Vocabulary: How Models Work

Each entry: the word, what it means, where it lives, and a sentence you can say in the room. Drill by listening, then test yourself: hear the term, say the definition before the narration does.

LLM (large language model) — a neural network trained on huge amounts of text to predict the next token; all of its "knowledge" is stored implicitly in its weights. Lives in: the foundation — everything else wraps around this. Say it: "An LLM generates; it doesn't look anything up — retrieval has to be added around it."

Parameters / weights — the billions of numbers inside the model that training adjusts; where learned patterns and facts actually reside. Lives in: model internals. Say it: "Facts fine-tuned into weights go stale — knowledge that changes belongs in an index, not in weights."

Token — the unit of text a model reads and writes: usually a word fragment. Pricing, context limits, and speed are all measured in tokens. Lives in: model internals; also the billing unit. Say it: "We pay per token, and output tokens cost more than input — so concise outputs are a cost lever."

Training vs. inference — training is the one-time process of adjusting weights on data; inference is every subsequent use of the frozen model to generate answers. Lives in: the model lifecycle. Say it: "Nothing we do at inference changes the model — it learns nothing from our prompts."

Next-token prediction — the single operation an LLM performs: given the sequence so far, produce the most plausible next token, repeatedly. Lives in: model internals — the mechanism behind every answer. Say it: "Even a tool call is just next-token prediction aimed at a schema."

Context window — the model's entire working memory: the maximum tokens it can consider at once — system prompt, conversation, and any documents included. Lives in: prompt assembly; the constraint RAG and summarization exist to manage. Say it: "The context window is the only thing the model knows for sure — engineering is deciding what goes in it."

Temperature — the sampling setting controlling how adventurously the model picks among plausible next tokens: low = deterministic and precise, high = varied and creative. Lives in: inference configuration. Say it: "For extraction and grounded answers we run low temperature — we want repeatable, not creative."

Hallucination — a fluent, confident output that is wrong; structural, because the model optimizes plausibility, not truth, and has no built-in fact-checker. Lives in: model behavior; the risk the whole grounding toolkit addresses. Say it: "Hallucination can't be fixed inside the model — truth is added from outside, with grounding and verification."

System prompt — the standing instruction block the developer controls: role, rules, tone, constraints — prepended invisibly to every conversation. Lives in: prompt assembly; the first and cheapest adaptation tool. Say it: "Behavior rules live in the system prompt; users shouldn't be able to override them."

Zero-shot / few-shot — zero-shot: the model performs a task from instructions alone. Few-shot: you include worked input-output examples in the prompt, and the model pattern-matches onto them. Lives in: prompt engineering. Say it: "When instructions alone underperform, we add few-shot examples before we ever consider fine-tuning."

Knowledge cutoff — the date the model's training data ends; it knows nothing after it except what you put in the context. Lives in: model limitations; another reason retrieval exists. Say it: "Anything after the cutoff — and anything internal to us — reaches the model only through retrieval."

Multimodal — a model that handles more than text: images, audio, documents in, or speech and images out. Lives in: model capabilities. Say it: "Transcript intake is a multimodal problem at the edges — scanned images in, structured data out."

Prompt engineering — designing the instructions, structure, examples, and context that reliably get the behavior you want from a model. Lives in: the development discipline around prompts. Say it: "We treat prompts as engineering artifacts — versioned, reviewed, and tested like code."