Tau
How AI works under the hood · 12 lessons

How does AI work? A plain guide to what's under the hood

See what happens behind the chat box: how AI learns, why it forgets or makes things up, and what each question costs. For curious people who want the why, without the maths.

Who it’s for: Curious users, students and professionals who want the mechanics.

You don't need to understand an engine to drive a car. But knowing a little about how AI works helps you predict when it will go wrong. This course looks behind the chat box, in plain English and without the maths. It's for curious users, students and professionals who want to know why AI answers the way it does.

You'll start with tokens, the small chunks of text AI reads and writes. In English, a token is about four characters long, and limits and prices are counted in tokens. Then you'll see how a chatbot is made. It's trained on a huge collection of text, fine-tuned on examples and shaped by feedback. That explains why it doesn't know last week's news, and why it sometimes needs a web search.

Next come the questions people ask most. Why does AI forget what I said earlier? That's the context window, the amount of text it can hold in mind at once. Why do I get different answers each time? Why does it make things up so confidently? You'll also learn how AI can search your own documents, and how much energy a single question uses.

By the end you’ll be able to…

  • Estimate how many tokens a document uses and why that matters
  • Explain pre-training, fine-tuning and human feedback in a line each
  • Predict when a long chat will outgrow its context window
  • Explain how RAG grounds an answer in your own documents

Short on time? Start with the essentials

These 4 lessons cover the most useful ideas in this course. Each one takes about a minute.

Essential 1

How does AI read my words?

Explain what tokens are and roughly estimate a token count (about four characters each in English).

Essential 2

How is a chatbot made?

Describe pre-training on huge text collections, fine-tuning on examples and learning from human and AI feedback.

Essential 3

Why doesn't AI know the latest news?

Explain training cutoffs and when web search or uploaded documents are needed for current facts.

Essential 4

Why does AI forget what I said earlier?

Explain the context window and spot signs a long chat is losing early details, then restart with a summary.

Every lesson in this course

This course has 12 video lessons of about a minute each, from the basics to expert habits. Each one ends with a question, three takeaways and its sources.

Foundations 3 lessons

  1. How does AI read my words?Essential

    Explain what tokens are and roughly estimate a token count (about four characters each in English).

    “AI doesn't read words the way we do. It reads pieces called tokens.”

  2. How is a chatbot made?Essential

    Describe pre-training on huge text collections, fine-tuning on examples and learning from human and AI feedback.

    “Chatbots like ChatGPT are made in three stages. Here's what each one does.”

  3. Why doesn't AI know the latest news?Essential

    Explain training cutoffs and when web search or uploaded documents are needed for current facts.

    “Ask AI about yesterday's match. Without web search, it's probably guessing.”

Building skills 4 lessons

  1. Why does AI forget what I said earlier?Essential

    Explain the context window and spot signs a long chat is losing early details, then restart with a summary.

    “Ever had AI forget something you said earlier? There's a reason.”

  2. Why do I get different answers each time?

    Explain sampling and temperature, and choose low or high randomness for factual versus creative tasks.

    “Ask the same question twice and you'll often get two different answers.”

  3. What are 'thinking' models doing?

    Explain how reasoning models spend extra computation on hidden working, and when that's worth the wait and cost.

    “Some AI models pause and think before they answer.”

  4. How does AI see my photos?

    Explain how images and audio are turned into tokens the same model can reason over alongside text.

    “When you send AI a photo, it often cuts it into tiny squares.”

Advanced 3 lessons

  1. How does AI search my documents?

    Explain embeddings as a map of meaning and how retrieval-augmented generation pulls in relevant passages before answering.

    “To AI, “car” and “automobile” sit right next to each other.”

  2. Why does AI make things up?

    Explain how next-word prediction plus scoring that rewards guessing produces confident errors, and what reduces them.

    “AI makes things up partly for the same reason students guess.”

  3. How much energy does one question use?

    Compare the energy of a typical text prompt with everyday activities and explain why reasoning, images and video cost more.

