During training, the model sees billions of examples and slowly adjusts these numbers until they encode the patterns of language, logic, and knowledge it was trained on.

You can think of each parameter as a tiny dial. Individually, one dial means nothing. But billions of dials — each fine-tuned through training — together create something that can write, reason, translate, code, and explain. When people say an AI has "7 billion parameters," they mean it contains 7 billion of these learned numbers.

  • Stored as floating-point numbers inside the model file
  • Adjusted during training via backpropagation
  • More parameters generally means more capacity to learn
  • GPT-4 estimated at ~1.8 trillion parameters
Input Output

Line thickness = parameter weight strength

GPT-2 (2019)
1.5B
GPT-3 (2020)
175B
PaLM (2022)
540B
GPT-4 (est.)
~1.8T

How they work together

When you send a message to an AI, all four pieces are in play simultaneously.

Step 1 — Tokenisation

Your text becomes tokens

The CPU receives your message and converts it into a list of tokens — numerical IDs the model understands. "Hello, how are you?" might become [15496, 11, 703, 389, 345, 30].

Step 2 — CPU Orchestration

The CPU manages the request

The CPU loads the token list into memory, coordinates with the GPU, and handles the surrounding infrastructure — network, memory management, security checks.

Step 3 — GPU Computation

Thousands of GPU cores crunch the maths

The GPU takes each token and runs it through the model — billions of multiplications and additions across the parameter matrix. This is the "thinking." The GPU can do many of these in parallel, which is why responses are fast.

Step 4 — Parameters activate

Billions of learned numbers shape the answer

As the GPU runs the computation, it passes data through the model's parameters — billions of learned numbers that encode the AI's knowledge of language, facts, and reasoning. These numbers determine what the next token should be.

Step 5 — Output tokens

Tokens become words, one by one

The model predicts the most likely next token. Then the next. Then the next — repeating the GPU computation for each one until the reply is complete. The CPU then assembles those token IDs back into readable text and sends it to you.