The AI glossary

12V-2x6 power connector

The 16-pin plug that powers modern high-end graphics cards. Push it all the way in.

Active parameters

In a mixture-of-experts model, the small part of the model actually used for each word. It decides speed; the whole model still has to fit in memory.

Adapter

A small trained add-on that changes how a model works. The base model's stored numbers stay fixed.

Agent harness

The code that runs an agent's tools. It can keep logs and limit what the agent may do.

Agent loop

The cycle an agent repeats: look, decide, act, check, until the task is done or it must stop.

Agentic AI

AI that plans and carries out several steps on its own instead of answering one question at a time.

AI agent

An AI system that takes steps to do a task. It may choose tools, search, or work with files. Its access should have clear limits.

AI disclosure to the court

Some courts require you to say if AI helped prepare a filing.

AI governance

The rules an organization sets for which AI tools may be used, for what, and with what checks.

AI model

The trained part of an AI system. It uses a large set of numbers to turn what you give it into a result.

AI slop

Poor or bland content made with AI. The word is a criticism, not a precise test of quality.

Alexi

A legal research tool made in Canada. It can draft memos that cite cases. Check the cases before you rely on them.

Alignment

Work to make a model act as people intend. For example, it should follow rules and avoid harm. This is a goal, not a promise.

Anthropic

The company that makes Claude.

AnythingLLM

A desktop app for chatting with your own documents using a local model.

API key

A secret password that proves a program is allowed to use a service. Keep it private.

API — Application Programming Interface

A doorway that lets one program talk to another. Apps use an API to send your request to the model and get the answer back.

Apple Silicon

Apple's M-series chips, which put the processor, graphics, and memory together. Good at running local AI.

Application

The app or website you actually use, like ChatGPT or Claude. It wraps around the model and decides what files, searches, and tools the model gets.

Approval gate

A point in an automated task where it must stop and wait for a person to say yes.

Artificial general intelligence (AGI)

A still-imaginary AI that could do most mental tasks as well as a person. Nobody agrees on how to test for it.

Artificial intelligence (AI)

Software that can do tasks such as finding patterns, working with language, or making plans. Those skills do not prove it thinks like a person.

Attention

A way for a model to weigh links between parts of its input. The weights help it process text. They do not show which claims are true.

Audit log

A record of what the system did, when, and on whose instructions.

Authentication

Proving who you are, for example by logging in.

Authorization

Checking what someone who has logged in is allowed to do.

Base model and instruct model

A base model just continues text. An instruct model has been trained to follow instructions and chat.

Batch processing

A group of tasks sent or run together. Some services charge less if the tasks can wait.

Benchmark

A standard test used to compare models. A high score on it does not mean the model is good at your task.

Bias

A pattern of skewed or unfair results. It can come from the data, the model, or the way a tool is used.

BIOS / UEFI

The settings screen built into the motherboard that appears before Windows starts.

Hardware & hosting Advanced

Bit depth and precision

How much detail each number in a model is stored with. Fewer bits means a smaller, faster, slightly rougher model.

Blue J

A legal research tool known for tax work. Check which laws and sources its current plan covers.

Calibration

Whether a system's stated confidence matches how often it is actually right. 'Ninety percent sure' should be right nine times in ten.

Case and airflow

The box that holds the parts. Its fans move air through it so the parts stay cool.

Certification of authorities

A statement a lawyer signs in a court filing saying they checked that every case cited is real. Ontario requires it in factums.

Chain of thought

Having the model write out its steps before the final answer. It often helps, but the written steps are not proof the answer is right.

Chatbot

A program you talk to in ordinary language.

ChatGPT

OpenAI's chat app, built on its GPT models.

Checkpoint

A saved point in a long task that work can be resumed from.

Chunking

Cutting a long document into smaller pieces so they can be searched and fed to the model.

Citation

A pointer to the source that supports a statement, so someone can check it.

Classification

Sorting items into groups. For example, put each email in a group based on its topic.

Claude

A family of AI models and a chat app from Anthropic. The app's tools depend on the plan and setup.