    “A typical AI text question uses less energy than nine seconds of TV.”

Expert 2 lessons

  1. AI keeps getting it wrong: which fix?

    Decide whether a problem needs better prompting, retrieval from documents or fine-tuning, based on cost and what's wrong.

    “Your AI keeps getting your product names wrong.”

  2. What does a model's fact sheet tell me?

    Use a model card to find context window, training cutoff, benchmarks, known limits and safety evaluations.

    “Most major AI models come with a fact sheet, often called a model card.”

Apps we recommend

These are the tools this course uses, and what each one is best for. We have no partnership with any of them.

OpenAI Tokenizer

Best for: Seeing how text is split into tokens

Paste any text and see it split into colour-coded tokens with a live count.

Free

Transformer Explainer

Best for: Watching a transformer predict the next word

Interactive visual of a real GPT-2 model in your browser, with a temperature slider.

Free

Google AI Studio

Best for: Experimenting with temperature and models

Change temperature, compare models and see token counts on your own prompts.

Freemium
Hands-on project

Explain the AI you use, under the hood

Pick one assistant you use often. Measure a real document in a tokenizer, find the model's context window and training cutoff in its official docs, ask one question three times, and test two failures you predict. Write it up for a friend or colleague.

What you’ll make: A one-page explainer of how your assistant works, with measured numbers and two tested predictions.

Check your own work

  • A real document's token count is measured and compared with the 4-characters-per-token rule
  • Context window and training cutoff are quoted from official docs or a model card, with a link
  • The same question asked 3 times, with the differences recorded and explained
  • Two predicted failures were tested, and each result is recorded
  • States whether the assistant used web search or your files for current facts, with evidence

Further reading and viewing

We’ve chosen these articles, research papers, videos and guides from other trusted sources.

Video · 3Blue1Brown · 8 min

Large Language Models explained briefly

Eight-minute visual primer on how LLMs predict text and how they are trained.

Video · Andrej Karpathy · 4 hours

Deep Dive into LLMs like ChatGPT

Full, accessible walkthrough of how chatbots are built, from data to reasoning.

Research · OpenAI · 15 min

Why language models hallucinate

Research summary arguing training and scoring reward guessing over saying 'I don't know'.

Research · International Energy Agency · 30 min

Energy and AI

Authoritative data on data-centre electricity use and where AI demand is heading.

Article · Google Cloud · 8 min

Measuring the environmental impact of AI inference

Google's measured energy, water and emissions per median Gemini text prompt.

Questions about this course

How do large language models work?

A large language model, or LLM, is the kind of AI behind chatbots like ChatGPT. It learned from a huge amount of text, and it writes by repeatedly predicting a likely next word, or more exactly the next token. It doesn't look facts up unless it's connected to search or your files. That's why its answers can sound fluent and still be wrong.

Why does AI make things up?

Because it's built to produce a likely-sounding answer, not to say 'I don't know'. Training has also tended to reward a confident guess over admitting uncertainty. So when it lacks the facts, it can fill the gap with something plausible, like a fake quote or source. People call this a hallucination. Grounding it in real documents or search, and checking sources, reduces the risk.

What is a context window?

The context window is how much text an AI can take into account at once, including your messages, its replies and any files. In a very long chat, early details can fall out of view or get less attention. If an assistant starts forgetting what you told it, ask it for a summary and start a fresh chat with that summary.

What is RAG in AI?

RAG stands for retrieval-augmented generation. Before answering, the AI searches a set of documents, such as a company handbook, and pulls in the most relevant passages. It then writes its answer from those passages. That makes answers better grounded and lets it show where they came from. It still isn't perfect, so check the quoted passages for anything important.

Do I need maths or coding for this course?

No. The course explains each idea with everyday examples, not equations or code. Each lesson is a short video of about a minute, followed by one question. Lessons run from foundations to expert level, so you can stop when you've learned enough or keep going further. Your first five lessons in Tau are free.

Get early access to this course

Tau opens soon. Join the waitlist and we’ll email you as soon as you can start this course.