AI is on the news, on the box your phone came in and in the chat window at work. It is also in the spam filter and the card-fraud check you have used for years without anyone calling them AI. The words around it are everywhere too: model, token, training, agent, open weights, AGI. The vocabulary is big. Really big. It grows every week, and the explanations, when they come at all, assume you already knew last week's.
Most of it is less mysterious than it sounds. The two letters name three different things: a field of study, a label on products, and the model inside them, which is a large file of learned numbers that another program reads and runs. Most of the confusion comes from not noticing which of the three someone means. Once you can hear the difference, most of the noise falls away, and what is left is a translation problem.
So I wrote the words down, in order, from the bottom.
What it is
A Hitchhiker's Guide to AI Galaxy, for the Curious, the Confused and the Computationally Lost. It is a map of AI vocabulary, not a course and not a dictionary: read it in order when you are new, look things up when you are not. You do not need programming, maths or machine learning to start.
It climbs four tiers, and each one assumes only the ones before it:
- Origin is where to start if AI is something you use every day but could not explain to a friend: what it is, where you already meet it, and what actually happens when you press enter.
- Vector is the look under the bonnet: what a model really is, what the letters and numbers in its name are telling you, and why "open" does not always mean open.
- Nexus is for choosing: the kinds of AI side by side, what each one is good at, what it needs to run, and the small print that catches people out.
- Apex is for reading the headlines and the launch posts without reaching for a dictionary: how models are made and tested, what an agent actually is, and why a benchmark score is never the whole story.
It stops there on purpose. It is not written for experts, or for anyone who intends to become one. If that is you, you are holding the wrong towel.
What is in it
The basics: what "artificial intelligence" actually claims and where the phrase came from, narrow AI and AGI, the AI you already use without choosing it, what happens when you type into a chat window and where your words go, how a model learned what it knows, and what it runs on.
The words: what a model, a token and a parameter are, what "thinking" and "multimodal" mean, what open weights are and whether they are open source, and which words mean two different things depending on who is saying them.
The kinds: AI that writes, draws, makes video, speaks, listens and sees, the search behind "find me something like this", and the older machine learning that has been quietly sorting and predicting for years. For each, what it does, how it works in a few sentences, and the one thing people get wrong about it.
The bigger picture: how a model is made and tested, what agents, RAG, MCP and prompt engineering are, what a world model is, why a benchmark table is harder to read than it looks, and whose names keep turning up.
The map: about forty figures. Two of them travel with you: a picture of what goes in and what comes out, drawn once at the start, and a table of the kinds that is redrawn at each tier with more in it, until by the end it is a chart of the whole territory. At the back there is an A to Z of over 160 terms, each pointing at the section that places it.
Why it exists
An answer is no use if you never understood the question. Most questions about AI have that problem: they are asked in words nobody explained. The vocabulary is the wall. Most people who give up on this subject do not give up on the ideas; they give up on the third unexplained word in one sentence. The guide gives each of those words a place. Where a word leads somewhere deeper than the guide goes, a box says so and names the subject to look up next.
Every number in it was checked against a source on 20 September 2026, and the sources are listed at the end with links, nearly two hundred of them. Where I could find only the vendor's own claim, the text says "reported". Where I could not find a source, the number is not there. The models it names will be superseded, probably soon. The words will last longer, and the words are the point.
It reads best on a computer. The figures are drawn wide, and on a phone they fold into a column.
Don't panic. Go and have a look.
