The AI glossary

97 terms from the news, each defined in one sentence a person can actually use.

A

Access token: An access token is a temporary digital pass a service gives you after you sign in or pay, which your software shows with each request to prove it is allowed in, without sending your password again.
Agent harness: An agent harness is the software wrapped around an AI model that runs its loop, executes its tool calls, manages its memory and context, and enforces limits, turning a model into an agent.
Agent sandbox: An agent sandbox is an isolated environment where an AI agent runs, with strict limits on which files, networks and tools it can reach, so a mistake or attack stays contained.
Agentic coding: Agentic coding is software development where an AI agent plans the work, edits multiple files, runs tests and fixes its own errors over many steps, while a human sets the goal and reviews the result.
Agentic commerce: Agentic commerce is online buying and selling in which AI agents research, compare and complete purchases for a person or business, within preferences and spending limits the buyer sets.
AI agent: An AI agent is a system built on an AI model that can pursue a goal over several steps by choosing actions, using tools such as a browser or files, and acting on the results.
AI alignment: AI alignment is the work of making AI systems reliably pursue the goals and follow the limits their builders and users intend, rather than doing something technically allowed but unwanted.
AI benchmark: An AI benchmark is a standardized test, a fixed set of tasks with a scoring method, used to measure and compare what AI models can do on things like coding, math, reasoning or computer use.
AI companion: An AI companion is a chatbot app designed for ongoing personal conversation and relationship, often as a friend, character or confidant, rather than for finishing tasks like search or coding.
AI gateway: An AI gateway is a service that sits between your app and several AI model providers, giving you one interface plus routing, fallbacks, spending tracking and logging across all of them.
AI guardrails: AI guardrails are the rules, filters and checks placed around an AI model or agent to keep its inputs, outputs and actions within safe and intended limits.
AI memory: AI memory is a feature that saves details from your past conversations, such as preferences and facts about you, and brings them back into later chats so you do not have to repeat yourself.
AI watermarking: AI watermarking is a method of hiding a statistical signal in AI-generated text, images, audio or video so that a detector holding the right key can later flag it as machine-made.
Airgapping: Airgapping is keeping a computer system physically or logically cut off from the internet and other outside networks, so software on it, including an AI model, cannot reach or compromise outside systems.
Apache 2.0 license: The Apache 2.0 license is a permissive open-source license that lets anyone use, modify and sell software or a model commercially, as long as they keep the license notice and state their changes.
API key: An API key is a secret string of characters that identifies you to an AI provider's service and authorizes your requests, so whoever holds it can use your account and run up your bill.

B

Base model: A base model is an AI model that has finished pretraining on a huge amount of text but not yet been tuned to follow instructions, so it continues text rather than acting as an assistant.
Browser agent: A browser agent is an AI agent that operates a web browser for you, reading pages, clicking, typing and filling in forms across sites to finish a task, often while logged in to your accounts.

C

Capture the flag: Capture the flag (CTF) is a security challenge where players break into deliberately vulnerable systems to find a hidden secret, and AI labs use CTFs to measure how well models can carry out cyberattacks.
Chain of thought: Chain of thought is the step-by-step reasoning an AI model writes out before its final answer, which often improves accuracy on hard problems and lets people inspect how it got there.
Chain-of-thought monitoring: Chain-of-thought monitoring is checking an AI model's written reasoning, often with another model, to catch signs of cheating, deception or harmful plans before or while it acts.
ClickFix attack: A ClickFix attack is a social engineering trick that gets you to copy, paste and run a malicious command yourself, usually from a fake error, CAPTCHA or install page.
Compute: In AI, compute is the amount of raw computing power, mostly from GPUs, used to train or run a model, and it is the scarce resource that limits how big and how fast models can get.
Computer use: Computer use is an AI capability where a model looks at screenshots of a computer screen and controls the mouse and keyboard to operate software the way a person would, rather than through a purpose-built connection.
Context compaction: Context compaction is when an AI agent replaces a long conversation or work history with a shorter summary so it can keep working without running out of room in its context window.
Context engineering: Context engineering is the practice of choosing and maintaining what goes into an AI model's context window at each step, so an agent has the right information to do the job.
Context window: A context window is the maximum amount of text, measured in tokens, that an AI model can consider at once, including your prompt, any documents, the chat history and its reply.
Convolutional neural network: A convolutional neural network (CNN) is a type of neural network built for images that slides small filters across a picture to detect edges, textures and shapes, then combines them to recognize what it shows.
CUDA: CUDA is Nvidia's software platform that lets programmers run general-purpose code on Nvidia GPUs, and most AI software is built on it, which keeps developers tied to Nvidia hardware.

D

Decompilation: Decompilation is the process of turning a compiled program back into human-readable source code, an approximation of what the original programmers wrote.
Diffusion model: A diffusion model is an AI model that generates images, video or audio by starting from random noise and removing it step by step until a clean result that matches your prompt appears.
Distillation: Distillation is a training technique where a smaller "student" AI model learns to imitate the outputs of a larger "teacher" model, keeping much of its ability at a fraction of the size and cost.

