AI fundamentals: what it is, how it works, where it fails

Understand how modern AI is built and why it fails, so you can judge what to hand to it and what to keep.

You probably use an AI assistant already, or your boss has told you to start. What nobody explains is how the thing works, so you cannot tell when to trust it. It writes a good email, then invents a regulation that does not exist, in the same confident tone. This course explains the machinery in plain language: how models learn from data, how a language model writes one token at a time, why it forgets, why it makes things up and what an agent actually does. There is no maths and no coding in it. You leave able to predict where AI will help you and where it will let you down.

What you'll be able to do

• Explain the difference between AI, machine learning, deep learning and generative AI using an example from your own work
• Describe how a model learns from labelled examples and why the quality of its data limits what it can do
• Explain how a language model produces text one token at a time and why the same question can get different answers
• Describe pretraining, fine-tuning and human feedback, and how each one shapes the assistant you use
• Predict when a long conversation or document will push past what an assistant can hold in view
• Check an AI answer for invented facts, sources and numbers using a short, repeatable routine
• Build a capability map of your own work tasks showing what to hand to AI, what to share with it and what to keep

Syllabus

Module 1: Know what people mean when they say AI

Separate AI, machine learning, deep learning and generative AI, and use the difference to read product claims and predict how a tool will fail.

• 1.1 AI is a family of techniques, not one thing
• 1.2 Why rules gave way to learning
• 1.3 The AI you already use without calling it AI
• 1.4 Sort the tools in your working week

Module 2: See how a model learns from data

Understand training as a process of adjusting millions of settings to reduce error on examples, and why the data a model learns from sets the limits of what it can do.

• 2.1 A model is a function with millions of dials
• 2.2 Learning means guessing, measuring the miss and adjusting
• 2.3 A model can only be as good as its examples
• 2.4 Train a tiny image classifier yourself

Module 3: Understand how a language model writes

Follow how a large language model turns your words into tokens, predicts the next token from probabilities, and samples its way to an answer, and what that explains about its behaviour.

• 3.1 Your words become tokens before the model sees them
• 3.2 It writes by predicting the next token, again and again
• 3.3 Why the same question gets a different answer
• 3.4 Watch prediction at work in an assistant

Module 4: Trace how a chatbot is made

Follow the three stages that turn raw text into a helpful assistant: pretraining, fine-tuning and human feedback, and see which of the assistant's habits each stage creates.

• 4.1 Pretraining: learning to predict text at huge scale
• 4.2 Fine-tuning turns a predictor into an assistant
• 4.3 Human feedback and why assistants sound so agreeable
• 4.4 Find the edges of what a model knows

Module 5: Work within what the model can hold in view

Understand the context window as the model's only working memory, why long chats and big files degrade answers, and how retrieval lets an assistant use information it was never trained on.

• 5.1 The context window is the model's only working memory
• 5.2 Why long chats drift and big files get skimmed
• 5.3 Retrieval: how an assistant looks things up
• 5.4 Test the limits with a long document

Module 6: Catch hallucinations before they cost you

Understand why language models produce confident false statements, where that is most likely, and a short checking routine you can run on any answer before you rely on it.

• 6.1 Why a fluent answer can still be false
• 6.2 Where invented answers are most likely
• 6.3 A five-minute routine for checking any answer
• 6.4 Fact-check a real answer from start to finish

Module 7: Use models that see, hear and act

Understand how image, audio and multimodal models work, and how tool use lets a model search, calculate and take actions as an agent, with the new kinds of failure that brings.

• 7.1 How AI makes and reads images
• 7.2 Speech, voice and models that handle everything at once
• 7.3 Tool use: when a model asks software to act
• 7.4 What an agent is and where it goes wrong
• 7.5 Test a multimodal assistant on real material

Module 8: Decide what to hand to AI and what to keep

Turn everything in the course into a capability map of your own work: tasks AI can do with a light check, tasks to do together, and tasks to keep, with reasons for each.

• 8.1 AI ability is uneven, and the edge is hard to see
• 8.2 Three questions that sort any task
• 8.3 Hand over, work together, or keep
• 8.4 Build your personal capability map

Frequently asked questions

How long does the course take?

About eight and a half hours across eight modules, including the exercises. Most people finish in three to four weeks at two or three hours a week.

Do I need a technical background or any maths?

No. Every idea is explained with everyday examples, and the exercises use free tools in a browser. If you can use email and a spreadsheet, you can follow it.

Which AI assistant do I need?

Any mainstream assistant works, such as ChatGPT, Claude, Gemini or Copilot, and the free version of most is enough. The course teaches how these tools work in general, so it stays useful when products change.

How is this different from Working with AI assistants?

This course explains how AI works and where it fails. Working with AI assistants: prompting that gets usable output teaches the practical skill of directing an assistant. Take this one first so the prompting techniques make sense.

Does it cover privacy, deepfakes and scams?

Only briefly. Those risks are taught in depth in AI risks: privacy, bias, deepfakes and scams.

Explore more

• All Academy courses