Claude Code

An AI tool that works with code and files. It can read, edit, and run commands within the access it has.

CLAUDE.md

A file named CLAUDE.md that gives Claude Code project guidance. Which files it reads depends on the setup.

CLI — Command-line interface

A text window where you type commands to the computer instead of clicking buttons.

Clio Duo

Software for running a law practice. Some of its tools use AI. Clio is the wider product, not just its AI helper.

Hardware & hosting Beginner

Cloud inference

Running a model on a remote server. Your input travels to that server, so check its data terms before use.

CoCounsel

A legal AI tool from Thomson Reuters. It helps with tasks such as research and document review. Its output still needs checking.

Codex

OpenAI's coding agent, which can read, write, and run code on a task you give it.

Compaction

Squeezing a long conversation into a short summary so there is room to keep going. Details can be lost in the squeeze.

Computer use

An AI system using screen controls. It can click, type, and move through apps, within the access it has.

Confidentiality

Keeping private material from people who may not see it. Check who can read, store, or share what you put into a tool.

Connector

A link between an AI app and a service you use, like your email or document store.

Container

A packaged-up program with everything it needs, so it runs the same way on any computer.

Context engineering

Choosing what a model gets to work with. Give it the facts and rules it needs. Leave out clutter that may confuse the task.

Context rot

A drop in answer quality as the context grows or fills with clutter. It does not happen in the same way in every model or task.

Context window

How much text the model can hold in mind at once: your question, the documents you added, and the conversation so far. Anything outside it is invisible to the model.

CPU cooler

A part that removes heat from the main chip. It may use a fan, a metal block, or a liquid loop.

CPU — Central processing unit

The main brain of a computer. It can run AI models, but slowly.

CSV — Comma-separated values

A plain-text spreadsheet: one row per line, values separated by commas.

CUDA

Nvidia's system for running AI maths on its graphics cards.

Cursor

A code editor with an AI assistant built in.

Custom GPT

A version of ChatGPT you set up with your own instructions and files for a particular job.

Data leakage

Data reaching a place it should not. In model tests, it can also mean test answers have reached the training data.

Data residency

Which country the data is stored and processed in.

Data retention

How long a service keeps your data after it has been used.

Database

A store of data that software can search and change. It keeps the data in a set structure.

Dataset

A collection of examples used to train or test a model.

DDR5

The current kind of computer memory. Install it in matched pairs so it runs at full speed.

Decode

The model writing its answer one word-piece at a time. How fast this goes depends mostly on memory bandwidth.

Deep learning

Machine learning that uses a big stack of simple maths layers. Stacking many layers lets the program learn complicated patterns, like speech or faces.

Deep research

A feature where the AI searches the web many times and writes a long report with links. Every link still needs checking.

Deepfake

A fake photo, video, or voice recording made by AI that looks or sounds real.

DeepSeek

A company in China that makes AI models. Some of its models can be downloaded and run on your own machine.

Dependency

Something a program needs installed to work.

Deployment

Putting a tool into use. This includes deciding who can use it and how it will be kept working.

Direct preference optimization (DPO)

A way to train a model with pairs of answers. Each pair shows which answer people preferred.

Distillation

Training one model with help from another. A smaller model may learn some of a larger model's skills, but can lose others.

eDiscovery

Finding and reviewing digital records for a legal case. It can include searching emails and deciding what must be produced.

Embedding

A list of numbers used to compare items such as texts. Texts with similar meanings may have similar lists.

Embedding model

A model whose only job is to turn text into embeddings for searching.

Endpoint

An address a program calls to do a set task. One address might accept a file; another might return a reply.

Entity extraction

Pulling out names, dates, amounts, and places from text.

EU AI Act

An EU law about AI. The rules depend on the system, its use, and the date. Check the current law for the task you plan.

Evaluation (eval)

Testing an AI tool on a set of tasks to see how well it actually does.

Explainability

Being able to understand why a system gave the answer it did.

Extended thinking

A mode where the model thinks for longer before answering. The visible 'thinking' is text it wrote, not a window into its mind.

Fabricated citation

A case or quotation the AI invented. It looks real and is not.

