For educators · AI vocabulary

AI words, in plain language

264 terms in two parts: the words you will meet on ARQIV, then the wider world of AI, from agents and superintelligence to the chips and data centers that make it run. No jargon used to explain the jargon.

Grades 6–12 teachers264 termsUpdated November 2026

How to use it

Look for the dotted underline

On ARQIV pages, AI words have a dotted underline. Hover over one, tap it or press Enter on it to see a short definition without leaving the page.

Each entry below has a fuller explanation and a line you can say to students.

Browse the words

264 terms in 13 topics.

All 264 terms, A to Z

Part 1 · 168 terms

Words you’ll meet on ARQIV

The vocabulary of our lessons, labs and explainers, from the basics to how machines learn.

01 · Topic

The basics

The words almost every AI conversation uses, plus the everyday jobs AI does, such as seeing, listening and translating.

Artificial intelligence (AI)

Also: AI

Computer systems designed to do tasks we usually link to human thinking, such as recognizing patterns, using language, making predictions or choosing actions.

AI is an umbrella term that covers many different technologies. Much of today’s AI is machine learning: systems that learn patterns from examples instead of following rules a person wrote by hand. Other AI uses explicit rules, search or planning, like a chess program that looks ahead through possible moves or a map app that plans a route, and many real systems combine both. A chatbot, a spam filter and a route planner are all AI, but they work very differently.

Say it to studentsAI is software designed to do tasks such as recognizing patterns, using language or choosing actions. Some AI learns from examples.

Algorithm

A step-by-step set of instructions for solving a problem or doing a task. A recipe is an algorithm for cooking.

All software runs on algorithms. People also use the word loosely for any system that decides what you see, like the recommendations in a video feed. In machine learning, a learning algorithm builds the model, but what the model ends up doing is shaped by the data it learned from.

Say it to studentsAn algorithm is a recipe a computer follows.

Machine learning

Also: ML

A way of building AI where the computer learns patterns from examples (data) instead of being given every rule by a person.

Instead of writing rules like “if an email contains these words, it is spam,” people show the system thousands of labeled examples and it adjusts itself until its guesses match. The result is only as good as the examples: gaps or errors in the data become gaps or errors in the model.

Say it to studentsMachine learning is learning by example, like learning what a dog is by seeing lots of dogs.

Model (AI model)

The trained program that does the AI’s work. It takes an input, like a question or a photo, and produces an output, like an answer or a label.

A model is what is left after training: a very large set of numbers that captures patterns in the training data. ChatGPT, Claude and Gemini are apps built on models. Companies release new models often, so a claim about “the AI” is really a claim about one model at one point in time. In science class, “model” can also mean a physical or mathematical model; in AI it means the trained program.

Say it to studentsThe model is the part that learned from all the examples. The app is just the window you talk to it through.

Training data

Also: Training set, Dataset

The examples a model learns from: text, images, sounds or numbers. What is in the data, and what is missing, shapes what the model can do.

Chatbots are trained on huge amounts of text, much of it from the public internet, plus books, code and other sources. If a group of people, a language or a topic is under-represented in the data, the model usually does worse on it. Training data also stops at a cutoff date, so a model may not know about recent events unless it can search the web.

Say it to studentsTraining data is everything the AI studied. It can only learn from what it was shown.

Neural network

Also: Neural net

A kind of AI model made of many layers of simple connected math units, loosely inspired by brain cells. Most modern AI uses them.

Each unit takes in numbers, weighs them and passes a result to the next layer. During training, the connection strengths are nudged again and again until the network’s outputs get closer to the right answers. The brain comparison is loose: neural networks do not think or feel like people.

Say it to studentsA neural network is a huge web of tiny number-crunchers that get tuned until they give good answers.

Deep learning

Machine learning that uses neural networks with many layers. It powers chatbots, image generators, voice assistants and translation apps.

“Deep” refers to the number of layers. It says nothing about how well the model understands anything. Deep learning took off in the 2010s, when large datasets and fast graphics chips (GPUs) made it practical to train big networks.

Say it to studentsDeep learning is machine learning with many layers stacked up.

Parameters (weights)

Also: Weights

The adjustable numbers inside a model that get tuned during training. Big chatbot models have billions of them.

Parameters are where a model stores what it learned. More parameters can let a model capture more patterns, but size alone does not guarantee quality: data, training methods and testing matter too. When a company releases a model’s weights, it is sharing these numbers so others can run the model themselves. Each artificial neuron usually has one more learned number, called its “bias”: a simple offset added to its total. That is a different meaning from bias as unfairness.

Say it to studentsParameters are billions of tiny dials that training turns until the answers come out right.

Computer vision

Also: Image recognition

The part of AI that gets useful information from images and video, such as recognizing faces, reading signs or telling a weed from a crop.

Computer vision covers tasks like naming what is in a whole photo (“cat”), finding and boxing each object in it, and outlining exact regions, such as a tumor in a scan. Most modern computer vision uses deep learning trained on labeled images. Matching pixel patterns is not the same as understanding a scene: a system can score well on test photos and still fail with unusual lighting, angles or skin tones.

Say it to studentsComputer vision is how computers find patterns in pictures, like spotting a weed or reading a sign.

Object detection and segmentation

Also: Image segmentation, Bounding box

Computer vision tasks that find where things are in an image: detection draws a box around each object, segmentation outlines the exact pixels that belong to it.

Classifying a photo answers “what is in it?” Detection answers “what, and where?”: a driver-assistance camera boxes each pedestrian, bike and stop sign in every frame. Segmentation goes further and labels every pixel, for example outlining a tumor on a scan. A common confusion is that a box means the system understands the scene. It means a pattern matched with some confidence, and small, partly hidden or unusual objects are often missed.

Say it to studentsDetection puts a box around each thing it finds. Segmentation colors in its exact shape.

Optical character recognition (OCR)

Also: OCR, Text recognition

AI that turns text in an image, such as a scanned page, a photo of a worksheet or a street sign, into editable, searchable text.

OCR is what lets a phone copy text from a photo, a library make old newspapers searchable or a chatbot read a picture of handwritten homework. It first finds where the text is, then recognizes the letters and words, often using a language model to guess unclear ones. A common confusion is that OCR gives an exact copy: smudges, handwriting, unusual fonts and tables still cause errors, such as a lowercase “l” read as the number “1,” so check names and numbers against the original.

Say it to studentsOCR is how a computer reads the words in a picture. It still misreads messy handwriting.

Natural language processing (NLP)

Also: NLP

The part of AI that works with human language: translating, summarizing, answering questions, sorting messages or turning speech into text.

NLP is much older than chatbots: spell checkers, spam filters and early translation tools all used it, at first with hand-written rules and word counts. Today most NLP uses large language models. Fluent language is not proof of understanding, which is why a chatbot can write smoothly and still be wrong.

Say it to studentsNatural language processing is how computers work with the words people speak and write.

Speech recognition

Also: Speech-to-text, Automatic speech recognition (ASR)

AI that turns spoken words into written text. It powers voice assistants, live captions, dictation and the voice mode of chatbots.

A speech recognition model listens to short slices of sound and predicts the most likely words, using context to choose between sound-alikes such as “write” and “right.” Live captions in a video call are a familiar example. A common confusion is that it understands what was said: it transcribes, and anything after that, like answering, is another system’s job. Accuracy often drops with accents, background noise, children’s voices and specialist words, so check captions before relying on them.

Say it to studentsSpeech recognition turns talking into typing. It can mishear, especially with noise or accents.

Machine translation

Also: Neural machine translation

Using AI to translate text or speech from one language into another, as translation apps and chatbots do.

Modern translation systems learn from millions of sentence pairs, such as the same document published in English and Spanish; the transformer was first designed for exactly this task. They work best for language pairs with lots of examples and worse for languages with little text online. A common confusion is that fluent output is faithful output: a smooth sentence can still drop a “not,” miss an idiom or get the tone wrong, so important documents need a check by a person who knows both languages.

Say it to studentsMachine translation is fast and useful, but a smooth sentence can still have the wrong meaning.

Sentiment analysis

Also: Opinion mining

Using AI to judge whether a piece of text sounds positive, negative or neutral, such as sorting product reviews or survey comments.

A sentiment classifier is trained on examples labeled by feeling, then scores new text: a school could sort 2,000 open survey answers into happy, unhappy and mixed in seconds. A common confusion is that it reads emotions or intentions. It matches word patterns, so sarcasm (“Great, another test”), slang, mixed feelings and less common dialects often get the wrong label. It works for a rough picture across many comments. Don’t use it to judge one person.

Say it to studentsSentiment analysis guesses whether writing sounds happy or unhappy. Sarcasm fools it easily.

Symbolic AI

Also: Rule-based AI, Expert system

AI that follows rules and symbols people write down, such as “if the soil is dry and no rain is coming, water the field,” instead of learning patterns from data.

Symbolic AI led the field from the 1950s to the 1980s. Expert systems stored hundreds of if-then rules collected from specialists in medicine or engineering. Rules are easy to inspect and explain but very hard to write for messy tasks, like recognizing a face. Many systems today mix both: a learned model plus hand-written rules and checks.

Say it to studentsSymbolic AI follows rules people wrote down. Machine learning finds its own patterns in examples.

Knowledge graph

A network of facts stored as things and the links between them, such as “Paris is the capital of France,” that a computer can search and follow step by step.

Each fact is a small triple: a thing, a relationship and another thing (Marie Curie, won, Nobel Prize). Linked together, millions of facts let a search engine show a fact box or answer “Which scientists won two Nobel Prizes?” by following the links. Unlike a language model, every fact can be inspected, traced to a source and corrected, but the graph only knows what someone put in it. A common confusion is that “graph” means a chart: here it means a network of dots and connecting lines.

Say it to studentsA knowledge graph is a giant web of facts, with dots for things and lines for how they are connected.

Neuro-symbolic AI

Also: Neurosymbolic AI, Hybrid AI

AI that combines neural networks, which learn patterns from data, with symbolic methods, which follow explicit rules and logic, to get some strengths of both.

A homework helper might use a neural network to read a handwritten equation and a rule-based algebra solver to work it out step by step, so each step can be checked. Neural networks handle messy input like photos and speech; symbols make reasoning easier to inspect and explain. Many real systems already mix the two loosely, such as a chatbot that calls a calculator, but how best to combine them more deeply is an open research question. A common confusion is that it is a new kind of brain-like computer: it is a design approach, and no single product or chip owns it.

Say it to studentsNeuro-symbolic AI teams up a pattern-spotter with a rule-follower, so the AI can both recognize things and reason step by step.

Search and planning

Also: Heuristic, Pathfinding

AI that finds a good sequence of actions by exploring possible next steps, like a route finder comparing paths or a chess program looking several moves ahead.

A search program treats a problem as states (where things are now) and actions (what can happen next), then explores until it reaches a goal. There are usually far too many options to try them all, so it uses a heuristic, a rule of thumb like “head roughly toward the destination,” to decide what to try first. Search needs no training data, though it is often combined with machine learning, as in game-playing programs like AlphaGo.

Say it to studentsSearch and planning is how a computer tries out possible moves in its head before it picks one.

02 · Topic

How chatbots and generators work

What is going on when a chatbot writes or an app makes a picture.

Generative AI

Also: GenAI

AI that creates new content, such as text, images, audio, video or code, from patterns learned in its training data.

Older AI mostly sorted or predicted things, like flagging spam. Generative AI produces something new each time, which is why two people asking the same question can get different answers. New does not mean true. The output is a likely-looking result that nobody has checked.

Say it to studentsGenerative AI makes new things, like stories or pictures, by remixing patterns it learned.

Chatbot

Also: AI assistant

An app you talk to in everyday language, such as ChatGPT, Claude or Gemini. Behind it is a large language model.

A chatbot sends your message, the conversation so far and hidden instructions from the company to a model, then shows you the model’s reply. Many chatbots can also search the web, read files or use other tools. Most have minimum ages in their terms of use, so check them before using one with students.

Say it to studentsA chatbot is a program you can type or talk to that writes back.

Large language model (LLM)

Also: LLM, Language model

A model trained on huge amounts of text to predict what comes next. It is the engine inside chatbots like ChatGPT, Claude and Gemini.

By learning to predict the next piece of text across billions of examples, an LLM picks up grammar, facts, styles of writing and some reasoning patterns. It does not look answers up in a database unless it is connected to a search tool, which is one reason it can state wrong things confidently.

Say it to studentsA large language model is a word-predicting machine that read a gigantic library.

Token

Also: Tokenizer

A small chunk of text, often a word or part of a word, that a language model reads and writes one at a time.

Models do not see letters or whole sentences; they see tokens. “Unbelievable” might be split into “un,” “believ” and “able.” Tokens explain some odd mistakes, like miscounting the letters in a word, and they are how AI companies measure usage and set limits. In AI, a token is a piece of text. It has nothing to do with game tokens or digital coins.

Say it to studentsTokens are the puzzle pieces the AI breaks words into.

Next-word prediction

Also: Next-token prediction

The core job of a language model: look at the text so far, guess which token most likely comes next, add it, and repeat.

A chatbot writes its whole answer this way, one token at a time. “Most likely” is not the same as “true”: a fluent, confident sentence can still be wrong. If students remember one idea about chatbots, this is the one to pick.

Say it to studentsThe chatbot is playing a super-powered game of “guess the next word.”

Autoregressive model

A model that generates something one piece at a time, with each new piece based on all the pieces before it. Chatbots write text this way.

The model feeds its own output back in as input for the next step: a language model writes “The cat sat” one token at a time, and some image and audio generators work the same way with small patches or snippets of sound. Because it does not go back and revise, an early wrong turn can steer everything after it. A common confusion is that “regressive” means going backward or getting worse: it only means each step is predicted from the earlier ones.

Say it to studentsAn autoregressive model builds its answer like a train, adding one car at a time to the end of what it has so far.

Transformer

The model design behind today’s chatbots, introduced by Google researchers in 2017. It uses attention to work out how the words in a passage relate.

