Words about how the model works
- Hallucination
- When a model produces fluent, confident text that is not grounded in reality — invented citations, made-up names, plausible-sounding but wrong dates. Not lying; the model has no concept of lying. It's pattern-completing in a region where the patterns it learned don't have a true answer.
- Chain of thought
- Asking a model to think step-by-step on a hard problem. Improves accuracy because it routes a single hard prediction through a series of easier ones.
Mechanism words you'll see in the wild (optional reference)
You do not need to know these to direct AI well, and the lessons don't require them. They turn up in news articles and other tutorials, so we define them here for when you encounter them.
- Neural network
- A kind of computer program loosely inspired by how neurons connect in the brain. For directing AI, what matters isn't how it's built — it's what it does: text in, text out, with billions of internal numbers in between. Every chatbot you've used has one underneath.
- Token
- A chunk of text — sometimes a whole word, often a piece of one, sometimes a single space or punctuation mark. When you read that a model has “a 200,000-token context window,” that's roughly 600 pages of text.
- Parameter
- One of the billions of internal numbers a model uses to compute its predictions. Training tuned all of them. The number of parameters is roughly how big a model is — more parameters generally means more capable, but slower and more expensive to run.
Words about the agent
- Agent
- A model + a loop + tools + state. An AI system that runs in cycles, calls tools, and keeps notes between turns. Different from a chatbot.
- Loop
- Code that runs the model more than once on a single task. After each model response, the loop decides whether to stop or run the model again with new information.
- Tool
- A specific action the model is allowed to take — searching the web, reading a file, sending an email, running a command. Sometimes called a "function." Without tools, the model can only produce text.
- State
- Everything the agent has learned, decided, or read so far during this task. The agent's working memory. Lives inside the context window during the task.
- Context window
- The maximum amount of text the model can see on any single turn — the original task, prior reasoning, tool results, all of it together. Different models can hold different amounts, from about 6 pages on a small model to about 600 pages on a large one. When a conversation grows past the window, the oldest content silently drops.
- Trace
- The visible record of what an agent did, turn by turn — what it reasoned, what tools it called, what the tools returned. Your main diagnostic instrument when something goes wrong.
- Termination
- How a loop ends. Four ways: task complete, turn budget exhausted, unrecoverable error, or user stop.
Words about directing
- Director
- Someone who works AI tools by understanding the parts (model, loop, tools, state) and diagnosing failures by part. Contrast with user — someone who treats the system as a single black box.
- Scope
- What a task includes and excludes. A well-scoped task has boundaries you could draw on a whiteboard.
- Grounding
- The real-world sources an agent consults before answering — files, searches, APIs, databases. An ungrounded answer comes from what the model learned during training; a grounded answer comes from a tool that pulls a fresh source.
- Tool surface
- The full set of tools available to an agent on a task. A small tool surface is predictable; a large one is powerful but has more ways to go sideways.
- Handoff
- The point where one agent, one turn, or one task passes its output to another. Handoffs are where state most often gets lost.
- The four director questions
- Model, loop-framing, tools, state. Asked in order before touching a tool. The whole course, compressed to one page.
Words about your computer (you'll need these for the activities)
- Working folder
- A folder on your computer where everything you build for this course lives. Make one now if you don't have one — Desktop is fine. Anything simple works as a name; AI Architect Academy is the example we use throughout.
- .md file
- A plain-text file in a format called Markdown. Just text with a few small punctuation conventions for headings, bold, and lists. Create with Notepad (Windows) or TextEdit-set-to-plain-text (Mac). Don't use Microsoft Word — it adds invisible formatting that breaks Markdown.
- Path
- A string of folder names that points to a file on your computer. On Mac and Linux, paths use forward slashes (capstone/my-first-loop-v1.md). On Windows, paths sometimes show with backslashes (capstone\my-first-loop-v1.md). Both refer to the same file.
Words you'll see in the rest of the course
- API key
- A long secret string that lets a program (your agent) call a paid AI service on your account. Module 2 covers these as an optional advanced path; most students don't need one because the Pro subscription and the desktop app cover the bulk of the course.
- Local model / cloud model
- A local model runs on your own machine — private, free per call, slower. A cloud model runs on a company's servers — faster and more capable, but you pay per call and your data leaves your machine.
- Recipe
- In this course, a tool-specific step-by-step section that's dated and refreshed every quarter. Spotted by the "Recipe" header and the verification date in a small box at the top. Compare to concept content (the durable kind) — most of what you read is concept; recipes are a smaller, faster-changing layer on top.