False negative

Something the system missed that it should have found.

False positive

Something the system flagged that it should not have.

Few-shot prompting

Showing the model a few examples of what you want before asking it to do the same thing.

Fine-tuning

More training on a chosen set of examples. It changes a model to suit a task or style.

Firewall

A gatekeeper that decides which programs can use the internet. You can block one app to prove it works offline.

Foundation model

A model trained for a broad range of tasks. Other tools can use it as a base.

Framework

A ready-made structure for building a certain kind of program.

Frontier model

A model near the leading edge of current AI skills. Which model leads depends on the task and the date.

Gemini

Google's name for a family of AI models and its AI app. The model and the app are different parts of the system.

Gemini Notebook (formerly NotebookLM)

A Google tool for asking questions about sources you add. Check that each answer is supported by those sources.

Gemma

A family of models from Google. You can download the model files and use them under the stated terms.

Generative AI

AI that makes content, such as text, images, or sound. What it makes may be wrong or may resemble existing work.

GGUF

The file format most local AI apps use for downloaded models.

Git

A tool that keeps a history of every change to a set of files, so you can go back or see who changed what.

GitHub

A website where code is stored and shared.

GitHub Copilot

An AI helper for programmers that suggests code as they type.

GPT — Generative Pre-trained Transformer

The family of models behind ChatGPT. The letters stand for Generative Pre-trained Transformer.

GPU offload

How much of the model sits on the graphics card. Whatever does not fit goes to slower system memory, which slows the answer.

GPU — Graphics processing unit

The chip that handles graphics, and also the maths AI needs. It is what makes AI fast.

Gradient

The direction to nudge the model's numbers so its mistakes get smaller. Training takes many small steps in that direction.

Graphics driver

The software that lets Windows or macOS use your graphics card. Get it from the card maker.

Grok

The AI assistant made by xAI, available in X and as an app.

Grounding

Basing an answer on a named source or set of facts. A link to a source does not prove the answer follows from it.

Guardrail

A rule or filter that stops an AI system from doing certain things.

Hallucination

When the model states something that sounds right but is made up, including cases, quotes, and facts that do not exist.

Hallucination rate

How often a system makes things up, measured on a set of test tasks.

Harvey

An AI platform made for law firms, used for research, document review, and drafting.

Hermes Agent

An open-source agent made by Nous Research.

HTTP — Hypertext Transfer Protocol

The set of rules computers use to send pages and requests over the web.

Hugging Face

A website where models and datasets are shared and downloaded.

Human in the loop

A design where a person reviews or approves what the AI does before it counts.

Hybrid search

Using keyword search and meaning search together.

IDE — Integrated development environment

A program for writing code, with helpers built in.

Idempotency

A property of an action: doing it twice has the same effect as doing it once. This helps when a system retries a task.

Inference

The act of running a trained model to get an answer. In law the word means something else: a conclusion drawn from evidence.

Input token

A piece of text you send to the model, including documents and instructions.

Jailbreak

A trick that gets a model to ignore its rules.

JSON — JavaScript Object Notation

A tidy text format for data, made of labels and values in curly brackets. Programs use it to pass information around.

Keyword search

Searching for the exact words you typed.

Knowledge base

A set of records a system can search. It might hold guides, case notes, or other source files.

Knowledge cutoff

The date the model's training data ends. It knows nothing after that unless you give it the information.

Knowledge graph

A map of things and how they connect, such as people, companies, and events, with lines between them.

Hardware & hosting Advanced

KV cache — Key-value cache

Stored results from part of a model's earlier work. Reusing them can speed up the next step. This cache also takes memory.

LangChain

A toolkit programmers use to connect models to data and tools.

LangGraph

A toolkit for building agents as a map of steps.

Large language model (LLM)

A program trained on enormous amounts of text so it can read and write language. ChatGPT and Claude are built on large language models.

Latency

The wait between sending a request and getting the answer back.

Least privilege

Giving a system only the access it needs for the job, and nothing more.

Legora

An AI workspace for legal work. It includes tools for research, drafting, and review. Check its current features and sources.