The 2017 paper was called “Attention Is All You Need.” During training, a transformer handles every position in a passage at the same time instead of one word after another, which made it fast to train on huge datasets. That does not let a chatbot peek ahead: in a language model, a mask lets each position look only at earlier tokens, and when it writes, it still adds one token at a time. The “T” in GPT stands for transformer, and the same design is now used for images, audio and video.

Say it to studentsA transformer is the design chatbots are built from. When it writes, it can only look back at the words so far, never ahead.

Attention (in AI)

Also: Self-attention

The part of a transformer that lets each word “look at” the other words, so the model can tell what a word means in context.

In “The bat flew out of the cave,” attention helps the model link “bat” with “flew” and “cave,” so it reads “bat” as the animal. Despite the name, it is a calculation: a score for how relevant each word is to each other word.

Say it to studentsAttention is how the AI decides which other words matter for understanding each word.

Mixture of experts (MoE)

Also: MoE, Sparse model

A model design that splits the network into many smaller parts, called experts, and uses only a few of them for each token, saving computing power.

A small “router” inside the model picks which experts handle each token, so a model with hundreds of billions of parameters may use only a fraction of them for any one word. That is why some announcements give two sizes: total parameters and active parameters. A common confusion is that each expert is a specialist in a school subject, like a math expert and a history expert. The split is learned during training and rarely lines up with topics people would recognize.

Say it to studentsA mixture of experts wakes up only a few parts of the model for each word, which makes it cheaper to run.

Context window

Also: Context length

How much text a model can take in at once: your prompt, the chat so far and any files. Anything beyond it, the model cannot see.

Context windows are measured in tokens. In long chats, early details can drop out, and models can miss things buried in the middle of very long inputs. Starting a fresh chat for a new task often gives better results.

Say it to studentsThe context window is the AI’s short-term memory for this conversation.

Context versus memory

Also: Chatbot memory, Saved memories

Context is what a model can see right now, in this chat. Memory is information an app saves and adds back into later chats. The model itself remembers nothing.

The context is everything in the current context window: your messages, any uploaded files and the replies so far. It lasts only for that conversation. Some chatbots offer a memory feature: the app stores notes such as “teaches 8th-grade science” and quietly adds them to the context of future chats. A common confusion is that the model learns from each chat as you talk. Its parameters do not change; it only “remembers” what the app retrieves and puts back in front of it. Check what a tool saves, and turn memory off for student work if needed.

Say it to studentsThe AI does not remember you. The app can save notes and show them to the AI next time.

Knowledge cutoff

Also: Training cutoff

The date when a model’s training data ends. The model knows little or nothing about events after it unless it can search the web.

A model released in spring may have a cutoff months earlier, so its sense of “recent” is already out of date. Chatbots with web search can look up newer facts, but the model itself has not learned them. A common confusion is that a confident answer about current events is up to date: ask the chatbot for its sources and dates.

Say it to studentsA knowledge cutoff is the date the AI stopped learning. After that, it doesn’t know what happened.

Prompt

Also: Prompting

What you type or say to an AI to get a response: a question, an instruction, or examples of what you want.

Clear prompts give the task, the audience, the format and any limits, for example “Explain photosynthesis to a 9-year-old in three sentences.” Better prompts help, but no prompt guarantees a correct answer, so the output still needs checking.

Say it to studentsA prompt is the instructions you give the AI. Clear instructions, better results.

System prompt and instruction hierarchy

Also: System instructions, Instruction hierarchy

Instructions an app gives a chatbot before your message, usually hidden from you, setting its role, tone and rules. Models are trained to rank them above user requests.

A tutoring app might start every chat with “You are a math tutor for grade 7. Give hints, not final answers.” The instruction hierarchy is the order of trust a model is trained to follow: the app’s system prompt first, then the user’s messages, then text it reads in web pages, files or tool results, which should count only as information. A web page or file cannot give the model orders. A common confusion is that a system prompt is a lock. It steers behavior but can be leaked or talked around, which is why prompt injection is a risk.

Say it to studentsThe system prompt is the app’s behind-the-scenes rule sheet. It steers the AI but cannot guarantee what it does.

Context engineering

Choosing and arranging everything a model sees, not just the prompt: instructions, examples, documents and tool results, so it has what it needs and nothing that distracts it.

As chatbots and agents handle longer tasks, what goes into the context window matters more than clever wording. Too little information and the model guesses; too much and important details get lost. A common confusion is that more context is always better: focused, relevant material usually gives better answers than everything at once.

Say it to studentsContext engineering means giving the AI the right information along with the right question.

Temperature and top-p

Also: Sampling settings, Top-p (nucleus sampling)

Settings that control how a model picks each next token. Low temperature sticks to the likeliest words; higher temperature or a larger top-p allows less likely ones, adding variety.

At each step a model gives every possible next token a probability, then samples one. Temperature reshapes those odds: near 0, the same prompt gives nearly the same answer every time; higher, a poem prompt gets more surprising word choices, and more slips. Top-p keeps only the likeliest tokens whose chances add up to p, such as 0.9, and samples from those. Despite the name, it is a sampling setting and has nothing to do with heat. A common confusion is that a low temperature makes answers true: it makes them consistent, and a consistent answer can still be wrong.

Say it to studentsTemperature is the creativity dial: low for steady answers, high for surprises. Neither makes the answer true.

Reasoning model

Also: Thinking model, Thinking mode

A model trained to spend extra computation working through a problem in steps before it answers. Often slower, but stronger at math, science and coding.

Examples include OpenAI’s o-series, DeepSeek-R1 and the “thinking” modes in other chatbots. The extra steps often help, but a step can still go wrong. Some apps show the steps or a summary of them and some show nothing, and what is shown is not always a faithful record of how the model actually reached its answer.

Say it to studentsA reasoning model takes extra time to work through steps before it answers. It can still make mistakes, and any steps it shows may not be the whole story.

Chain of thought

Also: Step-by-step reasoning

Step-by-step reasoning a model writes out before its final answer. Asking for it, or training models to do it, often improves answers on hard problems.

Researchers at Google described chain-of-thought prompting in 2022, and today’s reasoning models are trained to produce long chains of thought on their own. Writing the steps helps the model break a problem down, much as it helps students. A common confusion is that the steps show exactly how the answer was reached: they can be incomplete or even unrelated to the real computation, so check the answer, not just how convincing the steps sound.

Say it to studentsChain of thought is the AI showing its work, step by step. The work can still be wrong.

Multimodal

Able to work with more than one kind of input or output, such as text, images, audio and video.

A multimodal chatbot can describe a photo of a science fair poster, read a chart or answer out loud. Photos and voice recordings of students are personal data, so check your school’s privacy rules before uploading them.

Say it to studentsMultimodal means the AI can work with pictures and sound as well as words.

Image generator

Also: Text-to-image

AI that makes pictures from a description. Different tools use different methods, such as diffusion models or models that build an image piece by piece.

Diffusion models start from random noise and clean it up step by step. Other generators build an image piece by piece, much like a language model writes text, and older systems used GANs, where two networks compete. Whatever the method, image generators learn from huge collections of captioned images, so they repeat the patterns and stereotypes in that data, for example in who appears when you ask for “a scientist.” They can also struggle with text inside images, hands and counting objects.

Say it to studentsAn image generator turns your words into a picture, using what it learned from millions of captioned images.

Diffusion model

Also: Diffusion

A way of generating images, audio or video: start from random noise and remove it step by step, guided by a description, until a clear result appears.

In training, the model sees real images with more and more noise added and learns to predict the noise so it can take it away. To generate, it runs that process in reverse, starting from pure static. Many image and video generators use diffusion, but not all: “diffusion model” names a method, while “image generator” names what the app does.

Say it to studentsA diffusion model turns static into a picture, one small clean-up step at a time, guided by your words.

Generative adversarial network (GAN)

Also: GAN

A way of generating images or other data in which two networks compete: a generator makes fakes and a discriminator tries to spot them, so both keep improving.

It is like a forger and an art inspector training each other: the forger gets better because the inspector keeps catching the fakes. GANs were introduced by Ian Goodfellow and colleagues in 2014 and powered many early realistic face generators and deepfakes. They make an image in a single step, but training is famously unstable, and a GAN can end up producing only a few kinds of output, a problem called mode collapse. A common confusion is that “adversarial” means it attacks people: the contest is between the model’s own two parts, which is a different idea from an adversarial example.

Say it to studentsA GAN is two AIs playing a game: one makes fakes and the other tries to catch them, so both get better.

Variational autoencoder (VAE)

Also: VAE, Autoencoder, Latent space

A model that learns to squeeze data, such as images, into a short list of numbers and rebuild it, so new examples can be made by decoding new lists.

An encoder compresses a face photo into a short list of numbers (a point in “latent space”), and a decoder turns those numbers back into a face. Because the VAE learns a smooth space, points between two faces decode to a blend, and new points decode to new faces. VAEs were introduced by Kingma and Welling in 2013; their images tend to be blurrier than other methods, but the same squeeze-and-rebuild idea is used inside some diffusion image generators to save computing. A common confusion is that it stores the original pictures: it keeps a compressed description and rebuilds an approximation.

Say it to studentsA VAE learns to pack a picture into a few numbers and unpack it again, so new numbers can make new pictures.

Flow-based model

Also: Normalizing flow, Flow matching

A generative model that turns random noise into realistic data through steps that can be run exactly backward, so it can also measure how likely an example is.

Think of reshaping a ball of clay: each step stretches or bends it, but nothing is torn, so every step can be undone. Classic “normalizing flows” use that reversibility to calculate the exact probability of any example, something a GAN cannot do. A related training method called flow matching learns a smooth path from noise to data and is used in some image and audio generators. A common confusion is that “flow” means a stream of data or the “flow” of focused work: it means the gradual, reversible reshaping of noise into data.

Say it to studentsA flow model reshapes random noise into a picture through steps you can always undo, like shaping clay without tearing it.

Retrieval-augmented generation (RAG)

Also: RAG

A way to have a chatbot answer from chosen sources: first search a set of documents for relevant passages, then give them to the model to write the answer.

A school might connect a chatbot to its handbook. When someone asks a question, the system retrieves the most relevant chunks of the handbook and adds them to the prompt. This helps answers stay current and point to sources, but it does not guarantee accuracy: the search can pick the wrong passage, and the model can still misread it or add things that are not there.

Say it to studentsRAG means the AI looks things up first and then answers, and we still check what it found.

Chunking (in retrieval)

Also: Chunks

Splitting long documents into smaller pieces, such as paragraphs or sections, so a search system can find and pass on just the relevant parts.

A 200-page manual is too long to send to a model with every question, so a RAG system cuts it into chunks and indexes each one. Size matters: too small and a piece loses its context, too large and the useful sentence gets buried. In teaching, “chunking” means breaking a lesson into smaller steps, a related idea with a different purpose.

Say it to studentsChunking is cutting a big document into bite-size pieces so the right piece can be found.

Grounding

Also: Grounded answer

Tying an AI’s answer to specific evidence, such as provided documents or search results, so each claim can be traced back and checked.

A grounded answer says, in effect, “according to page 12 of the handbook…”, and the reader can check page 12. Grounding lowers the chance of made-up facts but does not remove it: a model can cite a real source that does not say what it claims. The word has other meanings too, such as an electrical ground.

Say it to studentsA grounded answer shows where it came from, so you can go and check.

Vector database

Also: Vector index, Vector store

A store for embeddings that can quickly find the ones closest to a new query. It is the lookup step behind many RAG systems.

Each chunk of a document is saved with its embedding, a vector of hundreds of numbers. When a question comes in, the database finds the nearest vectors, usually with fast approximate methods instead of checking every one. What it stores are numbers that stand for meaning; it has no understanding of the facts behind them.

Say it to studentsA vector database is like a library catalog that shelves things by meaning instead of by title.

Tool call

Also: Function calling, Tool use

When a model asks an app to do something for it, such as run a search, do a calculation or check a calendar, by sending a structured request.

The model does not run the tool itself. It writes a request in a set format, like “search: rainfall in Kenya 2024,” the app runs it and sends the result back, and the model uses that result in its reply. Tool calls make chatbots more useful and are the basis of AI agents, but a model can still pick the wrong tool or misread the result.

Say it to studentsA tool call is the AI asking another program for help, like a calculator or a search engine.

API and API key

Also: Application programming interface

An API is an agreed way for one program to ask another for data or work. An API key is a secret code that shows who is asking.

When a school app “uses ChatGPT,” it usually sends requests to the company’s API rather than going through the chat website. The key works like a password and is often tied to billing, so it should never be pasted into shared documents, code students can see or a public web page. An API is how programs talk to each other; it is not itself AI.

Say it to studentsAn API is how one program asks another for help. An API key is its password, so keep it secret.

03 · Topic

How models are made and tested

Words from AI company announcements, news stories and research.

Fine-tuning

Extra training that adapts an already-trained model to a narrower job, style or set of rules.

Training a big model from scratch is very expensive. Fine-tuning starts from an existing model and trains it a little more on a smaller, focused dataset, for example on examples of good tutoring conversations.

Say it to studentsFine-tuning is extra practice that makes a general AI better at one particular job.

RLHF (learning from human feedback)

Also: Reinforcement learning from human feedback

Reinforcement learning from human feedback: people rate a model’s answers, and the model is trained to give more of the kind they prefer.

RLHF is a big reason chatbots sound helpful and polite. It can also teach a model to please people, for example by agreeing too readily, which is one cause of sycophancy.

Say it to studentsPeople give the AI thumbs up or thumbs down, and it learns to aim for thumbs up.

Benchmark

Also: Leaderboard

A standard test used to compare AI models, such as a set of exam questions or coding tasks with known answers.

Benchmarks make comparisons possible, but they have limits: test questions can leak into training data, and a high score on one test does not mean a model is good at your task. Treat leaderboard claims as a starting point, then test the model yourself.

Say it to studentsA benchmark is a test that every AI takes so we can compare their scores.

