Vocabulary
Concepts
The terms people drop as though you already know them. Each one explained once, properly, with the practical consequence spelled out. Prices, limits and model names are kept off these pages and left to Model facts, so they stay correct as the models change.
Published · Last verified
Foundational#
The six that explain why these systems behave the way they do. If you use AI and something it did has ever puzzled you, the explanation is almost certainly on one of these pages. No coding required for any of them.
-
Context windows
Why a model seems to forget things mid-conversation, and why a bigger window is not automatically better.
-
Hallucination
Why confident invention is a property of how these systems work, not a bug awaiting a patch.
-
Training cutoff
The date the model's knowledge stops. There are actually two dates, and the model doesn't know its own.
-
Tokens
The unit you're billed in. Not words, not characters, and the rule of thumb everyone repeats is right on some models and wrong on others.
-
Thinking and reasoning
What's happening when a model "thinks before answering", when it's worth paying for, and why the visible reasoning isn't a confession.
-
Tool use
How a model that only produces text ends up searching the web and editing your files — and the risk that comes with it.
Building#
The four that only start to matter once you are putting a model inside something. They assume the foundational pages above, and they are not worth your time otherwise — a reader who wants to use AI well rather than build with it can stop at the six.
-
Agents
The most abused word in the field. A model in a loop with tools, and why the loop is where all the difficulty lives.
-
RAG and retrieval
Answering questions about documents the model never saw. Mostly a search problem, which is why most of it fails at the search.
-
Prompt caching
The largest cost lever most builders have, and the easiest one to break with a single careless line.
-
Fine-tuning
Teaches a model how to behave, not what is true. Why it is rarely the answer to "it doesn't know about my business".
What's next#
The original ten are written, so there is no queue to publish here, and a list of pages that don't exist yet isn't worth much anyway. New concepts get added when something is genuinely missing rather than to hit a number, and each one is dated and recorded in the changes log when it lands. The more useful ongoing work is keeping these correct, which is the harder half.
Something that should be here and isn't? That counts as a gap worth reporting, same as a wrong fact.
Numbers referenced by these pages → Model facts · Quick definitions → Glossary · The limits → What AI is actually bad at · The method → Is what you're reading out of date?