Lexis+ AI and Protégé

AI tools within Lexis+. They use legal content to help with research and drafting. Check the sources and the plan's scope.

License

The terms that say how you may use a piece of software or a model.

Llama

A family of AI models from Meta. Model files can be downloaded under a licence. Check the terms for the version you use.

llama-bench

A small tool that measures how fast a model reads and writes on your computer.

llama.cpp

Software that runs AI models on your machine. Many local chat apps use it behind the scenes.

LLM-as-a-judge

Using one AI to grade another AI's work.

LM Studio

A desktop app for running models locally, with a chat window.

Hardware & hosting Beginner

Local model

A model that runs on your own computer, so your documents never leave it.

Loss function

The score training uses to measure how wrong a model's guess was. Training tries to make this number smaller.

Low-rank adaptation (LoRA)

A cheap way to fine-tune: instead of changing all the model's numbers, you train a small add-on that sits beside them.

M.2 NVMe SSD

A small, very fast storage stick that plugs straight into the motherboard. Model files load from it quickly.

Machine learning (ML)

Building a program by training it on examples. The training shapes how it handles new input.

Markdown

A simple way to add headings, lists, and bold text using plain characters like # and *. AI tools read and write it easily.

MCP — Model Context Protocol

A shared standard that lets AI apps plug into outside tools and data sources in one agreed way.

Memory

A feature that saves facts between conversations so you do not have to repeat them. It is a saved note, not the model learning.

Memory bandwidth

How fast data moves out of memory, in gigabytes per second. Think of it as the width of a loading door: wider means the model writes words faster.

Memory profile (XMP / EXPO)

A switch in the BIOS (called XMP or EXPO) that makes your memory run at the speed printed on the box.

Metadata

Information about a file rather than its contents, such as its author, date, and page count.

Metal

Apple's system that lets software use the graphics part of a Mac's chip. Local AI apps use it automatically.

Microsoft Copilot

A name Microsoft uses for several AI products. Features vary by app and plan, so name the one you mean.

Mistral AI

A company in France that makes AI models. Some can be downloaded. Others are used through a hosted service.

Foundations Advanced

Mixture of experts (MoE)

A model built from several specialist sub-networks, with only a few switched on for each piece of text. It saves work without losing size.

MLX

Apple's own software for running AI models on Macs with Apple chips. Often the fastest choice on a Mac.

Model card

A short document that describes a model: what it was trained on, what it is good for, and its limits.

Model serving

Running a model so that other programs can send it requests.

Model weights

The full set of numbers that make up a trained model. Download the weights and you have the model.

Motherboard

The main board every other part plugs into. It decides which processor, memory, and cards you can use.

Multimodal

Able to work with more than text, such as images, audio, or video.

MXFP4

A compact 4-bit format some newer models are released in, which keeps their files small.

Natural language processing (NLP)

The field of getting computers to work with human language.

Neural network

A model made of linked layers of maths. Training adjusts its numbers so it can learn patterns from data.

OAuth

A standard way to let one app use another on your behalf without giving it your password.

Observability

Being able to see what an AI system is doing inside, so problems can be found and explained.

OCR — Optical character recognition

Turning a picture of text, such as a scanned page, into real text a computer can search.

Offline test

A test that shows a task can run with no network link. It does not prove that the app will never send data later.

Ollama

Software for downloading and running AI models. It can run local models, but you still need to check which model and service you chose.

Open source

Software whose code is published for anyone to use, study, and change.

Open WebUI

A chat interface you can host yourself. It connects to a model service, which may be local or remote.

Open weights

A model whose trained numbers are published, so anyone can download and run it.

OpenAI

The company that makes ChatGPT and the GPT models.

OpenClaw

An open-source personal AI agent that runs on your own computer.

Orchestration

Managing the steps in a workflow. It sets which step runs next and how results pass between tools.

Output token

A piece of text the model writes. Output is usually priced higher than input.

Overfitting

When a model fits its training examples too closely. It may then do poorly on new examples.

Parameter

One of the billions of adjustable numbers inside a model. Training sets them; using the model does not change them.