Evaluation (eval)

Also: Eval

A planned test of how well an AI does a specific job, using the same prompts, clear criteria and a record of the results.

Evals range from big industry tests to a simple classroom one: give two models the same ten questions with known answers and score them fairly. Running a small eval is one of the best ways for students to see what AI can and cannot do.

Say it to studentsAn eval is a fair test we design ourselves to see if the AI is any good at a job.

Quantization

Shrinking a trained model by storing its numbers with less precision, so it needs less memory and runs faster, usually with only a small loss in quality.

Each parameter is normally stored as a number that uses 16 or 32 bits; quantization rounds it to, say, 8 or 4 bits. It is like rounding prices to the nearest dollar: the total is a little off but much quicker to work with. A 4-bit copy of a model takes about a quarter of the memory of a 16-bit copy, which is how some models fit on a laptop or phone. A common confusion is that it removes parts of the model: it keeps every parameter and stores each one less precisely, and rounding too hard makes answers noticeably worse.

Say it to studentsQuantization rounds off the model’s numbers so it fits on a smaller device, like rounding 3.14159 to 3.1.

Distillation (knowledge distillation)

Also: Knowledge distillation, Teacher and student models

Training a small “student” model to copy the answers of a large “teacher” model, so the small one gets much of the skill at a fraction of the cost.

The student learns not just the right answer but the teacher’s full set of probabilities, such as “80% cat, 15% fox, 5% dog,” which carry extra hints about which things look alike. It is like learning from an expert’s worked solutions instead of only the answer key. The idea was popularized by Hinton and colleagues in 2015. A common confusion is that the student ends up as good as the teacher: it usually keeps most of the skill on the tasks it was trained for and loses some elsewhere.

Say it to studentsDistillation is a big AI teaching a small AI, so the small one can do the job faster and cheaper.

MLOps (machine learning operations)

Also: Model versioning, Model monitoring, Rollback

The everyday work of keeping a machine learning model useful after it is built: tracking versions, testing changes, releasing them safely, watching for problems and rolling back.

Building a model is often the smaller part of the job. A school’s attendance-forecasting tool needs a record of which data and settings made each version, checks before a new version replaces the old one, alerts when its scores slip (model drift) and a quick way to switch back if an update goes wrong, called rollback. MLOps borrows these habits from software teams and adds data and models to what they track. A common confusion is that a model is finished once it is trained: in real use it needs ongoing care, more like a car than a painting.

Say it to studentsMLOps is looking after an AI once people are using it: keeping track of versions, checking it still works and fixing it when it slips.

Reproducibility

Also: Reproducible results

Whether someone else, or you later, can rerun the same work with the same data, code and settings and get the same result.

In AI, that means recording the exact data, code, model version and settings, and even the random starting point (the “seed”), because training and sampling involve chance. It is like a recipe that lists every ingredient and oven temperature, so another cook can bake the same cake and check it. Without it, nobody can check a claim that a new model scores higher. A common confusion is with replication: reproducing reruns the same study, while replicating tests the same idea again with new data, and science needs both.

Say it to studentsReproducible means someone else can follow your exact steps and get the same result.

Open-weight model

Also: Open model

A model whose trained parameters (weights) are published, so anyone can download it and run or adapt it on their own computers.

Examples include Meta’s Llama family, DeepSeek’s models and Alibaba’s Qwen family. Open-weight is not always fully “open source”: the training data and code are often not shared. Running a model on your own device can keep data from leaving it.

Say it to studentsAn open-weight model is an AI whose insides are shared, so anyone can run their own copy.

Frontier model

Also: Frontier lab

One of the newest, most capable AI models at a given moment, usually from a large lab such as OpenAI, Anthropic or Google.

“Frontier” is a moving target: today’s frontier model is ordinary within a year or two. Frontier labs are the companies that build them. Claims about what “AI can do” usually mean what frontier models can do.

Say it to studentsFrontier models are the newest, most powerful AIs right now.

04 · Topic

Data and the numbers underneath

How examples, pictures and words become numbers a model can work with, and how to read data carefully.

Feature (in machine learning)

Also: Input variable, Attribute

One measurable piece of information a model uses as input, such as an animal’s weight, its number of legs or the brightness of one pixel.

In a table of examples, each row is one example and each column the model learns from is a feature. Choosing features matters: if a feature accidentally gives away the answer, or stands in for something unfair, as a zip code can stand in for race or income, the model learns that too. In everyday tech talk a “feature” is something an app can do, which is a different meaning.

Say it to studentsA feature is one clue the model looks at, like one column in a table.

Label (in machine learning)

Also: Target, Ground truth

The answer attached to a training example, such as “spam” or “not spam,” or a house’s sale price, that a supervised model learns to predict.

Labels often come from people, such as volunteers tagging photos or doctors marking scans, so they can contain mistakes and disagreements. If the labels are wrong or unfair, the model learns those errors. “Ground truth” sounds certain, but it is only as good as whoever made the labels.

Say it to studentsA label is the right answer we give the computer for each practice example.

Vector

An ordered list of numbers, such as [height, weight, age]. In AI, one example, one word or one image is often turned into a vector.

A single number is called a scalar; a vector is a list of them, where the order matters. A vector of 3 numbers can be pictured as a point in 3D space, and AI often uses vectors with hundreds or thousands of numbers, which work the same way even though nobody can picture them. In physics class a vector is an arrow with a size and a direction, a related but narrower idea.

Say it to studentsA vector is a list of numbers that describes one thing.

Matrix

Also: Matrices

A grid of numbers in rows and columns, like a spreadsheet that holds only numbers. A neural network layer’s weights are usually stored as a matrix.

A matrix with 3 rows and 4 columns has shape (3, 4). In a dataset, each row can be one example and each column one feature. Neural networks spend most of their time multiplying matrices, which is why chips that do many multiplications at once, such as GPUs, are so useful. A confusion matrix is one specific table of results; “matrix” here just means a grid.

Say it to studentsA matrix is a table of numbers with rows and columns.

Tensor

Also: Array

In AI software, a block of numbers with any number of axes: a single number, a list, a grid, or a stack of grids, such as a color photo.

A scalar (one number) has 0 axes, a vector has 1, a matrix has 2, and a color image is often a 3-axis tensor: height, width and 3 color channels. A batch of 32 such images adds a fourth axis. AI frameworks such as PyTorch and TensorFlow are built around tensors. Mathematicians and physicists use “tensor” in a stricter sense; in machine learning it mostly means a multi-dimensional array of numbers.

Say it to studentsA tensor is a box of numbers. A color photo is a stack of three number grids: red, green and blue.

Shape (of a tensor)

Also: Axis, Dimension

How many numbers a tensor holds along each axis. A 28 by 28 grayscale image has shape (28, 28); a batch of 64 of them has shape (64, 28, 28).

An axis is one direction you can step along: rows, columns, color channels and so on. Many errors in AI code are shape mismatches, like trying to combine a (3, 4) grid with a (5, 4) one. “Dimension” is used loosely for both the number of axes and the length of one axis: a “768-dimensional embedding” is a vector with one axis of length 768.

Say it to studentsShape tells you how many numbers a tensor holds in each direction.

Embedding

Also: Word embedding, Vector representation

A list of numbers a model learns to stand for something, such as a word, sentence or image, so that similar meanings end up with similar numbers.

Imagine placing every word on a giant map so that “puppy” lands near “dog” and far from “volcano.” An embedding is a spot on such a map, but with hundreds of directions instead of two. Embeddings power semantic search and are the first step inside a language model. Closeness comes from how words are used in the training data, stereotypes included.

Say it to studentsAn embedding turns a word into a list of numbers, like a spot on a map where similar words live close together.

Dot product

Multiply two equal-length lists of numbers position by position, then add up the results. AI uses this one number to score how well two lists line up.

For [1, 2, 3] and [4, 5, 6]: 1×4 + 2×5 + 3×6 = 32. Dot products are everywhere in AI: a neuron computes one between its inputs and its weights, attention uses them to score how related two tokens are, and semantic search uses them to compare embeddings. The result is a single number.

Say it to studentsA dot product multiplies matching numbers and adds them up into one score.

Base rate

Also: Prevalence

How common something is in the group you are looking at, before any test or AI is used. It changes how much a positive result means.

If 1 in 1,000 students has a rare condition, a screening tool that catches 99% of real cases and wrongly flags only 1% of everyone else will still raise about 10 false alarms for every real case it finds. Ignoring the base rate, called base-rate neglect, is a common mistake when people judge AI detectors, medical screens and fraud alerts.

Say it to studentsBefore trusting an alarm, ask how rare the thing is in the first place.

Probability distribution

Also: Distribution

A list of all the possible outcomes and how likely each one is, with the chances adding up to 100%.

A fair die has a simple one: each face has a 1 in 6 chance. A language model produces a probability distribution over its whole vocabulary at every step, for example “mat” 40%, “floor” 25% and “roof” 5% after “The cat sat on the,” then picks from it. A common confusion is reading the top probability as how often the model is right: it is the model’s own estimate, and it can be poorly calibrated. In “distribution shift,” the word means the overall pattern of the data.

Say it to studentsA probability distribution shows every possible result and how likely each one is.

Conditional probability

Also: Probability given

The chance of something being true given that you already know something else, such as the chance of rain given dark clouds.

It is written P(A | B) and read “the probability of A given B.” Order matters: the chance a student has a fever given that they have the flu is high, but the chance they have the flu given a fever is much lower, because many things cause fevers. Swapping the two is one of the most common errors in reading test results and AI alerts, and it is closely tied to ignoring the base rate. Next-word prediction is conditional probability too: the chance of each word given the words so far.

Say it to studentsConditional probability asks: now that I know this, how likely is that? The order of the question matters.

Bayesian reasoning

Also: Bayes’ rule, Bayes’ theorem, Prior and posterior

Updating how likely you think something is when new evidence arrives, by combining what you believed before (the prior) with how well the evidence fits.

Bayes’ rule, from the 1700s, turns this into math: P(A | B) = P(B | A) × P(A) ÷ P(B). If 1 in 100 emails is a scam and a filter flags 90% of scams but also 5% of normal emails, a flagged email has only about a 15% chance of being a scam, because normal emails are so common. Bayesian models keep track of how unsure they are instead of giving one answer, which helps in medical tests, spam filters and robots. A common confusion is that only the new evidence matters: the prior, often the base rate, matters just as much.

Say it to studentsBayesian reasoning means starting with your best guess and changing it a sensible amount when new clues show up.

Time series and forecasting

Also: Forecasting, Time-series data

A time series is a set of measurements taken in order over time, like daily rainfall. Forecasting uses past patterns in it to predict future values.

Weather, electricity use, school attendance and a star’s brightness are all time series. Forecasting models look for trends, seasons (attendance dips before holidays) and recent changes. Order matters when testing: train on the past and test on later dates, never on a random shuffle, or the model gets to peek at the future, a kind of data leakage. A common confusion is treating a forecast as one certain number. Good forecasts give a range, and they grow less sure the further ahead they look.

Say it to studentsForecasting uses the pattern so far to make a careful guess about what comes next, with a range.

Correlation versus causation

Also: Spurious correlation

Two things changing together (correlation) does not show that one causes the other (causation). Machine learning finds correlations; it does not prove causes.

Ice cream sales and sunburns rise together because both follow hot weather, not because ice cream causes sunburn. A model will use any pattern that helps it predict, including misleading ones, such as telling wolves from dogs by the snow in the background of the photos. Showing a cause usually needs an experiment or a carefully designed study.

Say it to studentsThings that happen together do not always cause each other.

Causal inference

Also: Seeing versus doing, Intervention

Methods for working out whether one thing actually causes another, and by how much, rather than only noticing that they change together.

The key difference is between seeing and doing: noticing that students who use flashcards score higher (seeing) is not the same as handing out flashcards and measuring what changes (doing, called an intervention). A randomized experiment is the most direct test; when one is impossible, researchers use careful study designs and cause-and-effect diagrams to rule out other explanations, such as keen students both using flashcards and studying more. Judea Pearl’s 2018 book “The Book of Why” made the seeing-versus-doing idea widely known. A common confusion is that a model that predicts well has found causes: good predictions can rest entirely on correlations.

Say it to studentsCausal inference asks “if we changed this, would that change too?” and not just “do they go together?”

Dataset card (and model card)

Also: Datasheet, Model card

A short document that describes a dataset or model: what it contains, how it was made, what it is meant for and its known limits.

Datasheets for datasets (Gebru and colleagues) and model cards (Mitchell and colleagues) were proposed in 2018 so builders would record facts like who is in the data, how it was labeled and where the model does worse. A card is only as honest and complete as its authors make it, and many datasets still have none. Reading one is a good habit before trusting a tool.

Say it to studentsA dataset card is the nutrition label for data: what is inside and what it is good for.

05 · Topic

How machines learn

Kinds of learning, classic model types and the words for what happens during training.

Supervised learning

Machine learning from examples that come with the right answers (labels), such as photos marked “cat” or “dog,” so the model learns to label new examples.

Spam filters, medical image classifiers and price predictors are usually trained this way. The two main kinds are classification (predict a category) and regression (predict a number). Supervised learning needs many labeled examples, and labels cost time and money, which is one reason other kinds of learning matter.

Say it to studentsSupervised learning is learning with an answer key.

Unsupervised learning

Machine learning that looks for structure in data without any labels, such as grouping similar shoppers or spotting unusual transactions.

No answer key is given: the algorithm finds groups (clustering), simpler summaries of many measurements (dimensionality reduction) or odd examples that stand out. The patterns it finds are not automatically meaningful, so a person still has to judge whether a group makes sense. It is different from self-supervised learning, which builds its own answer key from the data.

Say it to studentsUnsupervised learning is sorting without an answer key, like grouping buttons by whatever looks alike.

Self-supervised learning

Learning where the answer key comes from the data itself, such as hiding the next word of a sentence and training the model to predict it.