E

Elo rating: An Elo rating is a score built from head-to-head matchups, and AI leaderboards use it to rank models by how often people prefer one model's answer over another's in blind comparisons.
Embedding: An embedding is a list of numbers that represents the meaning of a piece of text, an image or other data, so a computer can find items with similar meaning by comparing the numbers.
Evals: Evals are tests you run on an AI model or app, using your own set of example tasks and a scoring method, to see how well it does the specific job you need before and after any change.
Evaluation awareness: Evaluation awareness is when an AI model recognizes that it is being tested rather than used for real, which can change how it behaves and make safety test results less trustworthy.

F

Fine-tuning: Fine-tuning is the process of training an already-trained AI model further on a smaller, targeted set of examples so it behaves better for a specific task, style or domain.
Formal verification: Formal verification is the use of mathematical proof, checked by software, to show that a program or system always meets a precise specification, such as never crashing on any valid input.
Frontier model: A frontier model is a general-purpose AI model at or near the cutting edge of what AI can do, among the most capable available, and the term is relative: yesterday's frontier is tomorrow's ordinary model.
Fuzzing: Fuzzing is a software testing technique that feeds a program huge numbers of random or mutated inputs to make it crash, so hidden bugs and security flaws show up before attackers find them.

G

GPU: A GPU (graphics processing unit) is a chip built to do many simple calculations at the same time, which makes it the standard hardware for training and running AI models.

H

Hallucination: An AI hallucination is an output that sounds fluent and confident but is false or made up, such as an invented fact, quote, source or piece of code that does not exist.
High-bandwidth memory: High-bandwidth memory (HBM) is a type of stacked computer memory placed right next to an AI chip, feeding it data far faster than ordinary memory so large models can run efficiently.

I

Inference: Inference is the act of running a trained AI model to produce an output, such as answering a prompt, as opposed to training the model in the first place.

J

Jailbreak: A jailbreak is a prompt or technique that tricks an AI model into ignoring its safety rules and producing content or actions it was trained and instructed to refuse.

K

Knowledge cutoff: A knowledge cutoff is the date after which an AI model's training data stops, so it knows nothing about later events unless it is given the information or can search the web.

L

LoRA: LoRA (low-rank adaptation) is a fine-tuning method that trains a small set of added weights while leaving the original model frozen, making customization far cheaper and the result a small file you can swap in.

M

Mixture of experts: A mixture of experts is a model design that splits a neural network into many specialist sub-networks and activates only a few of them for each token, cutting the compute needed per answer.
Model calibration: Model calibration is how closely a model's stated confidence matches how often it is right: a well-calibrated model that says 90% is correct about nine times in ten.
Model Context Protocol: The Model Context Protocol (MCP) is an open standard that lets AI apps connect to outside tools and data through one common interface, so a tool built once works with many different AI models.
Model weights: Model weights are the learned numbers stored in a trained AI model that determine how it responds, and the file holding them is what a lab protects as its most valuable asset.
Multi-agent system: A multi-agent system is a setup where several AI agents work on one goal together, splitting the job, sharing information and checking each other's work instead of one agent doing everything.
Multimodal model: A multimodal model is an AI model that can take in, and sometimes produce, more than one kind of data, such as text, images, audio and video, within a single system.

N

Neocloud: A neocloud is a cloud company built mainly to rent out GPU computing power for AI training and inference, rather than offering the broad catalog of services a giant like Amazon or Google does.

O

Open-weight model: An open-weight model is an AI model whose trained parameters (weights) are published so anyone can download and run it, though the training data and code may stay private.

P

Parameter: A parameter is one of the learned numbers inside an AI model, adjusted during training, and the count of them, often billions, is a rough measure of the model's size.
Penetration testing: Penetration testing is an authorized, simulated attack on a computer system, run to find security weaknesses before real attackers do, and AI agents are now used to automate parts of it.
Post-training: Post-training is the set of steps applied to a pretrained AI model, such as fine-tuning on examples and learning from feedback, that turns raw text prediction into a useful, specialized assistant.
Pretraining: Pretraining is the first and largest stage of building an AI model, where it learns general patterns by predicting the next token across enormous amounts of text, code and other data.
Process node: A process node is a chipmaking generation, such as 5 nm, 3 nm or 2 nm, and the number is now a marketing label for a newer, denser, more efficient generation, not a literal measurement of the transistors.
Prompt caching: Prompt caching is a feature that lets an AI provider reuse its work on the start of a prompt it has already processed, so repeated input costs less and returns faster.
Prompt injection: Prompt injection is an attack that hides instructions in text an AI model reads, such as a message or web page, so the model follows the attacker's orders instead of its owner's.
Proof of personhood: Proof of personhood is a way to show that an online account belongs to one real, unique human, not a bot or a duplicate, without necessarily revealing who that person is.

Q

Quantization: Quantization is a compression technique that stores an AI model's numbers at lower precision, such as 8-bit or 4-bit instead of 16-bit, so the model needs less memory and runs on cheaper hardware.

