What is AI?

Open a chat interface, type a question in plain English, and get a coherent answer in seconds. It feels like magic — and that feeling is exactly what makes AI tools easy to misuse. If you believe the output comes from something that understands, you stop checking it. If you know it comes from very sophisticated pattern matching, you stay appropriately sceptical.

This lesson is about building the right mental model before you touch the tools. A good mental model does not require any mathematics. It just requires being honest about what the system is actually doing.

AI as a field, not a product

"Artificial intelligence" is the name of a research field, not a single technology. For most of its history — from the 1950s through the early 2010s — AI was dominated by rule-based systems: programs where humans wrote explicit if-then rules that encoded expert knowledge. A chess engine, a spam filter built from keyword lists, a tax-filing wizard that walks you through fixed decision trees. These work well in narrow, well-defined domains. Add a rule for every case and you get a reliable system. The catch is that you have to anticipate every case yourself, and many real-world tasks have too many cases to enumerate.

Machine learning is a different approach: instead of humans writing the rules, the system infers them from examples. You feed the system thousands of labelled emails — spam and not-spam — and it discovers patterns that distinguish them. The result is not a list of rules you could read; it is a set of numerical weights that encode statistical regularities in the training data.

This distinction matters because it explains two key properties of modern AI:

1. It can handle tasks no human could fully specify. Recognising a cat in a photo involves so many edge cases that writing explicit rules is hopeless. Learning from millions of cat photos is tractable.

2. It can fail in ways no human would. If the training data had a bias, the model inherits it. If a situation falls outside the distribution of training examples, all bets are off — and the model has no way to know it is out of its depth.

Narrow AI versus general AI

Almost every AI system in production today is narrow AI: it does one thing, or a small cluster of related things, extremely well. An image classifier classifies images. A recommendation engine recommends content. A translation model translates. These systems do not generalise: a translation model cannot recommend movies even if both tasks deal with "language."

Artificial general intelligence (AGI) — a system that can reason across arbitrary domains the way humans do — remains an unsolved research problem. No system deployed today qualifies, regardless of what marketing says.

So where do large language models (LLMs) fit? They are still narrow in the technical sense: they predict text. But predicting text well enough and at large enough scale turns out to be surprisingly general in practice. A model trained on a broad slice of human writing picks up enough patterns to help with coding, drafting, explaining, translating, summarising, and reasoning through problems — not because it was explicitly trained on each task, but because all those tasks involve language. That is the practical sense in which LLMs feel general-purpose: they are a single tool you can point at many tasks.

Why this moment feels different

For most of AI's history the technology was useful only to specialists: you needed to know which model architecture to use, how to prepare training data, and how to deploy the result. The shift that happened around 2022–2023 was natural-language interfaces on top of powerful models. Now you direct the system in plain English. That dramatically lowers the floor.

It also raises the stakes for the learner, because the ease of use makes it tempting to skip understanding what is happening underneath. The rest of this track is about closing that gap: understanding the mechanism, knowing where it breaks, and building the habits that let you get consistent value from these tools without being blindsided by their failures.

Check your understanding

Where to go next

You now have a framework for thinking about what AI is and where it comes from. Next: how large language models work — the mechanism behind the text prediction that makes LLMs feel so capable, and why that same mechanism produces confident-sounding errors.