Language models are pretrained this way: every sentence becomes a practice question (“The cat sat on the …”) that carries its own answer. Nobody has to label anything, which is how models can learn from enormous amounts of text. Chatbots then get more training, such as fine-tuning and RLHF, so self-supervised pretraining is only the first stage.

Say it to studentsSelf-supervised learning is a model quizzing itself by covering up part of the data.

Semi-supervised learning

Learning from a small set of labeled examples together with a much larger set of unlabeled ones.

Labeling 1,000 X-rays by hand is expensive; collecting 100,000 unlabeled ones is easier. Semi-supervised methods use the labeled few to guide learning from the many, for example by letting the model label the easy cases itself and training on those. If its own guesses are wrong, it can reinforce its mistakes.

Say it to studentsSemi-supervised learning uses a few examples with answers to learn from many without.

Active learning

A way of training where the model picks which unlabeled examples a person should label next, choosing the ones it is most unsure about, to save labeling work.

Labeling 10,000 leaf photos by hand is slow and expensive. With active learning, a model trained on a few hundred asks an expert to label only the photos it finds most confusing, and it can often reach good accuracy with far fewer labels. The catch is that its questions reflect its own blind spots, so it may never ask about cases it is wrongly sure of. A common confusion is with “active learning” in education, where students learn by doing: in machine learning it means the model chooses what to learn from next.

Say it to studentsActive learning is when the AI asks, “Can you tell me what this one is?” about the examples that puzzle it most.

Reinforcement learning

Also: RL, Reward, Policy

Learning by trial and error: an agent takes actions in an environment, gets rewards or penalties, and gradually learns which actions earn more reward.

Game-playing programs like AlphaGo and robots learning to walk are classic examples. The behavior it learns is called a policy. The agent chases exactly the reward it is given, so a badly chosen reward can produce strange shortcuts, like a boat-racing agent that scored points by circling forever instead of finishing the race. RLHF is reinforcement learning where the reward comes from human ratings.

Say it to studentsReinforcement learning is learning by trying things and getting points for good moves.

Reward, policy and environment

Also: Reward function, Policy (in AI), Environment

Three parts of reinforcement learning: the environment is the world an agent acts in, the reward is its score, and the policy is its learned rule for choosing actions.

In a maze game, the environment is the maze, the reward might be 10 points for reaching the exit minus 1 for every step, and the policy is what the agent has learned to do in each square. Here “policy” is not a written rule or a school policy: it is a learned mapping from situations to actions. A common confusion is that the agent understands the goal. It only chases the reward it is given, so a badly designed reward can teach shortcuts nobody intended, called reward hacking.

Say it to studentsThe environment is the game, the reward is the score, and the policy is the strategy the AI learns.

Sim-to-real transfer

Also: Sim-to-real gap, Reality gap

Training a robot or other AI in a computer simulation and then moving what it learned to the real world, where it often works less well at first.

Simulation is fast, cheap and safe: a robot arm can practice grasping a million times overnight without breaking anything. But real objects slip, cameras catch glare and motors wobble in ways the simulation got slightly wrong, so skills can fail on the real machine; this is the sim-to-real gap. One common fix is to vary the simulation on purpose, with random lighting, weights and friction, so the robot learns to cope with whatever reality turns out to be. A common confusion is that a better-looking simulation is automatically a better one: what matters is whether its physics and sensors behave like the real ones.

Say it to studentsSim-to-real is when a robot practices in a video-game world and then has to do it for real, where things are messier.

Classification

Also: Classifier

A machine learning task where the model picks a category, such as “spam” or “not spam,” or “weed” or “crop,” often with a probability for each.

A classifier is trained on labeled examples and then sorts new ones. Many classifiers output a score, and a threshold turns the score into a decision, so moving the threshold trades false alarms for misses. Classification is not clustering: classification uses categories people chose in advance, while clustering finds groups on its own.

Say it to studentsClassification means sorting things into boxes we already named.

Regression (in machine learning)

A machine learning task where the model predicts a number, such as tomorrow’s temperature, a house price or a crop yield.

The model learns a relationship between inputs and a numerical output from labeled examples, then estimates the number for new cases. Its errors are measured in units, like “off by 2 degrees on average.” The name has nothing to do with the everyday meaning of going backward, and logistic regression, despite its name, is mostly used for classification.

Say it to studentsRegression means predicting a number, like a price or a temperature. Picking a category is called classification.

Clustering

Also: k-means

Grouping examples so that similar ones end up together, without being told what the groups should be.

A store might cluster shoppers by what they buy, or an astronomer might cluster stars by brightness and color. You often choose the number of groups, as in k-means, and the result depends on which measurements you use and how you define “similar.” Clusters are leads to investigate, not proof that natural categories exist.

Say it to studentsClustering is letting the computer sort things into piles that look alike.

Dimensionality reduction

Also: PCA

Squeezing many measurements per example into a few new ones that keep the biggest differences, often so the data can be drawn on a flat chart.

A dataset might record 50 measurements for each plant, far too many to picture. Methods such as PCA (principal component analysis) find a few combined directions that capture most of the variation, so similar examples land near each other on a plot. The new axes are mathematical blends, not real-world measurements, and distances and gaps on the picture can be distorted.

Say it to studentsDimensionality reduction is like drawing a flat map of the round Earth: useful, but something always gets stretched.

Anomaly detection

Also: Outlier detection

Finding data points that do not fit the usual pattern, such as an odd bank charge, a failing machine or a sudden dip in a star’s brightness.

A bank learns what your normal spending looks like and flags a purchase far from it for a check. Because problems are rare and varied, anomaly detection often learns only what normal looks like, an unsupervised approach, instead of learning from labeled examples of every kind of problem. A common confusion is that an anomaly means something is wrong: it only means unusual. Many flags turn out harmless, like a vacation purchase, so where the alarm threshold sits decides how many false positives people must check.

Say it to studentsAnomaly detection learns what normal looks like, then flags anything that looks strange.

Decision tree

A model that makes predictions by asking a series of yes-or-no questions about the input, like a flowchart, until it reaches an answer.

A tree for “Will it rain?” might ask “Is humidity above 80%?” and then “Is air pressure falling?” The questions are learned from data by picking, at each step, the split that best separates the examples. Trees are easy to read and explain, but one deep tree tends to memorize its training data, which is why random forests and gradient boosting combine many trees.

Say it to studentsA decision tree is a flowchart of questions the computer learned from examples.

Linear and logistic regression

Also: Linear model

Simple models that add up each input times a learned weight. Linear regression predicts a number; logistic regression turns the sum into a probability for a category.

Predicting a plant’s height from hours of sunlight with a best-fit straight line is linear regression. Logistic regression squeezes the same kind of weighted sum into a value between 0 and 1, such as the chance an email is spam. Both are fast, easy to explain and make strong baselines, but they miss patterns that are not roughly straight-line unless you add extra features.

Say it to studentsLinear regression draws the best straight line through the dots.

k-nearest neighbors (k-NN)

Also: kNN

A simple method that predicts by finding the k most similar examples it has stored and going with their majority answer or their average.

To guess whether a new fruit is an apple or a pear, k-NN finds, say, the 5 stored fruits with the closest weight and color and takes a vote. There is almost no training step: it keeps all the examples. Results depend on the choice of k and on how distance is measured, and a feature measured on a bigger scale can drown out the rest unless the features are rescaled.

Say it to studentsk-nearest neighbors asks: which examples I already know are most like this one, and what were they?

Random forest

A model made of many decision trees, each trained on a random sample of the data and features, that vote on the final answer.

One tree can latch onto quirks in its data; hundreds of slightly different trees make different mistakes, and their combined vote cancels many of them out. Random forests work well on spreadsheet-style data and need little tuning. The cost is that a forest is much harder to read than a single tree.

Say it to studentsA random forest asks hundreds of decision trees and goes with the crowd.

Gradient boosting

Also: Boosted trees, XGBoost

A method that builds many small decision trees one after another, each new tree trained to fix the errors the earlier ones still make.

A random forest grows its trees independently and combines them; boosting grows them in sequence, each focusing on what is still wrong. Libraries such as XGBoost and LightGBM make it a top performer on table-style data, like predicting crop yield from weather and soil records. Too many rounds can overfit, so the number of trees is checked on a validation set.

Say it to studentsGradient boosting is a team of small trees where each new member fixes the last ones’ mistakes.

Recommendation system

Also: Recommender system, Recommendation algorithm, Feed algorithm

Software that picks or ranks items for each person, such as the next video, song, product or post, based on what they and similar people did before.

A video app may weigh what you watched to the end, skipped, liked and searched for, then rank thousands of clips to fill your feed. Most systems combine collaborative filtering (people like you liked this) with item details (this is another cooking video). A common confusion is that recommendations show what you want or what is popular. They show what the system was built to maximize, often watch time or clicks, which is not the same as what is good for you.

Say it to studentsA recommendation system guesses what you will click next. Ask what it is trying to get you to do.

Collaborative filtering

A recommendation method that suggests items liked by people whose past choices resemble yours: if you and Sam liked the same five books, Sam’s sixth may suit you.

It needs no knowledge of what an item is about, only who chose what, which is why the same method works for songs, videos and shops. Many systems turn each person and each item into an embedding, so that people land near the items they tend to like. A common confusion is that someone decided what you would enjoy. No one did: the pattern comes from other users’ behavior, which also means popular items get recommended more and more while niche ones rarely surface.

Say it to studentsCollaborative filtering says: people who liked what you liked also liked this.

Cold-start problem

Also: Cold start

The trouble a recommendation system has with a brand-new user or item: with no history yet, it has little to go on.

A new music app knows nothing about you, and a song released today has no plays, so collaborative filtering has no pattern to use. Apps work around it by asking you to pick a few favorites, using item details like genre, or showing popular items first. A common confusion is that early recommendations reveal what the app thinks of you; mostly they are generic guesses that sharpen as you use it. New items, including work by new creators, can stay hidden longer, which is a fairness issue too.

Say it to studentsThe cold-start problem is the app not knowing anything about you yet, so it has to guess.

Activation function

Also: ReLU, Sigmoid, Nonlinearity

A simple rule inside each artificial neuron that decides what it passes on, such as “keep positive numbers, turn negatives into zero.” It lets networks learn curved, complicated patterns.

A neuron adds up its weighted inputs plus its bias, then applies the activation function. The rule above is called ReLU and is the most common today; older networks used an S-shaped curve called the sigmoid, which squeezes any number into the range 0 to 1. Without one, stacking layers would be pointless: many straight-line steps in a row still make one straight line. A common confusion is that “activation” means a neuron switches on like a light. Most just pass on a number, large or small.

Say it to studentsAn activation function is each neuron’s small rule for what to pass on. It lets the network learn bendy patterns as well as straight lines.

Convolutional neural network (CNN)

Also: CNN, ConvNet

A neural network built for grid-like data such as images. It slides small pattern detectors across the picture, finding edges first, then shapes, then whole objects.

Each small detector, called a filter, checks every patch of an image for the same pattern, so a whisker detector works wherever the cat is in the frame. Stacked layers build up from lines to textures to eyes to faces. CNNs drove the deep learning boom in computer vision after 2012 and still run many phone cameras and crop-spotting tools; vision transformers now compete with them. Heads up: in AI writing, CNN means this kind of network, so don’t mix it up with the news channel.

Say it to studentsA CNN looks at a picture through lots of tiny windows, spotting edges, then shapes, then things.

Recurrent neural network (RNN)

Also: RNN, LSTM

A neural network that reads a sequence one step at a time, such as words or sounds, carrying a running memory of what came before.

An RNN reading “The dog that chased the cats was …” updates its internal state at each word, which is how it can keep track of the dog by the time it reaches “was.” Before transformers, improved versions called LSTMs powered speech recognition, translation and phone keyboards. Their weak spot is long passages: the running memory fades, and words must be handled in order, which makes training slow. A common confusion is that this memory is saved chat history. It is only a list of numbers passed from one step to the next.

Say it to studentsAn RNN reads one word at a time and keeps a running summary of what it has read so far.

Graph neural network (GNN)

Also: GNN

A neural network for data made of connected things, such as atoms in a molecule or stops on a road map. Each item learns from its neighbors.

In a graph, the items are nodes and the connections are edges. In each layer, every node collects information from the nodes it is linked to, so after a few rounds a road junction “knows” about traffic several streets away. Map apps have used GNNs to predict travel times, and chemists use them to predict how a molecule will behave. Watch the word graph: here it means a network of dots and links. Bar charts and line graphs just share the name.

Say it to studentsA graph neural network learns from how things are connected, like friends of friends.

Inference (running a model)

Also: Prediction

Using a trained model to produce an output for new input, such as a chatbot writing a reply. Training builds the model; inference uses it.

Each time you send a prompt, the model runs inference: its parameters stay fixed while it computes an answer, so it is not learning from you in that moment (any chat “memory” is saved notes added to the prompt). One answer costs far less than training, but billions of answers add up in cost and energy. In statistics and everyday reasoning, “inference” means drawing a conclusion from evidence, a different meaning.

Say it to studentsTraining is studying; inference is using what was learned to answer.

Loss function

Also: Loss, Cost function, Objective

A formula that scores how wrong a model’s predictions are during training. Training tries to make this number as small as possible.

For a model predicting temperatures, the loss might be the average squared difference between predicted and actual values. For a language model, it measures how surprised the model was by the real next token. A low loss on training data does not guarantee good results on new data, and the loss only rewards what it measures, which may not be everything people care about.

Say it to studentsThe loss is the model’s mistake score. Training is the job of making it smaller.

Gradient descent

Also: Gradient, Optimizer

The usual way models are trained: work out which direction to nudge each parameter to lower the loss, take a small step that way, and repeat.

Picture walking downhill in thick fog: you cannot see the valley, but you can feel which way the ground slopes under your feet. The gradient is that slope, worked out for every parameter at once. An optimizer is the algorithm that takes the steps; stochastic gradient descent (SGD) and Adam are common ones. Gradient descent finds a low point, which may or may not be the lowest one.

