Glossary · AI Tools Field Guide

Open edition · no CAC needed

Learn more

The tools

Term	One line
ChatGPT / Gemini / Grok / Claude	The four general chat assistants this guide uses, from OpenAI, Google, xAI and Anthropic, each sold as a free or personal consumer plan and as a business plan.
Anthropic	The company that makes Claude, and sells it as the claude.ai app, a business plan and an API.
Claude	Anthropic's assistant and model family: Haiku is the fastest, Sonnet the everyday balance, Opus the careful tier.
Consumer vs business plan	The same chat window under two contracts: business plans do not train on your text by default and let an administrator set the rules.
Project	Files plus standing instructions that persist across chats; the tool's memory for a piece of work.
Custom GPT / Gem / agent / playbook	The vendors' names for a packaged set of instructions, sources and a model that somebody else can run.
Agent Designer	Google's builder in Gemini Enterprise for agents described in plain language, which can be shared and scheduled.
Artifact	Claude's name for a document, chart or small app that opens beside the chat and can be edited and shared.
Canvas	A document or code file that opens beside the chat and that the model edits in place while you talk. Better than re-pasting a draft every turn, and it exports when you are done.
Deep Research	A mode that plans a search, reads what it finds and returns a report with its sources, in minutes rather than seconds. Use it for a background paper or a comparison of options, not for a one-line fact.
Terra, Sol, Luna, Astra	OpenAI's names for variants within one GPT generation, shown in the model picker beside the version number.

The model

Term	One line
LLM	A model that predicts the next token; fluent, confident, memoryless between calls.
Token	About ¾ of a word; the unit everything is measured in.
Reasoning / thinking mode	A setting where the model works through steps before answering; slower, better for hard multi-step tasks.
Multimodal	A model that reads images, documents, audio or video as well as text.
Hallucination	Confident, fluent and wrong. It follows from how the method works, so no version patches it away. "How the model decides" below has the reason: a guess is graded better than a blank.
Parameters / weights	The learned numbers inside a model, counted in billions; the model file holds nothing else.
Precision / quantization	How many bits store each parameter; fewer bits means a smaller, faster model with a small quality cost.
Open-weight vs closed-weight	Whether the parameter file is published, so anyone can run it, or reached only through the vendor's service.
Knowledge cutoff	The date a model's training data ends; anything later must come from search or your documents.
Tier (Flash, Pro, mini)	Size-and-speed variants of one model generation.
Benchmark	A score on a public test, usually vendor-reported; use it to shortlist, never to choose.
Realtime voice model	A model that takes speech in and gives speech out through one model, instead of chaining speech-to-text, a text model and text-to-speech. It predicts what sound comes next, so it handles pauses and interruption naturally but has different failure modes from a text model.
"35B"	The parameter count: 35 billion learned numbers in the file. It sets how much memory the model needs and is a rough guide to capability within one family, not across families.
Q4 / Q8	Bits per weight after quantization: Q8 is near-lossless and about twice the size of Q4. Q4 roughly halves the memory for a small quality cost, which is why most models people run themselves are Q4.
imatrix	An importance matrix: a calibration pass over sample text that records which weights matter most to the output. The quantizer then spends its bits on those weights, which is what keeps low-bit versions usable.
GGUF	A single-file container for a quantized model that holds the tensors plus standard metadata, so one file loads and runs. It is the format you will see on a download page for a model somebody runs on their own hardware.
Mixture of experts (MoE)	An architecture where each token is routed to a few expert sub-networks instead of the whole network. "235B-A22B" means 235 billion parameters in total with 22 billion active per token: total predicts the memory you must buy, active predicts the speed.
Open-weight examples	Models whose weight files are published, so anyone can run them: Qwen3, gpt-oss, Kimi K2, DeepSeek V3, Llama 4. None of them are the hosted chat apps; they are what people run on their own hardware, and Llama 4 is the last Llama.

How the model decides

Term	One line
Probabilistic	The answer is drawn rather than looked up, so the same request twice gives the same sense in different words.
Deterministic	Gives the same answer every time for the same input: a script, a spreadsheet. The opposite of a language model.
Sampling	The step that takes the next word from the weighted shares; the draw itself.
Temperature	One number that sets how warm the draw is: low keeps the leading words dominant, high gives the tail a real chance.
Temperature zero / greedy	No draw at all: the top-scoring word wins every time, so the same input gives the same answer. It makes the answer repeatable and no more correct.
Top-p	Keep only the leading words that together cover that share of the draw, and throw the tail away before drawing.
Seed	The number that fixes the random draws, so a run can be repeated.
Logits	The raw scores the model produces, one for every word it could write next, before they are turned into shares.
Softmax	The step that turns those scores into shares that add up to one.
Entropy	How spread the draw is, in one number: low where one word is nearly certain, high where many would fit.
Perplexity	Entropy read as a number of effective choices; lower means the text surprised the model less.

