Plain-language AI vocabulary for the SPARK practicum.

Artificial intelligence (AI)

A broad category of computer systems that perform tasks associated with reasoning, language, perception, prediction, or decision-making.

Generative AI

AI that creates new text, images, audio, code, summaries, plans, or other content from a prompt.

Large language model (LLM)

A model trained on large amounts of text so it can predict and generate language, including answers, drafts, explanations, and code.

AI model

The trained system that receives input and produces output. In SPARK, this often means an LLM such as ChatGPT, Claude, Gemini, or a local model.

Model card

A summary of a model's capabilities, limitations, evaluation results, and intended uses.

Model provider

A company or organization that hosts AI models and provides software, websites, desktop tools, or APIs for using them.

Model selection

Choosing which model to use for a task based on capability, cost, speed, data policy, and the amount of context needed.

Open model

A model whose weights, code, license, or distribution terms make it possible for people outside the original company to run or adapt it.

Local model

An AI model that runs on your own computer instead of sending every request to a provider's data center.

On-prem model

An AI model run on servers controlled by an institution rather than on a public cloud service.

Prompt

The instruction, question, example, or source material you give an AI system so it can produce a response.

Context

The information the model can currently use, including your prompt, earlier conversation, uploaded files, tool results, and system instructions.

Context window

The maximum amount of text and other information a model can consider at one time.

Token

A small unit of text that a language model reads or writes. A token may be a word, part of a word, punctuation, or whitespace.

Token cost

The price or usage impact of sending tokens to a model and receiving tokens back from it.

Token budget

The practical limit on how many tokens you want an AI system or agent to use for a task.

Output distribution

The model's calculated list of possible next tokens and their probabilities before it chooses what to write next.

Next-token probability

The model's estimate of how likely each possible next token is, given the context it has so far.

Sampling

The process of choosing a token from the model's output distribution and adding it to the response.

Few-shot examples

A small set of examples included in a prompt to show the model the desired pattern, format, or tone.

Chain-of-thought prompting

Asking a model to work step by step or reason carefully before giving a final answer.

Tree-of-thought prompting

Asking a model to explore multiple possible approaches before selecting or comparing answers.

Context compaction

Replacing a long conversation history with a shorter summary so the model can continue working within its context window.

Agent

An AI system that loops through steps, responds to prompts, uses tools when needed, and continues working toward a goal.

Agentic workflow

A workflow where an AI agent plans, takes actions, observes results, and revises its next steps.

Tool call

An action a model takes through a connected tool, such as searching the web, reading a file, running code, or creating a task.

API

An application programming interface: a set of rules that lets one computer program use another computer program.

MCP

The Model Context Protocol, a standard way for AI tools to connect with external systems and data sources.

Connector

A configured connection that lets an AI tool access another service, such as email, calendars, files, GitHub, or Linear.

Plugin

An add-on that extends a tool with new capabilities, integrations, or workflows.

Skill

A packaged set of instructions or capabilities that helps an AI assistant perform a specialized task.

Desktop tool

Software installed on your computer that can work with local files, accounts, development tools, or operating-system features.

IDE integration

A connection between an AI tool and a programming environment so the tool can read, edit, explain, or test code in a project.

AI-assisted workflow

A repeatable work process where AI helps with part of the task, such as drafting, organizing, checking, summarizing, or transforming information.

AI-assisted development

Using AI tools to help plan, write, edit, test, document, or maintain software and digital artifacts.

Prototype

An early working version of a tool or workflow used to test an idea before polishing or deploying it.

Artifact

A durable file or output created during work, such as a plan, document, spreadsheet, slide deck, website, script, or test result.

Markdown

A simple plain-text format for headings, lists, links, and notes, commonly saved in files ending with .md.

Repository

A folder of project files tracked by version control, often hosted on a service such as GitHub.

Version control

A system for recording changes to files so you can review history, collaborate, and return to earlier versions.

Git

A widely used version control system that tracks changes to project files on your computer.

Commit

A saved snapshot of changes in Git, usually with a short message explaining what changed.

Push

Sending local Git commits from your computer to a remote repository such as GitHub.

Branch

A separate line of work in Git that lets you make changes without immediately changing the main version.

Merge

Combining changes from one Git branch into another.

Pull request

A proposed set of changes in GitHub that others can review, discuss, and merge into a project.

Issue

A tracked task, bug, question, or feature request in a project-management system such as Linear or GitHub.

Milestone

A larger project checkpoint that groups related issues or tasks.

Acceptance criteria

Specific conditions that must be true before a task or feature is considered complete.

Deployment

Making a tool or website available where intended users can actually use it.

Validation testing

Checking whether a tool's outputs are accurate, useful, and appropriate for the task it is supposed to support.

Roadmap

A plan for future work that lists major next steps, priorities, and improvements.

Proprietary data

Information owned or controlled by an organization that should not be shared without permission.

FERPA-covered data

Student education records protected by the Family Educational Rights and Privacy Act.

De-identification

Removing or changing details that could identify a person, group, or protected record before data is shared or processed.

Accuracy verification

Checking AI output against reliable sources, original data, or human expertise before relying on it.

Security review

Examining a tool, workflow, or integration for risks involving data access, accounts, permissions, storage, and misuse.

Red-teaming

Deliberately testing a tool, workflow, or AI system to find weaknesses, misuse paths, incorrect outputs, or unintended consequences.

Red team

A person or group authorized to act like an adversary or skeptical tester to expose problems before real users encounter them.

Blue team

The defenders or maintainers who use red-team findings to improve security, reliability, or response.

Failure mode

A specific way a system, workflow, or AI output can go wrong.

Data loss recovery

The plan for restoring files, records, or work if a tool deletes, overwrites, corrupts, or loses data.

Personal account access

Permission for an AI tool or connector to interact with an individual's email, calendar, files, or other personal services.

Model evaluation

Testing a model or AI-assisted workflow against defined tasks or criteria to judge capability, reliability, safety, or fit.

System card

A broader report on an AI system's behavior, safeguards, risks, and evaluations.

Eval

A repeatable test used to measure whether an AI model or workflow performs a task correctly or safely.

Benchmark

A test or evaluation used to compare systems on a defined set of tasks.

Benchmark dataset

A collection of tasks or questions used consistently to compare model performance.

GSM8K

A benchmark of grade-school math word problems often used to test mathematical reasoning.

GDPval

A benchmark for testing AI performance on realistic professional knowledge-work tasks.

SWE-Bench

A software-engineering benchmark that tests whether AI systems can solve real coding issues.

Humanity's Last Exam

A difficult multidisciplinary benchmark used to test expert-level reasoning across many subjects.

METR

A research organization that evaluates advanced AI capabilities and risks.

AI regulation

Laws, policies, executive actions, or institutional rules that shape how AI systems may be built, purchased, used, or evaluated.