Under the Hood
This trained "engine" — like Claude or GPT — is the underlying system doing the predicting and generating.
What is an AI model?
Roughly four characters each, these text chunks are what AI reads, generates, and gets billed by.
What are tokens?
Unlike a chatbot that only answers, this type of AI takes actions on its own to accomplish a goal.
What is an agent?
When an AI "forgets" the beginning of a long chat, it's hit the edge of this working memory, measured in tokens.
What is the context window?
The most important AI limitation to know: confidently stating something false or made-up
What is a hallucination?
Trained on massive amounts of text, this is the most common type of AI model — every chatbot is powered by one.
What is an LLM / Large Language Model?
An AI that can handle images, audio, or video — not just text — earns this label.
What is multimodal?
Web search, running code, checking a calendar — these external capabilities let an AI actually do things, not just describe them.
What are tools?
Tying an AI's answers to verified source material — a response backed by real documents is said to be "well" this.
What is grounded / grounding?
The rules and safety limits built around an AI to keep it on-topic, safe, and appropriate.
What are guardrails?
Training is teaching the model; this is the model actually running to produce answers — and why businesses talk about its "costs."
What is inference?
This three-letter interface is how software connects to a model programmatically, rather than through a chat window. HINT: Data delivery coming for Goodwill, WGU, maybe others?
What is an API?
Like a manager delegating to team members, a main agent hands pieces of a big task to these specialized helpers.
What are sub-agents?
This acronym describes fetching relevant documents before answering, so responses come from real sources instead of memory alone
What is RAG / Retrieval-Augmented Generation?
A standardized test for comparing models — topping the "leaderboard" means scoring high on these
What is a benchmark?
A model described as "70 billion" of these is referring to its internal values — a rough measure of size and capability.
What are parameters?
The hypothetical milestone where AI matches or exceeds human ability across most tasks — the "G" stands for "general."
What is AGI?
A packaged, reusable "how-to" that teaches an AI to handle a specific task well, like following a company process
What is a skill?
Unlike a regular database that matches keywords, this special kind lets an AI search information by meaning.
What is a vector database?
Crafting prompts designed to trick an AI into ignoring its safety rules — named after bypassing a phone's restrictions.
What is jailbreaking?
A step beyond prompting: further training an existing model on specialized data to customize it for a specific company or purpose.
What is fine-tuning?
Models like Llama that publish their internals for anyone to download and run are called this, versus proprietary "closed" models.
What are open-source / open-weight models?
This three-letter protocol is the standard way to connect an AI to outside apps like email, calendars, and company files.
What is MCP / Model Context Protocol?
Converting text into a list of numbers that captures its meaning — the machinery that makes vector search possible.
What is an embedding?
When a model's outputs skew unfairly toward or against certain groups because of patterns in its training data. HINT: "We're the best product, but I'm ****, I work for Jobscan."
What is bias?