For the Curious, the Confused and the Computationally Lost.

You were holding this phrase before you opened the guide, so it goes first. Artificial means made by people rather than grown: an artificial lake, an artificial hip. Intelligence means learning, understanding and reasoning, and putting what you know to use. Put together, the phrase claims a made thing that does what intelligence does. It first appears in a proposal dated August 1955 for a summer workshop at Dartmouth College the following year, and the proposal said plainly that it rested on a conjecture: that every feature of learning or intelligence could in principle be described precisely enough for a machine to imitate it. That conjecture is still the name of a field of study and the hope behind it. It has never been a description of any particular program.

There is still no agreed definition of what a program must do to earn the phrase. The working one most people use is a program that does something we used to think needed a person, and its known flaw is that the boundary moves. Once a program does a thing reliably, correcting your spelling, finding a route, playing chess, people stop calling it AI and start calling it software, and the phrase slides on to whatever is still surprising. So "AI" on a box is a label, not a test. Usually it means that a model which learned from examples is somewhere inside; sometimes it means less than that. It does not tell you which kind, how good, or whether the thing understands anything, and whether any program understands is a philosophical argument this guide names and leaves alone.

The two letters name three different things, and most of the confusion about AI comes from not noticing which. The first is the field of study, with its conferences and journals, since 1956. The second is a label on products. "An AI", as people say it, is nearly always a product: a website, an app or a feature with a model somewhere in it, and this guide calls those products by what they are, a chat product, a helper in a document editor, a coding agent. The third is the model, and it is the sense this guide uses from the first tier on. A model, in the plain sense, is a stand-in built to answer for something real, a model of the weather or of a bridge. A model here is a file of numbers learned from examples, and it stands in for the patterns in those examples; a program reads the file and runs it. When someone says "the AI got it wrong", they mean the product. When a page says "this model has 27 billion parameters", they mean the file.

Two more words you will hear on the news. Narrow AI is a system that does one task: it sorts mail, or it writes text, or it drives. General AI, or AGI, means a system that could do any intellectual task a person can. No such system is agreed to exist, whether the current wave is on the way to one is contested, and there is no agreed test that would settle it; the Vector tier says the same at 2.7, when the word turns up on a slide.

This is a map of AI vocabulary, not a course and not a dictionary. Learn it in order when you are new; look things up when you are not. You do not need programming, maths or machine learning before starting; each tier assumes only the ones before it.

• If you have used a chat product and want to know what is underneath it, or if the words currently sound like alphabet soup, start at Origin and read forward.
• If you know the basics and have met a term you do not recognise, use The Index on The Reference page; every entry names the tier and section that places it.
• If you are choosing something to download, start at Nexus.
• If you are reading a release post, a paper or a vendor deck, start at Apex.
• Boxes marked "New here" stop deliberately before the mechanism and name the subject that carries on from there.
• Boxes marked "Same word, different thing" flag a term that means something else one family over.
• Boxes marked "You have met this" tie a word to something you have already done in a product.

Every family section in Nexus follows one template, in this order: how it works, the words, the anchors, what runs it, memory floor, the licence trap, what people get wrong. Skip to the part you need.

What ages and what does not. The vocabulary, the mechanisms and the traps are the durable part. The anchor tables are a dated snapshot: the models named will be superseded, and the tables exist to show what a card looks like when read properly, not to recommend. Prices, provider-specific limits and per-model rules are left out on purpose; where a vendor page is the source, the guide points at it rather than copying numbers that will be wrong by the time you read them.

• Tier 01, The Origin: the AI you met today; where you meet one; what a model is; what "AI" means on the box; the map, first drawing; the pieces it reads; what it does when it runs; how it learned; what it runs on; what you now know; the question it leaves you with
• Tier 02, The Vector: reading a filename; inference; tokens; thinking models; parameters; dense versus MoE; multimodal; precision and quantisation; the files in the folder; checkpoints and LoRA; open weights versus open source; same word, different thing; where to get them; the families, second column; the question it leaves you with
• Tier 03, The Nexus: language, and its dials; image; video; speech and audio; vision understanding; embeddings and rerankers; classical machine learning; the families, third column; the question it leaves you with
• Tier 04, The Apex: how models are made; architecture words; what someone did to the base; world and robot models; what gets bolted on; words on the invoice; evaluation and what goes wrong; reading the table; whose names these are; the families, final column; where to go next
• The Reference: the index; the sources; corrections

---