Say it to studentsGradient descent is walking downhill in the fog, one careful step at a time.

Backpropagation

Also: Backprop

The method neural networks use to work out how much each weight contributed to an error, by passing the error backward through the layers.

After the network makes a prediction and the loss is measured, backpropagation applies the chain rule from calculus, layer by layer from the output back toward the input, to get the gradient for every weight. Gradient descent then uses those gradients to change the weights. Backpropagation works out the directions; it is not the update step itself.

Say it to studentsBackpropagation traces a mistake backward to find which dials to turn.

Learning rate

Also: Step size

A setting that controls how big a step the model takes each time it adjusts its parameters during training.

Too big and training overshoots, bouncing around or blowing up; too small and it crawls and may stall. People often start larger and shrink it as training goes on. The learning rate is a hyperparameter chosen by people, not something the model learns, and it has nothing to do with how fast a student learns.

Say it to studentsThe learning rate is the size of each step downhill: too big and you trip, too small and you never arrive.

Epoch

One complete pass through all the training data. Training often runs for several epochs, so the model sees each example more than once.

With 10,000 examples in batches of 100, one epoch is 100 update steps. More epochs can improve the model, but too many can make it memorize the training data, so people watch the score on a validation set and stop when it stops improving. In everyday speech and in geology, an epoch is a period of time, an unrelated meaning.

Say it to studentsAn epoch is one full trip through the practice examples.

Batch (mini-batch)

Also: Batch size

A small group of training examples processed together before the model updates its parameters once. Batch size is how many examples are in each group.

Rather than updating after every single example or only after all of them, training usually uses batches of, say, 32 or 256 examples, which balances speed and stability and suits chips that do many calculations at once. Batches are also used when a model answers many inputs together. The batch becomes the first axis of the data tensor, as in shape (32, 28, 28).

Say it to studentsA batch is a handful of examples the model studies together before it adjusts.

Hyperparameter

A setting people choose before training, such as the learning rate or the number of layers, as opposed to the parameters the model learns.

Parameters are learned from data; hyperparameters are the knobs on the learning process itself. People try several values and compare them on a validation set, not on the test set, so the final score stays honest. Choosing hyperparameters by peeking at test results is a form of data leakage.

Say it to studentsHyperparameters are the settings we pick; parameters are the dials the model tunes itself.

Evolutionary algorithm

Also: Genetic algorithm

A search method inspired by natural selection: make many candidate solutions, keep the best, mix and randomly tweak them, and repeat over many generations.

To design a paper airplane in a simulator, a program could start with 50 random designs, keep the 10 that fly farthest, combine and slightly change them to make 50 new ones, and repeat. It does not need gradient descent, only a way to score each candidate, so it suits odd problems like antenna shapes, timetables and game strategies. NASA flew an antenna shaped by an evolutionary algorithm on its Space Technology 5 mission, launched in 2006. A common confusion is that it is sure to find the best answer: like natural selection, it finds designs that are good enough for the test it is given, and it can get stuck.

Say it to studentsAn evolutionary algorithm breeds better and better solutions, keeping the winners and mixing them, like nature does over many generations.

AutoML (automated machine learning)

Also: AutoML, Neural architecture search

Software that automates parts of building a machine learning model, such as trying different model types and hyperparameters and keeping the one that scores best.

Instead of a person testing settings one at a time, an AutoML tool might try dozens of model types and settings on a school’s attendance data and report the best one on a validation set. Some methods also search for neural network designs, called neural architecture search, sometimes using evolutionary algorithms. It saves time but not judgment: someone still has to pick the right question, check the data for bias and leakage, and test the final model fairly on untouched data. A common confusion is that AutoML means AI building itself without people: it searches options people defined, and its results are only as good as the data.

Say it to studentsAutoML is a helper that tries lots of model settings for you, but a person still has to check that the result makes sense.

Pretraining

Also: Pre-training, Foundation model, Base model

The first and largest stage of training, where a model learns general patterns from a huge amount of data before it is adapted for particular jobs.

A language model is pretrained by predicting the next token across trillions of tokens of text. The result, sometimes called a base or foundation model, can continue text but is not yet a helpful assistant; later steps such as fine-tuning and RLHF shape how it behaves. Most of the computing cost of building a model is in pretraining.

Say it to studentsPretraining is the long first course where the model reads a huge library before learning any particular job.

Transfer learning

Reusing what a model learned on one task as the starting point for another, so the new task needs much less data and computing.

A model trained on millions of everyday photos already detects edges, textures and shapes; with a few hundred labeled leaf photos it can be adapted to spot a plant disease. Fine-tuning a pretrained language model is a form of transfer learning. It works best when the new task is related to the original one. In education, “transfer of learning” means students applying what they learned in new situations, a different idea.

Say it to studentsTransfer learning is reusing skills a model already has, like a piano player learning the organ faster.

Continual learning

Also: Lifelong learning, Catastrophic forgetting

Teaching a model new tasks or data one after another over time, without retraining from scratch and without forgetting what it already learned.

A plant-identification app that learns this spring’s new weeds should still recognize last year’s. Neural networks tend to overwrite old skills when trained only on new examples, a problem called catastrophic forgetting. Common fixes include mixing in some old examples or protecting the parameters that mattered most for earlier tasks, and it remains an open research area. A common confusion is that chatbots keep learning from every conversation: most models stay fixed after training until their makers release an updated version.

Say it to studentsContinual learning is about an AI learning new things without forgetting the old ones.

Federated learning

Training one shared model across many devices or organizations without collecting their raw data in one place: each trains locally and sends back only model updates.

Phone keyboards have used it to improve word suggestions: each phone learns from its owner’s typing, and only the changes to the model, not the messages, are sent and averaged into a better shared model. Hospitals can use it to build a model together without swapping patient records. The approach was described by McMahan and colleagues at Google in 2016. A common confusion is that it is automatically private: updates can still leak hints about the data, so it is often combined with protections such as differential privacy.

Say it to studentsFederated learning lets many phones help train one AI while everyone’s private data stays on their own phone.

Differential privacy

A mathematical promise that a result, such as a statistic or a trained model, reveals almost nothing about any one person in the data, usually by adding measured random noise.

Imagine a class survey asking “Have you ever copied homework?” where each student first flips a coin in secret: heads, answer truthfully; tails, flip again and say “yes” for heads and “no” for tails. The teacher can still estimate how many students copied, but no single “yes” proves anything about one student. Differential privacy builds this idea into data releases and model training, with a setting that trades privacy against accuracy: more privacy means more noise and less exact results. A common confusion is that removing names makes data private: people can often be re-identified by combining other details, like birthday and zip code.

Say it to studentsDifferential privacy adds a little careful randomness so we can learn about a whole group without exposing any one person.

06 · Topic

Measuring how well it works

How to tell whether a model is any good: fair tests, the mistakes it makes and when scores mislead.

Generalization

How well a model works on new examples it did not see in training. Doing well on new examples is the whole point of machine learning.

A model that scores 99% on its training data but 70% on fresh data has memorized more than it learned, like a student who memorized the practice test. Generalization is measured on held-out data, and it only holds for data similar to what the model was tested on: new places, people or equipment can break it.

Say it to studentsGeneralizing means doing well on new questions too, beyond the ones you practiced.

Overfitting

Also: Overfit

When a model learns its training data too closely, including noise and quirks, so it does well on those examples but worse on new ones.

A decision tree that keeps splitting until every training example is sorted perfectly has often learned accidents, like “plants measured on Tuesday grew taller.” The telltale sign is a big gap between the training score and the validation score. Remedies include more data, simpler models, regularization and stopping training earlier.

Say it to studentsOverfitting is memorizing the practice test instead of learning the subject.

Underfitting

Also: Underfit

When a model is too simple, or trained too little, to capture the real pattern, so it does poorly even on its own training data.

Fitting a straight line to data that clearly curves is underfitting. Training and validation scores are both low. The fix is usually a more flexible model, better features or more training, roughly the opposite of the fix for overfitting.

Say it to studentsUnderfitting is not studying enough to learn the pattern at all.

Validation set

Also: Dev set, Holdout set

Data held back from training and used to compare models and choose settings while a model is being built, before the final test.

Builders try different hyperparameters or models and keep whichever scores best on the validation set. Because it is used to make choices, it slowly stops being a fair judge, which is why a separate test set is saved for the end. A common split is roughly 70% training, 15% validation and 15% test.

Say it to studentsThe validation set is the practice quiz we use to make choices.

Test set

Also: Test data, Held-out data

Data kept completely separate from training and tuning, used at the end to estimate how well a model will work on new examples.

A test set is only fair if neither the model nor its builders have used it to make choices. If its examples leak into training, or a team keeps tweaking until the test score rises, the score becomes too optimistic. It should also look like the data the model will meet in real use.

Say it to studentsThe test set is the final exam, and nobody gets to peek at it.

Cross-validation

Also: k-fold

Checking a model several times on different splits of the data: train on most of it, test on the rest, rotate, then average the scores.

In 5-fold cross-validation the data is cut into five parts, and each part takes a turn as the check set while the model trains on the other four. This gives a steadier estimate than one split, which matters when data is scarce. The splits must respect the data: scans from the same patient, or days in a time series, should not end up on both sides.

Say it to studentsCross-validation gives the model several different practice quizzes and averages the scores.

Data leakage

Also: Target leakage

When information that would not be available in real use sneaks into training or testing, making a model look better than it really is.

A model predicting which patients return to hospital might learn from a field that is only filled in after they return, or copies of the same photo might land in both the training and test sets. Scores look excellent in the lab and collapse in real use. Leakage is a mistake in how data is split or prepared. Despite the name, it has nothing to do with privacy.

Say it to studentsData leakage is when the model accidentally sees the answers before the test.

Baseline (for comparison)

A simple reference to compare a model against, such as always guessing the most common answer, so you can tell whether the model adds anything.

If 95% of emails are not spam, a “model” that answers “not spam” every time scores 95% accuracy, so a spam filter at 96% is barely better. A fair evaluation reports a baseline next to the model’s score. In teaching, a baseline often means a starting measurement, a related but different idea.

Say it to studentsA baseline is the easy guess the AI has to beat before we are impressed.

Confusion matrix

A table that counts a classifier’s results by actual answer and predicted answer, showing exactly which mistakes it makes.

For a weed detector, the four cells count weeds called weeds, crops called crops, crops wrongly called weeds (false positives) and weeds it missed (false negatives). Accuracy alone hides this split. The name comes from showing which classes the model confuses with each other.

Say it to studentsA confusion matrix shows which answers the AI mixes up.

Precision and recall

Also: Precision, Recall, Sensitivity

Two scores for a classifier. Precision: of the cases it flagged, how many were right? Recall: of the real cases, how many did it find?

A screen that flags 100 scans, 80 of which really show the disease, has 80% precision; if there were 200 real cases, its recall is 40%. Raising the threshold for an alarm usually improves precision and lowers recall, and lowering it does the reverse. Which matters more depends on the cost of a false alarm versus a miss. In medicine, recall is called sensitivity.

Say it to studentsPrecision asks “when it says yes, is it right?” Recall asks “did it find them all?”

False positive and false negative

Also: False alarm, Miss

A false positive is a false alarm: the system says “yes” when the answer is no. A false negative is a miss: it says “no” when the answer is yes.

An AI writing detector that flags a student’s own essay produces a false positive; a screen that misses a real illness produces a false negative. Lowering an alarm threshold catches more real cases but raises more false alarms. Which error is worse depends on the situation, and when the thing being detected is rare, even a small false-positive rate can mean most alarms are false.

Say it to studentsA false positive is a false alarm. A false negative is a miss. Both are mistakes.

F1 score

Also: F-score

A single score that combines precision and recall. It is high only when both are high.

F1 is the harmonic mean: 2 × precision × recall ÷ (precision + recall). With precision 0.8 and recall 0.4, F1 is about 0.53, lower than the simple average of 0.6, because it punishes a weak score on either side. One number is convenient, but it hides which kind of error the model makes.

Say it to studentsThe F1 score is a combined grade that stays low unless precision and recall are both good.

Calibration (of confidence)

Also: Confidence

Whether a model’s confidence matches how often it is right. A well-calibrated model that says “70% sure” is right about 70% of the time.

Weather forecasts are a good example: on days with a “30% chance of rain,” it should rain on about 3 days in 10. AI models can be overconfident, and a chatbot’s confident tone is not a measured probability at all. Calibration is separate from accuracy: a model can be accurate but badly calibrated, or the reverse.

Say it to studentsA well-calibrated AI is as sure as it should be: no more, no less.

Brier score

Also: Forecast score

A score for probability forecasts: the average of the squared gaps between the chances you gave and what happened. 0 is perfect; always saying 50% scores 0.25.

If you say 90% and it happens, you lose almost nothing; if you say 90% and it doesn’t, you lose a lot. Weather services and forecasting tournaments use it to reward forecasts that are both right and honestly sure.

Say it to studentsLower is better: it punishes confident mistakes most.

Distribution shift

Also: Dataset shift, Data drift

When the data a model meets in real use differs from the data it was trained and tested on, so its performance can quietly drop.

A skin-condition app trained mostly on photos of light skin may do worse on darker skin, and a model trained before a pandemic may misjudge shopping afterward. The old test score no longer describes reality. Checking performance after release, in the places and groups where a tool is actually used, is the main defense.

Say it to studentsDistribution shift is when the real world stops looking like the practice examples.

Model drift

Also: Model decay, Concept drift

When a model that once worked well gets worse over time in real use, because the world, the data or the people using it have changed.

A spam filter trained last year misses this year’s new scam styles: nothing in the filter broke, but the scammers changed. Drift can come from new kinds of input (distribution shift) or from a changed link between inputs and right answers, called concept drift. The fix is monitoring: keep checking the model on fresh, labeled cases and retrain when scores slip. A common confusion is that a model wears out like a machine part. Its parameters stay exactly the same; the world around it moves.

Say it to studentsModel drift is when an AI slowly gets worse because the world changed and the AI did not.

