A reference site, not a course
AI, explained by level.
Nothing to sell you.
Search “learn AI” and almost every result is published by somebody with a course to sell or a tool to promote — the platforms' own catalogues, the labs' own academies. Some of it is free and some of it is good. The few results not selling anything are forum threads: disinterested, and unsourced. Nothing on that first page tells you whether any of it still holds. This site points at whatever is actually best, says when it's free, and tells you the date it was last checked so you know how much to trust it.
Published · Last verified · Next scheduled review:
Pick where you actually are
The usual mistake is starting in the wrong place. People who just want to use these tools well get pushed into machine-learning theory and quit; people who want to build get handed prompt tips. Three honest tracks:
-
Track 01
Curious
You've maybe typed a question into a chatbot. You don't know what a token is and you shouldn't have to yet. Goal: understand what this thing is, what it's bad at, and why people keep talking about it.
No technical background needed -
Track 02
Practical
You use AI most days and suspect you're using maybe 20% of it. Goal: get genuinely good at the tools in your actual work, and learn enough of the machinery to know why things fail when they fail.
Where most people are, and least served -
Track 03
Builder
You want to put a model inside something you're making. Goal: APIs, tool use, agents, retrieval, and the cost and reliability decisions nobody mentions until your bill arrives.
Some coding assumed
Or start with one of these
If you'd rather read something than place yourself first, these three are where to start: what to actually do differently, what these systems can't do, and how to tell whether anything you're reading still holds. Behind them sit checking an answer written for you, knowing whether a change actually helped, the glossary, the myths and where each one is answered, how to choose an assistant, what happens to what you type, and two written for people building things: designing for the ways it fails and why it does things nobody asked for. They are all listed on Guides.
-
Getting better results
Eight habits that measurably change what you get back, each traced to the reason it works. The page for the Practical track.
-
What AI is actually bad at
The limits as mechanisms rather than a task list, so it stays true as the tasks change.
-
Is what you're reading out of date?
This site's own method, turned outward: the tells, and a sixty-second primary-source check.
How this site works
Every page carries two dates, published and last verified, because AI writing rots faster than anything else in tech and matching dates mean a page hasn't been re-checked since it was written. How to read them. Every hard number lives on one page, Model facts, so one page rots instead of a hundred. Nothing here is an affiliate link, free resources come first, and any figure that couldn't be checked against a primary source is flagged in orange rather than filled in with something plausible-looking.
Who writes this, what it costs to run, and the one commercial connection it has: about this site. Every correction is logged and dated on the changes page.
The terms everyone drops without explaining
Short, plain explainers for the vocabulary that gets used as though you already know it. Written to actually teach, not to rank.
-
Context windows
Why a model “forgets” mid-conversation, and why bigger isn't automatically better.
-
Tokens
The unit you're actually billed in, and why the rule of thumb for turning words into tokens is right on some models and wrong on others.
-
Hallucination
Not a bug that will be patched. A property of how these systems work.
-
Training cutoff
Where the model's knowledge stops. Two dates, not one, and it doesn't know its own.
-
Tool use
The model is an author, not an actor. How text becomes actions, and what that risks.
-
Agents
A model in a loop with tools. The rest is marketing, and the loop is the hard part.
-
RAG and retrieval
Answering from your own documents, and why the search step decides everything.
-
Fine-tuning
Changes how a model behaves, not what it knows. The distinction costs people money.
-
Prompt caching
The biggest cost lever most builders have, and the easiest to break by accident.
-
Thinking and reasoning
What "thinking before answering" really is, and why the working shown isn't a confession.
Not sure which track? → Start here and answer three questions. · New to the vocabulary? → Glossary · The reference pages → Guides · Where the facts come from → Sources