PCIe

The long slot on the motherboard where the graphics card plugs in.

Permission

What an AI system is allowed to see or do, set by the people who run it.

Perplexity (metric)

A score for how surprised a model is by a piece of text. Lower is better, but it does not measure accuracy on your task.

Perplexity (product)

An AI search app that answers questions with links to web sources.

Personal information

Data about a person who can be identified. Which privacy rules apply depends on the data, its use, and the law.

Power supply (PSU)

The box that turns wall power into power for the computer. It must be big enough for the graphics card.

Precision

Of the things the system returned, how many were right.

Prefill

The model reading your question and documents all at once before it starts answering. Long documents mean a longer wait for the first word.

Pretraining

An early, broad stage of model training. The model learns patterns from a large set of data before later training for some uses.

Privilege and AI tools

A legal shield for some private talks and records. For lawyer-client privilege, the legal purpose matters. Not every private message is covered.

Prompt

What you give a model to guide its reply. This may include a question, rules, and source text.

Prompt caching

Reusing some work on input the model has seen before. When supported, this can cut time or cost.

Prompt engineering

Writing and testing clear requests for an AI tool. The aim is an answer that meets your task and can be checked.

Prompt injection

Text in a source that tries to redirect an AI tool. It may tell the tool to ignore your rules or leak data.

Provenance

The record of where a piece of information came from and how it got here.

Python

A popular programming language that is easy to read and widely used for AI work.

Quantization

Storing model numbers with less detail. This can save memory. The effect on quality depends on the model and the task.

Quantization labels (Q4_K_M and others)

The codes in model file names, like Q4_K_M. The number is roughly how many bits each setting uses; smaller is lighter and slightly less precise.

Qwen

A family of AI models made by Alibaba. Some can run on your own machine. Check the terms and size of the model you choose.

RAG — Retrieval-augmented generation

Letting the model look things up in your documents before answering, and show where the answer came from. Like an open-book exam.

RAM — Random-access memory

A computer's short-term working memory. Programs and data live here while in use.

Rate limit

A cap on how many requests a service will accept in a given time.

Reasoning model

A model trained to spend extra work on some answers. It may help with hard tasks. It can still be wrong, and speed and cost vary.

Recall

Of the things it should have found, how many it actually found.

Red teaming

Trying to make a system fail or break its rules. The aim is to find weak points before real use.

Redaction and de-identification

Removing private details before sharing a file. A black box on a page may hide text without deleting it.

Regression test

A test you rerun after every change to make sure nothing that used to work has broken.

Reinforcement learning from human feedback (RLHF)

Training that uses people's ratings of model answers. It helps steer the model toward replies people prefer.

Repository

A folder of files whose history is tracked by Git.

Reranking

A second pass that puts the best search results at the top.

Retrieval

Searching a larger set of records for useful parts. Finding a passage does not prove it answers the question.

Retry

Trying a failed step again, usually after a short wait.

Reward model

A helper model that scores answers so the main model can learn what people like.

ROCm

AMD's system for running AI maths on its graphics cards.

Routing

Choosing where a task goes next. For example, send a hard question to a stronger model or to a person.

Safetensors

A file format for stored model numbers. Some local apps use another format called GGUF. Check which format your app accepts.

Sampling

How the model picks the next word from its list of likely candidates. It can always take the top one, or roll the dice among the top few.

Sandbox

A sealed-off space where an AI or a program can run without touching your real files or systems.

Schema

A description of the shape data must have: which fields exist and what goes in each.

Script

A short program that does one job, like renaming a folder of files.

SDK — Software Development Kit

A kit of code for using a service. It helps a developer call that service from a program.

Semantic search

Searching by meaning rather than by exact words, so 'delay' can find 'postponed'.

Server

A computer that runs somewhere else and answers requests from your device.

Shadow AI

AI use that a firm has not approved or does not know about. A staff member's personal chat account is one example.

Skill

A reusable pack of instructions, files, or code for an AI task. It can guide work but does not guarantee a good result.

Small language model (SLM)

A smaller language model that may fit on a phone or laptop. Its speed, skills, and privacy depend on how it is used.