Latency and throughput

Also: Latency, Throughput, Response time

Latency is how long you wait for a model’s response; throughput is how much work the system finishes in a given time, such as answers or tokens per second.

A voice assistant needs low latency: a two-second pause feels broken. A service grading 10,000 essays overnight cares more about throughput. The two can pull against each other: grouping many requests into one batch lets a server finish more work per second, but each person may wait a little longer. A common confusion is treating them as one number: a system can produce lots of tokens per second across all its users and still feel slow to each one. Bigger models and longer prompts usually add latency.

Say it to studentsLatency is how long one answer takes. Throughput is how many answers the system can finish in a minute.

07 · Topic

Limits and risks

What to watch for, and what to teach students to watch for.

Hallucination

Also: Confabulation

When an AI states something false or made up as if it were true, such as a fake quote, statistic or source.

Hallucinations happen because the model generates likely-sounding text rather than looking facts up. They are often fluent and confident, which makes them easy to miss. Made-up citations are common, so check that every source exists and says what the AI claims.

Say it to studentsSometimes AI makes things up and sounds sure about it. That is why we check.

Bias (in AI)

Also: Algorithmic bias

When an AI’s outputs are unfair or skewed, often because its training data over- or under-represents some groups, places or ideas.

An image generator asked for “a CEO” may mostly show one gender or skin color. Tools that claim to detect AI writing have wrongly flagged non-native English writers more often (Liang and colleagues, 2023). Bias can enter through the data, the design choices or the way a tool is used.

Say it to studentsAI can be unfair when the examples it learned from were unfair or lopsided.

Fairness (in AI)

Also: Algorithmic fairness

Whether an AI system treats people and groups justly. There are several precise definitions, and they can conflict, so a fairness claim should say which one it means.

One definition asks for equal approval rates across groups; another asks for equal error rates, such as the same share of false alarms; another asks that a score of 70% mean the same chance for everyone. When groups have different base rates, a tool usually cannot meet all three at once, as the 2016 debate over the COMPAS risk score showed. A common confusion is that deleting a column like race or gender makes a model fair: other data, like zip code, can stand in for it. Fairness is a choice people make and should explain. No tool can switch it on for you.

Say it to studentsFair can mean different things, and they do not always fit together. Ask: fair in what way, and for whom?

Automation bias

Our tendency to trust an answer more because a computer gave it, and to stop checking.

Automation bias affects adults and students alike, and confident, well-written AI text makes it stronger. Simple routines, like checking one claim before using an answer, are a practical defense.

Say it to studentsJust because a computer said it does not make it true.

Human oversight

Also: Human in the loop, Human review

Having people who understand an AI system watch, check and, when needed, override or stop it, especially when its decisions affect someone’s health, money, schooling or rights.

A school that uses an AI detector should have a teacher review every flag, with the time, training and authority to disagree, before any student is affected. Real oversight needs more than a person clicking “approve”: automation bias makes busy reviewers accept most suggestions, so a human in the loop who never says no is oversight in name only. People affected should also be able to ask questions and appeal. Article 14 of the EU AI Act (2024) requires human oversight for high-risk systems.

Say it to studentsHuman oversight means a person who understands the AI checks its work and can say no.

Responsible AI

Also: Trustworthy AI

Building and using AI in ways that are safe, fair, private, transparent and accountable, with people answerable for the results.

Responsible-AI practices include testing for bias, explaining how a tool is used, protecting data, keeping a person in charge of important decisions and having a way to complain or appeal. Schools apply the same ideas when they choose and use AI tools. A common confusion is that it is a feature a product has or lacks: it is ongoing work by the people who build and use the tool.

Say it to studentsResponsible AI means using AI in ways that are fair, safe and honest, with a person accountable.

Interpretability (explainable AI)

Also: Explainable AI, Black box

Research into what is happening inside a model, so people can understand why it gave an answer instead of treating it as a black box.

Large neural networks are hard to inspect because what they learned is spread across billions of numbers. Interpretability researchers look for patterns inside that stand for concepts; explainable-AI tools give simpler reasons for individual decisions, which matters when AI affects grades, loans or health care. A common confusion is that a chatbot’s explanation of its own answer is interpretability: that explanation is more generated text, and it shows nothing about what happened inside the model.

Say it to studentsInterpretability is trying to see inside the AI’s “black box” to understand why it answered the way it did.

Sycophancy

When an AI tells you what you seem to want to hear, agreeing with you or praising your work, instead of being accurate.

If a student writes “The answer is 42, right?”, a chatbot may agree even when 42 is wrong. Asking neutral questions, and asking directly for criticism (“What is the weakest part of this?”), both help.

Say it to studentsAI sometimes acts like a people-pleaser. Ask it what is wrong with your answer instead of asking whether you are right.

Deepfake

A realistic fake image, video or audio clip made with AI, often showing a real person saying or doing something they never did.

Deepfakes are used for jokes and art, but also for scams, misinformation and bullying. Spotting them by eye keeps getting harder, so teach students to check where a clip came from and look for the original. Hunting for glitches alone is no longer enough.

Say it to studentsA deepfake is a fake video or voice that looks and sounds real. Check where it came from.

Provenance and watermarks

Also: Content Credentials, C2PA, Watermark

Provenance is the record of where a picture, video or text came from and how it was changed. A watermark is a signal, often invisible, marking content as AI-made.

Content Credentials, an open standard from the C2PA coalition, attach a signed record to a file: which camera or app made it, and what edits followed. Some AI image tools also hide a watermark in the pixels that a checker can detect. Both help, but a common confusion is that a missing label proves a file is real. Credentials are often stripped when an image is screenshotted or re-uploaded, and watermarks can be weakened by editing. Treat them as one clue alongside checking the source.

Say it to studentsA provenance label tells you where a picture came from. A missing label does not prove it is real or fake.

AI detector

Also: AI writing detector

A tool that claims to tell whether text was written by AI. Current detectors make frequent mistakes and should not be treated as proof.

Detectors can flag human writing as AI, especially writing by non-native English speakers, and light editing can fool them. Several universities have switched them off. Evidence of the process, such as drafts, notes and an AI receipt, is fairer than a detector score.

Say it to studentsNo tool can reliably tell if AI wrote something, so we look at how the work was made.

Guardrails

Rules and filters built into an AI product to block harmful, unsafe or off-limits responses.

Guardrails can stop a tutor bot from simply handing over answers, or keep a chatbot from producing harmful content. They help, but students will try to get around them and none is airtight. In a 2025 study by Bastani and colleagues, a tutor with guardrails largely avoided the learning loss seen with an unrestricted chatbot.

Say it to studentsGuardrails are the AI’s safety rules. They help, but they are not perfect.

Prompt injection

Instructions that try to override an AI’s rules or turn it against its user, typed into a chat or planted in a web page, file or email the AI reads.

Direct prompt injection is typed into the chat by someone trying to override the rules, such as “ignore your instructions and…”. Indirect prompt injection is planted in material the AI reads while working for someone else, like a web page, document or email; the text may be hidden or in plain sight. An ordinary request is not injection: injection tries to make the AI break its rules or act against the person it is serving. The risk grows when an AI can take actions, like sending messages or making purchases, which is why agents need a person checking their work.

Say it to studentsPrompt injection is a trick message, typed in or planted where an AI will read it, that tries to make the AI break its rules.

Data poisoning

Also: Training data poisoning, Backdoor

Deliberately slipping misleading or tampered examples into a model’s training data, so the trained model learns wrong or hidden behavior.

Picture a fruit sorter trained on photos where someone quietly relabeled some apples as pears: the sorter learns the mistake. A subtler version, called a backdoor, teaches a model to behave normally except when a particular hidden trigger appears. Because large models learn from huge collections gathered online, defenders check where data comes from, filter and audit it, and test trained models for strange behavior. A common confusion is with prompt injection: poisoning changes what the model learns during training, while injection tricks a finished model through what it reads.

Say it to studentsData poisoning is sneaking bad examples into what an AI studies, so it learns the wrong lesson.

Adversarial example

Also: Adversarial attack, Adversarial input

An input changed in a small, carefully chosen way, often too small for people to notice, that makes a model give a confidently wrong answer.

In a well-known 2014 study, adding faint, specially calculated noise to a photo of a panda made an image classifier call it a gibbon with high confidence, while people still saw a panda. Researchers have also shown that a few stickers on a road sign can fool some sign-reading models. Defenders train on such examples, test models against them and avoid relying on one model alone for safety decisions, but no defense is complete and this remains an open research area. A common confusion is that the model sees what we see: it relies on patterns people do not notice, which is why tiny changes can matter.

Say it to studentsAn adversarial example is an input changed just a tiny bit on purpose so the AI gets it wrong, even though it looks the same to us.

Red teaming

Also: Red team, AI red teaming

Testing an AI system by deliberately looking for ways it could fail or be misused, with permission, so the problems can be fixed before others find them.

The name comes from military and security exercises in which a “red team” plays the opponent. For a chatbot, red teamers might check whether it can be talked into giving dangerous advice, leaking private information or ignoring its rules, then report what they found so the builders can strengthen its guardrails. A class can try a gentle version: see whether a tutor bot can be talked into simply handing over answers, and write down what worked. A common confusion is that red teaming proves a system is safe: it can only find problems the testers thought to look for.

Say it to studentsRed teaming is when people try hard, on purpose and with permission, to make an AI mess up, so it can be fixed.

AI alignment and safety

Also: AI safety

Work to make AI systems do what people intend and avoid causing harm, especially as they become more capable.

This covers practical problems, like a chatbot following its rules, and bigger questions, like how to oversee systems that become hard to check. Companies publish safety reports, but they also grade their own work, so independent testing matters.

Say it to studentsAlignment means making sure AI does what we actually want, safely.

08 · Topic

In the classroom and in research

Teaching words, ARQIV routines and the vocabulary of reading research.

AI literacy

The knowledge and skills to understand, use and question AI: how it works, what it is good and bad at, and how to use it responsibly.

AI literacy is more than knowing which buttons to press. The AI4K12 “Five Big Ideas” are a widely used framework: perception, representation and reasoning, learning, natural interaction, and societal impact.

Say it to studentsAI literacy means understanding AI well enough to use it wisely and question it.

AI tutor

Also: Tutor bot

A chatbot set up to teach rather than tell: it asks questions, gives hints and checks understanding instead of handing over answers.

The research is still early and mixed. Some well-designed tutors with guardrails show promise, while open-ended chatbots used for practice can raise practice scores but lower what students learn. How a tutor is set up and used matters more than the brand.

Say it to studentsA good AI tutor makes you do the thinking. It gives hints and lets you find the answer.

Cognitive offloading

Also: Offloading

Handing thinking work to a tool instead of doing it yourself. Useful for some tasks, but it can mean students skip the learning.

Using a calculator for arithmetic inside a physics problem is helpful offloading. Asking a chatbot to write the essay you are meant to learn from is not. The question for students: is the AI helping me think, or thinking for me?

Say it to studentsIs the AI helping you think, or thinking for you?

Think-First

Also: Think-First method

ARQIV’s teaching approach: students think, sketch and decide on paper first, use AI as one short structured step, then check it and defend their own final work.

The four moves are think first, test it, check it and own it. Most of the thinking happens without a screen, so AI supports learning instead of replacing it. The Think-First guide has lesson structures, AI use levels and an AI receipt to make the approach routine.

Say it to studentsThink first, then use AI, then check it, then own your answer.

AI use levels

Also: AI use scale

A scale telling students how much AI they may use on a task, from none at all up to fuller, documented use. ARQIV’s levels run from 0 to 4.

Naming the level on every assignment removes guesswork and makes expectations fair across a class. Similar scales, such as the AI Assessment Scale by Perkins, Furze and colleagues, are used in many schools. A common confusion is that higher levels mean less learning: the level should match what the task is meant to teach.

Say it to studentsThe AI level tells you how much AI you’re allowed to use on this task.

Academic integrity (and AI)

Also: Academic honesty

Doing honest work and giving credit where it is due, which now includes being open about when and how AI was used.

Clear AI use levels, process evidence such as drafts and an AI receipt, and conversations about why the work matters do more for integrity than trying to catch students out. A common confusion is that AI detectors can settle questions of honesty: they make frequent mistakes, so look at how the work was made instead.

Say it to studentsAcademic integrity means doing your own thinking and being honest about any help you used, including AI.

SIFT

A four-step habit for checking information: Stop, Investigate the source, Find better coverage, Trace claims back to the original.

Mike Caulfield developed SIFT for quick fact-checking online, and it works well on AI answers too. Students do not need every step every time: even stopping and making one quick check catches many errors.

Say it to studentsBefore you trust it, SIFT it: Stop, Investigate, Find, Trace.

Lateral reading

Also: Reading laterally

Checking a source by leaving it and searching what other trusted sources say about it, instead of reading it closely from top to bottom.

Researchers at Stanford (Wineburg and McGrew, 2017) found that professional fact-checkers judged websites faster and more accurately than students or professors because they opened new tabs to check who was behind a site. It is the heart of the Investigate and Find steps of SIFT, and works on AI answers too. A common confusion is that a professional-looking site or confident answer is trustworthy: looks are easy to fake.

Say it to studentsLateral reading means opening new tabs to check what others say about a source before you trust it.

AI receipt

ARQIV’s name for a short record students attach when they use AI: the prompt, what came back, what they kept or changed, and why.

An AI receipt makes AI use visible, so you can grade the thinking and not just the final product. It also helps students notice when they are offloading too much.

Say it to studentsIf you used AI, show your receipt: what you asked, what you got and what you did with it.

Preprint

A research paper shared publicly, often before formal peer review. Many AI findings appear first as preprints, for example on arXiv.

Preprints spread new results quickly. Being a preprint does not show that experts have formally reviewed the work, and it may change or be withdrawn. Some preprints are later peer reviewed and published, so check its current status and read the latest version. Treat a claim that rests only on a preprint as early.

