Tokens, context windows, agents, RAG, hallucinations — what they actually mean in plain English, and why each one matters to you.
Grouped by topic so you can skim what you need — or search when you hear something in a meeting and don't want to ask. No textbook tone, just what each word means and why it matters.
The words everyone uses before anything else makes sense.
Software that does things we used to think required a human — understanding language, recognizing images, making decisions. Today it usually means systems that learn patterns from huge amounts of data rather than following hand-written rules.
The engine behind ChatGPT, Claude, and Gemini. It’s a system trained on enormous amounts of text that predicts what words should come next — which, at scale, lets it answer questions, write, and reason.
When you chat with ChatGPT, you’re talking to an LLM.
A single trained “brain” with a name and version — like Claude Opus or GPT-5. Newer or bigger models are generally more capable (and usually cost more to run).
Whatever you type or say to the AI — your question, instruction, or request. The quality of your prompt largely decides the quality of the answer.
The small chunks of text an AI reads and writes in — roughly three-quarters of a word each. AI usage and pricing are measured in tokens, not words.
“Piazza Consulting” is about 4 tokens.
How much text the model can hold in mind at once — your message, the documents you paste, and the conversation so far. Go past it and the earliest parts start dropping off.
Why a very long chat can “forget” what you said at the start.
The mountain of text and images a model learned from. It shapes what the model knows — and it has a cutoff date, which is why a model may not know recent events unless it can search.
The internal dials a model adjusts as it learns — often billions of them. More parameters can mean more capability, but it’s not the only thing that matters.
What actually happens when you hit enter — and how models behave.
The act of actually running the model to get an answer. “Training” is teaching the model once; “inference” is using it, every time you hit enter.
Hidden background instructions that set the AI’s role and rules before you ever type — its “job description” for the conversation.
A support bot has a system prompt telling it to stay polite and only discuss that company’s products.
A setting for how predictable vs. creative the output is. Low temperature = safe and consistent; high = more varied and surprising.
A model that handles more than text — images, audio, sometimes video. It’s why you can show an AI a photo and ask what’s in it.
Taking a general model and training it a little more on your specific data or style, so it specializes — like sending a smart generalist to a short, focused course.
A newer kind of model that “thinks” before answering — working through steps internally. Slower, but much better at hard logic, math, and multi-step problems.
Getting a model to show or work through its reasoning step by step instead of blurting a final answer. It reliably improves accuracy on tricky questions.
Adding “think it through step by step” to a prompt.
A standardized test used to compare models on things like coding or reasoning. Useful, but real-world performance can differ from the leaderboard.
Agents, RAG, MCP — the terms dominating current conversations.
AI that doesn’t just chat — it takes actions. It can use tools, browse, run code, and complete multi-step tasks on its own, checking its work as it goes. The big shift of 2025–2026.
“Book the cheapest flight and add it to my calendar” — and it actually does it.
A setup that lets a model pull answers from your documents instead of only its training. It looks up the relevant material first, then answers from it — which cuts down on made-up answers.
A company chatbot that answers from your internal handbook.
A way of turning text into numbers that capture meaning, so a computer can tell that “car” and “automobile” are close. The quiet machinery behind smart search and RAG.
A specialized store for those meaning-numbers, so an AI can instantly find the most relevant passages across thousands of documents.
A common standard that lets AI assistants plug into other tools and data sources — your calendar, your files, a database — like a universal adapter for connecting AI to the things you use.
Running AI on your own computers instead of a vendor’s cloud, so sensitive data never leaves your network. The serious answer for regulated industries.
A hospital running AI in-house so patient data stays put.
“Open” models can be downloaded and run yourself; “closed” ones are used through a company’s service. Open gives control and privacy; closed often gives top performance with less hassle.
The specialized chip that makes AI fast. Originally built for video games, now the workhorse of training and running large models — and the reason AI infrastructure is expensive.
Know these before you trust output blindly or paste sensitive data.
When an AI states something false as if it were fact — a fake citation, a wrong number, an invented quote — delivered with total confidence. The single most important thing to watch for.
Why you verify any name, figure, or legal citation it gives you.
A security attack where hidden instructions buried in a webpage, email, or document trick an AI into doing something it shouldn’t — like leaking data. A real risk as agents start reading the open web.
Trying to phrase a prompt so the AI ignores its own safety rules. Providers work hard to block it, and for any serious or regulated use, relying on jailbreaks is a liability, not a feature.
When a model’s answers reflect skews in its training data — favoring some groups, views, or phrasings over others. Worth keeping in mind anywhere fairness matters.
The rules and filters that keep an AI from producing harmful or off-limits output. Necessary, imperfect, and a big part of what separates a toy from an enterprise tool.
Employees using AI tools their employer hasn’t approved — usually free, personal accounts — often with company data. Widespread, invisible to IT, and the leading source of AI-related data leaks.
Pasting a client contract into a personal ChatGPT account “just to summarize it.”
For anyone wiring AI into products, workflows, or internal tools.
A way for software to talk to an AI model directly, instead of a person typing in a chat box. It’s how companies build AI into their own apps and workflows.
The craft of writing instructions that consistently get good results from a model. Less about magic words, more about being clear and giving the right context.
The next step up — deciding what information to put in front of a model at each moment so it has exactly what it needs and nothing that distracts it. Crucial for building reliable agents.
Since usage is billed per token, longer prompts and longer answers cost more. Trimming unnecessary context is how teams keep AI bills under control.
Pass the 10-question test and earn the Certified AI Fluent credential from Piazza Consulting Group.