Software library

A set of ready-made code other programs can use.

Speaker diarization

Working out who said what in a recording.

Speech to text (STT)

Software that turns speech into text. Check names, numbers, and words it may have misheard.

Spellbook

An AI tool for contract work in Word. It helps draft and review text. A lawyer still needs to check the result.

SQL — Structured Query Language

The language used to ask a database questions and change what it holds.

SSD — Solid-state drive

A fast storage drive with no moving parts, where files and model downloads live.

Stop condition

The rule that tells an agent when to stop.

Streaming

Showing the answer word by word as it is written, instead of waiting for the whole thing.

Structured output

An answer in a set format, such as named fields. Software can check the format, but the facts still need review.

Subagent

A helper agent that a main agent hands part of a job to.

Sycophancy

A model tending to agree with the user, even when the user is wrong. Pushback can make it change a correct answer.

Synthetic data

Examples made rather than collected from real events. They can help train or test a model, but may miss real-world problems.

System instruction

Standing instructions the app gives the model before your conversation starts, such as how to behave or what to refuse. Also called a system prompt.

Technological competence

A lawyer's duty to understand the tools they use well enough to use them safely.

Technology-assisted review (TAR)

Software that helps rank documents for review. Some forms learn from a reviewer's choices. The process still needs testing and oversight.

Temperature

A setting that affects how varied the output may be. Lower values often make it less varied. They do not ensure the same answer each time.

Foundations Intermediate

Tensor

A grid of numbers (a table, or a stack of tables). Everything inside a model is stored and calculated as tensors.

Test set

Examples held back from training so you can check how the model does on things it has not seen.

Text extraction

Pulling the words out of a file like a PDF so a program can use them.

Text to speech (TTS)

Software that reads text aloud. Some tools can also imitate a voice.

Thermal paste

A thin grey paste between the processor and its cooler that helps heat move across.

Throughput

How much work a system can get through in a given time.

Time to first token

How long you wait before the first word of the answer appears. It grows with the length of what you gave the model to read.

Token

A piece of text, often part of a word, that the model actually works with. Models read and write tokens, not whole words.

Tokenizer

The tool that chops text into tokens and gives each piece a number. Different models chop differently.

Tokens per second

How many word-pieces the model reads or writes each second. Around 20 or more feels comfortable to read along with.

Tool

Something an AI system can use to do a task. Examples include a search service, a calculator, and a file reader.

Tool calling

The model asking for a tool to be run, and the app running it and handing back the result.

Top-p sampling

A setting that narrows the next-token choices. It keeps a group with a set share of the total chance. Lower values usually narrow that group.

Training

Adjusting a model with examples and a learning rule. The aim is to improve how it handles tasks.

Training on user data

Whether the company may use what you type to teach its future models. Many free apps do unless you turn it off.

Transformer

The design most modern language models use. Its trick is letting every word in a passage look at every other word to work out what matters.

Unified memory

One big pool of memory shared by a computer's processor and graphics, as on Apple Silicon Macs. Big models fit; they just run a bit slower than on a fast graphics card.

Validation

Checking that something is correct or in the right form before relying on it.

Vector database

A store built for embeddings, so you can quickly find texts with a similar meaning.

Vibe coding

Describing a program you want to an AI and running what it writes without reading the code.

Virtual machine (VM)

Software that acts like a separate computer. It shares the real machine's hardware. Its separation is not a complete security guarantee.

Vision model

A model that can look at images as well as read text.

vLLM

Software for running models quickly for many users at once.

VRAM — Video random-access memory

The memory on the graphics card. A model has to fit in it to run quickly.

Vulkan

A standard that lets software use graphics chips. Support for a model or task still depends on the program and the chip.

Webhook

A message one service sends when an event occurs. For example, a new file can trigger a message to another app.

Workflow

A fixed sequence of steps a task goes through, which may include AI steps and human checks.

XML

A way of labelling parts of a text with tags, like <date>12 March</date>, so a program knows what each part is.

Zero data retention

A promise from the provider not to keep your prompts and files after answering.