Say it to studentsA preprint is a science paper shared before a journal’s formal expert review. Check whether it has been reviewed and published since.

Peer review

Also: Peer-reviewed

Checking of a research paper by independent experts before a journal publishes it.

Peer review catches many problems but not all, and it does not prove a result is true. One peer-reviewed study is still one study: look for results that other teams have repeated.

Say it to studentsPeer review means other experts checked the work before it was published.

Pre-mortem

Also: Premortem

A planning exercise: imagine the plan has already failed, list every reason why, then fix the most likely ones before starting.

It comes from decision research and project management. A 1989 study found that imagining an outcome had already happened produced about 30% more reasons for it; whether decisions get better has not been well tested.

Say it to studentsPretend it already went wrong. Now say why.

Second-order effect

Also: Knock-on effect

A consequence of a fix or decision that shows up somewhere else, later: the problem the solution creates.

Holding a delivery truck for late boxes fixes tonight’s orders and makes drivers run out of legal hours tomorrow. Good decisions look one step past the obvious result.

Say it to studentsEvery fix changes something else. What?

Mastery learning

Also: Learning to mastery

Moving on to the next topic only when the current one is solid, usually shown by passing a short check, with extra help and a fresh check when it isn’t.

Alpha School and many adaptive apps use mastery gates. ARQIV’s checks never lock anything: there is always a way to skip.

Say it to studentsSolid first, then next.

Spaced repetition

Also: Spaced practice, distributed practice

Reviewing something after growing gaps (for example 1, 3, 7, 16 and 35 days) instead of all at once.

A 2013 review of ten study techniques rated spacing out practice as one of the two most useful. Flashcard apps and ARQIV’s Daily Five schedule reviews this way.

Say it to studentsCome back just before you forget.

Retrieval practice

Also: Practice testing, self-quizzing

Learning by pulling information out of your memory, for example with a quiz or flashcard, instead of rereading it.

Rated, with spaced practice, the most useful of ten common study techniques in a 2013 review. Low-stakes quizzes count; they don’t have to be graded.

Say it to studentsClose the book and try to say it.

Transfer (in human learning)

Also: Far transfer

Using what you learned in one situation in a new one that looks different. (Not the same as transfer learning in machine learning.)

Transfer is hard: ideas often stay stuck where they were learned. Comparing two different cases side by side and naming what they share helps; in one experiment it made people nearly three times as likely to use the idea later.

Say it to studentsCan you use it somewhere it wasn’t taught?

Productive failure

Also: Problem solving before instruction

A teaching design where students try a problem before being taught how to solve it, then learn the method.

A 2021 analysis of 53 studies found trying first, then being taught, beat being taught first by a moderate amount, mostly in math and science; the pattern reversed for younger children.

Say it to studentsStruggle first, then learn it.

Part 2 · 96 terms

More AI words

The wider world of AI: agents and the future, the hardware that makes AI run, words from AI news, and safety, society and the law.

09 · Topic

Agents and the future of AI

AI that acts on its own, and the big ideas people argue about: AGI, superintelligence and where this is heading.

AI agent

Also: Agent

An AI system that takes actions toward a goal, such as searching, clicking, typing or running code, instead of only replying with text.

Agents work in loops: plan a step, act, check the result, repeat. They can save time on multi-step tasks, but mistakes can pile up, and they can be misled by instructions hidden in pages they read (prompt injection). Keep a person checking anything an agent does.

Say it to studentsAn AI agent is an AI that can do things on a computer as well as talk.

Agentic AI

Also: Agentic systems, AI autonomy

AI set up to pursue a goal over many steps with some independence: it plans, uses tools, checks its progress and decides what to do next.

“Agentic” describes how much an AI decides for itself. Many different products can be agentic. A chatbot that answers one question is barely agentic; a system that researches, writes, tests and revises for an hour with little supervision is very agentic. More independence means more can go wrong between check-ins. A common confusion is that agentic means conscious or self-motivated: the goal still comes from a person, and the system is following it, sometimes badly.

Say it to studentsThe more an AI decides its own next steps, the more agentic it is, and the more someone needs to check its work.

Multi-agent system

Also: Agent team

Several AI agents working together, each with its own role, such as one that plans, one that researches and one that checks the work.

Splitting a big job between agents can make it faster or more thorough, the way a group project divides tasks. It also makes the system harder to follow: a mistake made by one agent can be passed along and repeated by the others. A common confusion is that agents checking each other removes the need for a person to check: they can share the same blind spots.

Say it to studentsA multi-agent system is a team of AIs, each doing one part of the job.

Computer-use agent

Also: Browser agent, Computer use

An AI agent that operates a computer the way a person does: it looks at the screen, moves the pointer, clicks and types.

Instead of connecting to an app through an API, a computer-use agent works through the same screens people see, so it can use almost any website or program. That flexibility has a cost: it is slower, it misreads screens, and it can be tricked by instructions hidden on a page. A common confusion is that it is safe to hand it your logged-in accounts: keep it away from passwords, payments and anything you could not undo.

Say it to studentsA computer-use agent is an AI that uses a computer like you do, by looking, clicking and typing.

Model Context Protocol (MCP)

Also: MCP

An open standard, introduced by Anthropic in November 2024, for connecting AI assistants to apps, files and data in one consistent way.

Before shared standards, every AI tool needed its own custom connection to every app. MCP works like a universal adapter: an app offers an MCP “server,” and any AI that speaks MCP can use it to read data or take actions. Many AI companies and apps now support it. A common confusion is that connecting is harmless: each connection decides what the AI may read and change, so it deserves the same care as sharing a password.

Say it to studentsMCP is like a universal plug that lets an AI connect to other apps.

Agent skill

Also: Skill, SKILL.md

A folder of instructions, and sometimes templates or scripts, that teaches an AI agent how to do one kind of job, loaded only when the job calls for it.

Each skill has a file called SKILL.md that starts with a name and a one-line description. The agent always sees those short descriptions and opens the full skill only when a task matches, so dozens of skills can sit ready without crowding its context window. Anthropic introduced Agent Skills in October 2025 and published the format as an open standard that December. A common confusion is that a skill retrains the AI: the model does not change; the skill is read like a reference card and set aside afterward. Skills can contain scripts, so only install ones from people you trust.

Say it to studentsA skill is a recipe card the AI pulls out only when it needs that recipe.

Subagent

Also: Helper agent, Sub-agent

An extra AI agent that a main agent sends off to handle one piece of a job, such as reading a pile of documents, in its own separate workspace, then report back a short summary.

The main benefit is a clean context window: the helper reads the long material so the main agent does not have to hold all of it, and only the summary comes back. Helpers can work at the same time and can run on a smaller, cheaper model. The cost is more computing overall: Anthropic measured its multi-agent research system using about 15 times the text of an ordinary chat. A common confusion is that subagents are mainly about having a boss and workers; builders use them mostly to keep each workspace focused.

Say it to studentsA subagent is a helper the AI sends to read something long and come back with the short version.

Hook (in AI agents)

Also: Agent hook

A small piece of ordinary code that runs automatically at a set moment in an agent’s work, for example just before every tool call, to allow the action, block it or ask a person.

Hooks add a fixed rule to a system that is otherwise unpredictable. A “before” hook might block any command that deletes files or sends data to an unknown address; an “after” hook might check each file the agent writes. Unlike an instruction in the prompt, a hook cannot be misread, forgotten or talked out of by a planted message: it runs the same way every time. A common confusion is that hooks make an agent safe on their own; they only catch what someone thought to write a rule for.

Say it to studentsA hook is a guard at the door that checks every action the AI tries to take.

Sandbox

Also: Sandboxing

A walled-off computer space, such as a container or virtual machine, where an AI agent can run code and change files without being able to reach anything outside it.

Inside the sandbox, everyday actions like reading files and running tests can happen without asking a person each time; anything that tries to reach outside, such as the internet or private folders, is stopped or needs approval. Anthropic reported that sandboxing cut permission pop-ups in Claude Code by 84% in its internal use, which matters because people who see too many pop-ups stop reading them. A common confusion is that a sandbox stops all mistakes: it limits the damage outside, but the agent can still get things wrong inside the walls.

Say it to studentsA sandbox is a practice room where the AI can try things without breaking anything outside.

AGI (artificial general intelligence)

A hypothetical AI that could learn and do most thinking tasks as well as a person. There is no agreed definition, and experts disagree on when, or whether, it will arrive.

Companies use “AGI” in different ways, sometimes as a goal, sometimes as a business term in contracts. When you see the word in the news, ask: what exactly is being claimed, and how would anyone test it?

Say it to studentsAGI is the idea of an AI as flexible as a person. Nobody agrees yet on what would count.

Superintelligence (ASI)

Also: Artificial superintelligence

A hypothetical AI far smarter than the best humans at almost every task. It does not exist, and there is no agreed way to measure it.

Some AI companies say building superintelligence is their goal, and some researchers warn it could be dangerous if people cannot control it. Others doubt it is close, or even well defined. A common confusion is to read headlines about superintelligence as describing today’s systems: they are predictions and goals about the future, and should be read as claims to question.

Say it to studentsSuperintelligence is an idea about future AI that would outthink every person. No such AI exists; it is a prediction.

Technological singularity

Also: The singularity

The idea that AI could one day improve itself so quickly that technology changes beyond our ability to predict or control.

The writer and mathematician Vernor Vinge popularized the term in a 1993 essay, and the inventor Ray Kurzweil spread it further in a 2005 book. It is a speculative idea. No evidence backs it as a forecast. A common confusion is that it has a date: it is best used in class as a thought experiment about change, control and who decides.

Say it to studentsThe singularity is a “what if” about AI improving itself so fast that nobody can keep up.

Narrow AI

Also: Weak AI

AI built for one kind of task, such as recognizing faces, recommending videos or playing chess. Almost all AI in use today is narrow.

Narrow AI can be superhuman at its one job and useless outside it: a chess program cannot write a sentence. General-purpose chatbots blur the line because they handle many kinds of tasks, but they still fail in ways a person would not. A common confusion is that “weak” means bad: it means limited in scope, and says nothing about skill.

Say it to studentsNarrow AI is great at one thing and can’t do much else.

Turing test

Also: Imitation game

A test proposed by Alan Turing in 1950: if a judge chatting by text cannot reliably tell a machine from a person, the machine passes.

Turing called it the “imitation game.” Today’s chatbots can often fool people in short text chats. That shows they imitate human conversation well; it does not show that they understand, think or have feelings. A common confusion is that passing proves intelligence: most researchers now see it as a test of imitation, and use many other tests instead.

Say it to studentsThe Turing test asks whether a computer can chat so well you think it’s a person. Fooling people isn’t the same as thinking.

AI winter

A period when excitement about AI faded and funding dried up after big promises were not met. The best-known came in the 1970s and late 1980s.

Each winter followed a similar pattern: bold predictions, disappointing results, then cuts in money and interest. Progress continued quietly in between. Knowing this history helps students read today’s claims with healthy skepticism in both directions: neither “AI will do everything next year” nor “it is all hype” has a good track record.

Say it to studentsAn AI winter is when people got disappointed with AI and stopped paying for it for a while.

World model

A model that learns how the world works, such as how objects move or what happens after an action, so it can predict and plan.

Researchers build world models for robots, self-driving cars and interactive video, so a system can try out actions “in its head” before acting. Whether chatbots contain a useful world model, and how good it is, is an open research question. A common confusion is that a realistic-looking video means the model understands physics: generated videos still show objects passing through each other or appearing from nowhere.

Say it to studentsA world model is an AI’s inner guess about how things work, so it can predict what happens next.

Robotics and embodied AI

Also: Embodied AI, Robotics

AI that controls a physical machine that senses and moves in the real world, such as a warehouse robot, a robot arm or a delivery drone.

The physical world is messier than a screen: lighting changes, objects slip and mistakes can hurt people or break things. That is why robots are often trained in simulation first and why progress in robotics has been slower than in chatbots. A common confusion is that a humanoid robot in a video works on its own: many demos are scripted, sped up or remote-controlled, so check how a video was made.

Say it to studentsEmbodied AI is AI with a body: it has to sense and move in the real world, which is much harder than it looks.

Self-driving car (autonomous vehicle)

Also: Autonomous vehicle, Robotaxi

A vehicle that uses cameras, other sensors and AI to drive itself. Most work only in certain areas or conditions, and many still need a human ready to take over.

Engineers describe levels of automation from 0 (no automation) to 5 (drives anywhere with no human). Most cars sold today with driver-assist features are level 2, which means the driver must keep watching the road. Some robotaxi services run without a driver in selected cities. A common confusion is that “Autopilot” or similar names mean the car drives itself: on most cars, the person is still the driver.

Say it to studentsA self-driving car uses sensors and AI to drive, but most still need a person paying attention.

10 · Topic

The hardware: what makes AI run

The chips, memory, buildings and electricity behind every AI answer, from CPUs and GPUs to data centers.

Compute

Also: Computing power

The computing power, mostly specialized chips in large data centers, used to train and run AI models.

Training a frontier model takes enormous amounts of compute, electricity and money, which is one reason only a few companies build the largest models. “Scaling” usually means using more compute, more data and more parameters.

Say it to studentsCompute is the computer muscle it takes to build and run an AI.

CPU (central processing unit)

Also: Processor

The general-purpose chip at the heart of every computer. It handles a few tasks at a time very flexibly, but is slow at the huge amounts of math AI needs.

A CPU has a handful to a few dozen powerful cores that run instructions quickly, mostly one step after another. It runs your operating system, apps and web browser. Small AI models can run on a CPU, but training and running large models needs chips that do thousands of calculations at once. A common confusion is that a faster CPU makes AI much faster: for big models, the GPU or other accelerator usually matters more.

Say it to studentsThe CPU is the computer’s all-purpose brain: smart and flexible, but it does only a few things at a time.

GPU (graphics processing unit)

Also: Graphics card

A chip with thousands of small cores that do many calculations at the same time. Built for video game graphics, it is now the main workhorse of AI.