Other kinds of AI

Term	One line
Reinforcement learning	Learning by trying actions and keeping what earns a reward; used for control and robotics, and for the training behind thinking modes.
Evolutionary method	Keep the best candidates, change them at random, try again; used where there is no formula to solve.
Search / solver	A method that computes the cheapest route or the best assignment under your constraints. Same input, same answer.
Diffusion model	An image or video model that starts from noise and removes it step by step; the randomness is the noise and the draw at each step.
Classifier	A model that sorts an input into one of a fixed set of categories: spam or not, this tag or that one.

The prompt

Term	One line
Prompt	The entire text the model receives this turn, not only what you typed.
System prompt / preamble	Vendor and organization instructions placed before your message.
Context window	The maximum tokens per call; your conversation shares it with the preamble, tools and files.
Compaction	Summarizing older turns to make room; detail is lost.
Harness	The product layer around the model: preamble, tools, memory, filters, compaction.
Prompt injection	Instructions hidden in data (a document, a web page, a tool result) that steer the model as if you typed them.

Building with it

Term	One line
Skill (file)	A description that triggers and a body of procedure, an example and checks, loaded into the prompt on demand.
Tool	A function the model can request; the harness runs it and returns text.
Function calling	The mechanism: the model emits a structured request naming a tool and its inputs.
Schema	The fixed shape of a tool's inputs or an output (fields, types); what a program can check.
Tool result	The returned text, appended to the context, possibly truncated.
MCP	Model Context Protocol, the standard connector format for tools; connectors are MCP servers.
Agent	An LLM in a loop with tools, until a stop condition.
Subagent	A fresh-context helper spawned by an agent; returns only a summary.
Session	One running transcript.
Pipeline	Deterministic steps with the model as one of them, and a check after it.
API	Calling the model from code; your prompt, your loop, your checks, none of the harness.
API key	The credential that lets code call the model; it acts as you.
Rate limit	The cap on calls per minute; queue and retry.

Documents and retrieval

Term	One line
RAG	Retrieve the relevant chunks of your documents, then generate from them with citations.
Chunk	A piece of a document, a few hundred tokens, cut on structure.
Embedding	A vector that places a chunk or a question by meaning; how retrieval finds "relevant".
Vector index	The store of embeddings that a question is searched against.
Hybrid search	Vectors for meaning plus keywords for exact strings, searched together.
Re-ranker	A second model that reorders retrieved chunks by relevance before the model reads them.

Training and evaluation

Term	One line
Training	Moving the weights, a little at a time, toward values that make predictions less wrong.
Loss	The one number that says how wrong the predictions were on a batch; training pushes it down.
Pre-training / post-training	Learning language from trillions of tokens; then learning to be an assistant from curated examples and preferences.
Checkpoint	A saved copy of the weights during training; the best one is chosen on held-out data, not the last.
Fine-tuning	Further training on your examples so the model behaves differently; teaches behavior, not facts.
LoRA	Fine-tuning by training small adapter matrices beside frozen weights; cheap, swappable.
Distillation	Training a small student model to imitate a large teacher on one narrow task.
Held-out set	Inputs the model never trained on, with reference answers fixed in advance; the only evidence that counts.
Baseline	What you do today, scored on the same set; the thing a candidate must beat.
Blind pairwise judging	Two outputs side by side, order randomized, labels hidden; a reader picks.

Staff and budget words

Term	One line
PPBE	Planning, Programming, Budgeting and Execution: the calendar-driven cycle that turns guidance into a budget, starting more than two years before the money is spent. "PB&E" is not a real term; when you hear it, somebody means PPBE.
POM	The Program Objective Memorandum, the programming phase's product: five years of programmed resources that adjust the future years defense program. Your department's data is what somebody else's POM argument is built from.
Issue paper	A real budget-review artifact, not shop jargon: analysts issue their marks as issue papers, and the command that disagrees answers with a reclama. Short, sourced and argued, which is what a project with the source documents can draft.
UFR	An unfunded requirement: something needed for which no appropriation was requested. The congressional unfunded priority list is a service-chief product; a command's internal UFR list is a different thing wearing the same name.
SMRD	Shore Manpower Requirements Determination: how a shore activity's manpower requirements are established from approved workload. Commands largely determine their own, which is why the workload arithmetic gets scrutinized on the way up.
AMD	Two different documents share this abbreviation. The Activity Manpower Document is the requirements side, and the Activity Manning Document is the distribution side; say which one you mean or a draft will mix them.
DRRS-N	The Navy's readiness reporting system, built on mission-essential tasks and weighted to the commander's own assessment. An approved task list has to exist before a unit can report against it.