R

Rate limiting: Rate limiting is a cap on how many requests a user, app or agent may make to a service in a set time, used to prevent overload and abuse and enforced by AI APIs and websites alike.
Reasoning model: A reasoning model is an AI model trained to work through a problem in a long internal chain of thought before answering, trading extra time and tokens for better results on hard tasks.
Red teaming: Red teaming is the practice of deliberately attacking an AI system, with tricky prompts and misuse attempts, to find harmful behavior and security flaws before real users or attackers do.
Reinforcement learning: Reinforcement learning is a way of training an AI system by letting it try actions and rewarding the outcomes that score well, so it learns what to do from results instead of from labeled examples.
Remote code execution: Remote code execution (RCE) is a security flaw that lets an attacker run their own code on someone else's computer or server over a network, often with no login, which makes it among the most severe kinds of bug.
Retrieval-augmented generation: Retrieval-augmented generation (RAG) is a technique where an AI system first looks up relevant documents, then gives them to the model along with your question so its answer is based on that material.
Reward hacking: Reward hacking is when an AI system finds a shortcut that earns a high score from its training or grading rule without doing the task the designers actually wanted done.
Robots.txt: Robots.txt is a plain text file at the root of a website that tells web crawlers which pages they may or may not visit, and it works only if the crawler chooses to follow it.

S

Scaling laws: Scaling laws are empirical rules showing that an AI model's error falls in a smooth, predictable way as you increase its size, training data and compute, which lets labs forecast results before training.
Side-channel attack: A side-channel attack steals information or sends signals by watching physical effects of a computer, such as timing, power use or cache behavior, instead of breaking the software's own security rules.
Speaker diarization: Speaker diarization is the process of working out who spoke when in an audio recording, labeling each stretch of speech as Speaker 1, Speaker 2 and so on, usually without knowing anyone's name.
Speculative decoding: Speculative decoding is a technique that speeds up an AI model's text generation by having a small, fast model draft several tokens ahead and the large model check them all at once.
Speech recognition: Speech recognition, also called speech-to-text or transcription, is technology that converts spoken audio into written text, and modern AI models do it across many languages and accents.
Superintelligence: Superintelligence is a hypothetical AI that would greatly outperform humans across virtually all areas of thinking, far beyond today's models, and it is a goal and a concern for some labs rather than an existing system.
Supply chain attack: A supply chain attack compromises software by tampering with something its builders trust, such as a code library, update or contributor account, so the malicious code reaches many users through a normal install.
SWE-bench: SWE-bench is a coding benchmark that asks an AI system to fix real software bugs taken from GitHub issues, and counts a fix as correct only if the project's own tests then pass.
Synthetic data: Synthetic data is information generated by a computer, often by another AI model, rather than collected from real people or events, and it is used to train, test or fine-tune AI models.
System card: A system card is a document an AI lab publishes with a model release describing what it can do, how it was tested, and the safety risks and limits found before launch.

T

Terminal-Bench: Terminal-Bench is a benchmark that tests AI agents on realistic command-line tasks, where each task runs in its own terminal environment and is graded by automated tests.
Test-time compute: Test-time compute is the extra computing an AI model spends while answering a question, such as thinking longer or trying several approaches, instead of during training, to get a better result on hard problems.
Text-to-speech: Text-to-speech (TTS) is technology that turns written text into spoken audio, and modern AI voices can sound natural, adjust tone and speak many languages.
Token: A token is the small chunk of text, often a word or part of a word, that an AI model reads and writes, and the unit that limits and prices most AI usage.
Tokenizer: A tokenizer is the component that splits text into tokens, the chunks an AI model actually reads, and it decides how many tokens a given piece of text costs.
Tool use: Tool use, also called function calling, is when an AI model asks the software around it to run an action, such as a search or a calculation, and then uses the result in its answer.
Training data: Training data is the collection of text, images, code, audio or other examples an AI model learns from, and it shapes what the model knows, how it writes and where it makes mistakes.

V

Vibe coding: Vibe coding is building software by describing what you want to an AI in plain language and accepting the code it writes, often without reading it closely, then fixing problems by asking for changes.
Voice agent: A voice agent is an AI system you talk to out loud that listens, decides what to do and answers in speech, often handling tasks like phone calls, bookings or support.
VRAM: VRAM is the dedicated memory on a graphics card, and for AI it sets the limit on how large a model you can run locally, since the model's weights and working data must fit in it.

W

Web crawler: A web crawler is an automated program that visits web pages and follows links to read and collect their content, used by search engines and by AI companies gathering data to train or ground models.
Web scraping: Web scraping is using software to automatically copy content from websites at scale, and AI companies use it to gather text and data for training models and for answering questions.
Word error rate: Word error rate (WER) is the standard score for speech-to-text accuracy: the share of words a system gets wrong compared with a correct transcript, counting substitutions, deletions and insertions, so lower is better.
World model: A world model is an AI system that learns how an environment behaves so it can predict what happens next and simulate it, such as generating interactive video, game scenes or robot training scenarios.

Z

Zero data retention: Zero data retention (ZDR) is an arrangement with an AI provider where your prompts and the model's outputs are not stored after the request is handled, usually approved per customer and limited to certain features.