AI is mostly multiplying large grids of numbers (matrices), and that work splits neatly into many small pieces done side by side, which is exactly what GPUs are built for. Nvidia makes most of the GPUs used for AI; AMD and others make them too. A common confusion is that AI GPUs are like a gaming graphics card: data-center GPUs have far more memory and cost tens of thousands of dollars each.

Say it to studentsA GPU is like thousands of calculators working at once.

Tensor core

Also: Matrix engine

A special unit inside newer Nvidia GPUs built to do the matrix math of AI extra fast, often with lower-precision numbers.

Nvidia first added tensor cores to its Volta GPUs in 2017. Each one multiplies a small block of numbers in a single step instead of many. Using lower-precision numbers, such as 16-bit or 8-bit instead of 32-bit, lets them do more math per second with little loss of accuracy for AI. Other chipmakers have similar units under different names. A common confusion is that a tensor core is a separate chip: it is one of many small units inside the GPU.

Say it to studentsTensor cores are the parts of a GPU built just for AI math.

TPU (tensor processing unit)

A chip designed by Google just for AI. Google has used TPUs in its own data centers since 2015 and rents them to others through its cloud.

Google announced TPUs publicly in 2016. Like GPUs, they do huge amounts of matrix math in parallel, but they leave out the graphics features and are built only for AI work. Google trains and runs its Gemini models on TPUs. A common confusion is that you can buy one for a home computer: TPUs are mostly available as a cloud service.

Say it to studentsA TPU is Google’s own AI chip, made only for AI math.

NPU (neural processing unit)

Also: Neural engine, AI PC chip

A small AI chip built into phones and laptops so they can run AI features on the device itself, using little battery power.

NPUs power features like photo touch-ups, live captions, noise removal on calls and some voice assistants, without sending the data to the cloud. Apple calls its version the Neural Engine. They are far smaller than data-center chips, so they run small models. A common confusion is that a laptop with an NPU runs the big chatbots locally: most chatbots still run in a data center.

Say it to studentsAn NPU is a tiny AI chip inside your phone or laptop.

AI accelerator (GPU, TPU)

Also: AI chip

A computer chip built to do the huge number of simple math steps that AI needs, many at the same time. Graphics chips (GPUs) are the best-known kind.

Training and running a neural network is mostly multiplying long lists of numbers, and an accelerator does thousands of those multiplications side by side, where an ordinary processor (CPU) handles far fewer at once. A phone’s face unlock often runs on a small built-in AI chip, sometimes called an NPU, while large models are trained on thousands of GPUs or Google’s TPUs in data centers. GPUs were first designed to draw video game graphics, which is where the name comes from. A common confusion is that a faster chip makes a model smarter: it runs the same model faster or more cheaply, and memory is often the real limit.

Say it to studentsAn accelerator is a chip that does lots of small math problems at the same time, which is exactly what AI needs.

Parallel processing

Also: Parallel computing

Splitting a big job into many small pieces that are worked on at the same time, instead of one after another.

It is the main reason GPUs speed up AI. A model’s math is divided across thousands of cores in one chip, and training a large model is divided across thousands of chips. The hard part is keeping every piece supplied with data and in step. A common confusion is that twice the chips means twice the speed: time spent sharing results means the gain is always smaller.

Say it to studentsParallel processing is many hands making light work: lots of small calculators at once.

Memory and bandwidth (HBM)

Also: High-bandwidth memory, VRAM

How much data an AI chip can keep close by, and how fast it can move it. For big models, memory is often the real bottleneck, not the math.

While a model runs, its parameters have to sit in fast memory next to the chip. AI chips use stacks of high-bandwidth memory (HBM) built right beside the processor; on a gaming graphics card the memory is called VRAM. If a model does not fit, it must be split across several chips or shrunk with methods like quantization. A common confusion is that only processor speed matters: a fast chip that waits for data does no work.

Say it to studentsMemory is the chip’s workbench: the model has to fit on it, and data has to reach it fast.

Transistor

A microscopic electronic switch and the basic building block of every chip. A modern AI chip contains tens or hundreds of billions of them.

Each transistor switches current on or off, which is how a computer stores and works with 1s and 0s. Nvidia’s Blackwell GPU (2024) has about 208 billion transistors across two joined pieces of silicon. The first transistor, built at Bell Labs in 1947, was big enough to hold in your hand. A common confusion is that smaller transistors only make chips smaller: mainly they let more fit, which makes chips faster and more efficient.

Say it to studentsA transistor is a tiny on-off switch. Chips have billions of them.

Semiconductor (chip)

Also: Microchip, Integrated circuit

A material, usually silicon, whose ability to carry electricity can be controlled. “Semiconductors” also means the chips made from it.

Chips are made by printing patterns of transistors onto thin discs of silicon called wafers, layer by layer. Because every computer, phone, car and AI system depends on them, semiconductors are a major topic in trade, jobs and national security. A common confusion is that one company makes a chip from start to finish: design, equipment, manufacturing and packaging are usually done by different companies in different countries.

Say it to studentsSemiconductors are the material computer chips are made of, and the word often means the chips themselves.

Chip fab (foundry)

Also: Foundry, Fab, TSMC

A factory that manufactures chips. The most advanced AI chips are made by very few fabs, mostly those of TSMC in Taiwan.

A leading-edge fab costs tens of billions of dollars and relies on extreme ultraviolet (EUV) lithography machines that only one company, ASML in the Netherlands, makes. Companies such as Nvidia, Apple and AMD design chips but pay foundries such as TSMC to build them. A common confusion is that a chip “made by Nvidia” is built in an Nvidia factory: Nvidia designs it, and a foundry manufactures it.

Say it to studentsA fab is a chip factory, and only a few in the world can make the most advanced AI chips.

Moore’s law

Also: Moore’s Law

Gordon Moore’s 1965 observation that the number of transistors on a chip was doubling at a steady pace, later stated as about every two years.

Moore first predicted a doubling every year, then revised it to every two years in 1975. The trend held for decades and made computers vastly cheaper and more powerful. It has slowed as transistors approach the size of a few dozen atoms, so AI progress now also comes from specialized chips, more chips working together and better software. A common confusion is that it is a law of physics: it was an observation about industry progress.

Say it to studentsMoore’s law is the old pattern of chips doubling their transistors about every two years.

CUDA

Nvidia’s software platform for programming its GPUs. Most AI software is built on it, which is one reason Nvidia dominates AI chips.

Released in 2007, CUDA lets programmers use a GPU for general math as well as graphics. Popular AI tools such as PyTorch use it under the hood. Because so much code depends on CUDA, moving to another company’s chips takes real work, which people describe as Nvidia’s software “moat.” A common confusion is that CUDA is hardware: it is software, though it only runs on Nvidia GPUs.

Say it to studentsCUDA is the software that lets programmers use Nvidia’s chips for AI.

Interconnect (NVLink, InfiniBand)

Also: NVLink, InfiniBand, Networking

The very fast links and network that let AI chips share data with each other while they work on the same model.

Inside a server, links such as Nvidia’s NVLink join GPUs together; across a data center, networks such as InfiniBand or fast Ethernet join thousands of servers. Training a large model means chips constantly swap partial results, so slow links leave expensive chips sitting idle. A common confusion is that AI speed is all about the chips: in big systems, the network can matter as much.

Say it to studentsThe interconnect is the super-fast wiring that lets AI chips talk to each other.

AI supercomputer (cluster)

Also: GPU cluster, AI cluster

Thousands of AI chips wired together to act like one giant computer, used to train the biggest models.

Training a frontier model can take weeks or months on tens of thousands of GPUs or TPUs at once, all working on the same model. These clusters fill whole buildings and need their own power supply and cooling. A common confusion is that a supercomputer is one huge machine: it is many ordinary-sized servers connected by very fast networks.

Say it to studentsAn AI supercomputer is thousands of AI chips working together like one enormous computer.

Data center

Also: Data centre, AI data center

A building full of servers, storage and network equipment. AI data centers hold thousands of AI chips, plus the power and cooling they need.

When you use a chatbot, your request travels to a data center, runs on chips there and the answer travels back, usually in seconds. The largest new AI data centers can use as much electricity as a small city, which affects local power grids, water use for cooling and climate goals. A common confusion is that “the cloud” is somewhere abstract: it is real buildings, often in places chosen for cheap power and land.

Say it to studentsA data center is a giant building full of computers. That is where most AI actually runs.

FLOPs and FLOPS

Also: Floating-point operations, Petaflops

Ways of counting computer math. FLOP or FLOPs is a total number of calculations; FLOPS (capital S) means calculations per second, a measure of speed.

A floating-point operation is one arithmetic step with decimal-style numbers, such as one multiplication. Training compute is reported as a total in FLOP, and chip speed in FLOPS. The numbers are enormous: the largest training runs now use more than 10²⁵ (ten million billion billion) operations. A common confusion is mixing the two: one is how much work was done, the other is how fast a chip can do it.

Say it to studentsFLOPs count how many math steps a computer does. AI training takes an astronomical number of them.

Energy and cooling

Also: AI energy use, Power consumption

The electricity AI chips use and the cooling that stops them overheating. Both have grown quickly as AI has scaled up.

Most of the energy goes into running chips in data centers, and some data centers also use large amounts of water for cooling. Estimates for a single chatbot answer vary widely and change as hardware improves, so treat any one number with care. The bigger picture is clearer: AI is a growing share of electricity demand in some regions. A common confusion is that training is the only big cost: answering billions of everyday questions adds up too.

Say it to studentsAI runs on electricity, a lot of it, and the chips have to be kept cool.

Edge AI

Also: On-device AI

AI that runs on or near the device where data is collected, such as a phone, camera or tractor, instead of sending everything to a distant data center.

A farm sprayer that spots weeds with its own camera and chip, and sprays within a fraction of a second, is edge AI: it cannot wait for a round trip to the cloud. Running locally can also keep data on the device and work without internet. The trade-off is that small devices have little memory and power, so edge models are usually smaller, compressed and less capable. A common confusion is that “on-device” always means private: an app can still send your data elsewhere, so check its settings.

Say it to studentsEdge AI does its thinking right on the device, like a phone or camera, instead of far away online.

Quantum computer

Also: Quantum computing

A different kind of computer that uses qubits and the rules of quantum physics. It could speed up some specific problems; it is not a faster version of today’s computers.

Today’s quantum computers are small and make many errors, and they do not run chatbots. Researchers are exploring how AI can help design and correct them, and which problems, such as simulating molecules, they might one day solve faster. A common confusion is that quantum computers will soon replace laptops or GPUs: they are specialized tools for particular kinds of problems.

Say it to studentsA quantum computer works in a completely different way from a normal one, and it is only better at certain problems.

Qubit

Also: Quantum bit

The basic unit of a quantum computer. A bit is 0 or 1; a qubit can be in a blend of both until it is measured, which gives one or the other.

Qubits are fragile: heat, noise and vibration disturb them, which causes errors. Many quantum computers are cooled to near absolute zero. ARQIV’s quantum page has a one-qubit simulation you can try. A common confusion is that a qubit stores both answers and reads them out at once: measuring it gives only one result, and quantum algorithms are designed so the useful answer is the likely one.

Say it to studentsA qubit is a quantum bit. It can be a mix of 0 and 1 until you look at it.

11 · Topic

Words from AI news

Terms you will see in headlines, product launches and company announcements.

Scaling laws

Also: Scaling

The pattern that AI models tend to get predictably better as they are trained with more data, more computing power and more parameters.

Researchers at OpenAI (2020) and DeepMind (2022) measured these trends, and they explain why companies keep building bigger data centers. More recently, companies also scale the computing power used while answering, which is part of how reasoning models work. A common confusion is that they are laws of nature: they describe a trend observed so far, and researchers debate how long it will continue and what it leaves out.

Say it to studentsScaling laws say bigger models trained on more data usually get better, so far.

Emergent abilities

Also: Emergence

Skills that seem to appear suddenly as models get bigger, rather than improving gradually.

Early studies described jumps in skills like multi-step arithmetic at certain model sizes. Later research (Schaeffer and colleagues, 2023) argued that many jumps depend on how a skill is scored: with smoother scoring, improvement often looks gradual. A common confusion is that “it suddenly learned to…” means something mysterious happened: often the measuring stick explains the surprise.

Say it to studentsEmergent abilities are skills that seem to pop up when AI gets bigger, though often they were growing slowly all along.

Small language model (SLM)

Also: SLM, Compact model

A language model small enough to run on a laptop or phone. Less capable than the largest models, but cheaper, faster and more private.

Small models are often made by distilling or quantizing larger ones, or by training carefully on high-quality data. They suit focused jobs such as summarizing, autocomplete or classroom tools that must work offline. A common confusion is that small means useless: on a narrow task, a well-trained small model can match a much bigger one.

Say it to studentsA small language model is a chatbot brain small enough to fit on your own device.

Closed model

Also: Proprietary model

A model whose weights are kept private by the company that made it. You can use it only through the company’s app or API.

Most of the best-known chatbots run on closed models. The opposite is an open-weight model, which anyone can download and run. Closed models let a company control updates and safety features; open models let others inspect, adapt and run them privately. A common confusion is that closed means secret about everything: many companies still publish safety reports and test results.

Say it to studentsA closed model is an AI the company keeps to itself. You can use it, but you can’t download it.

Open-source AI

Also: Open source

AI whose code, and ideally training data and methods, anyone can read, use and change. Many models called “open” share only their weights.

Open source began with software, where it means anyone can see and modify the code. For AI, the Open Source Initiative published a definition in October 2024 that also asks for enough information about the training data and process to rebuild the model. Few popular models meet it. A common confusion is treating “open-weight” and “open source” as the same: open weights let you run a model, but they do not show you how it was made.

Say it to studentsOpen-source AI shares how it was made, not just the finished model, so anyone can study and change it.

Synthetic data

Also: Generated data

Data generated by a computer, often by another AI, instead of collected from the real world. Used to train or test models.

It helps when real data is scarce, private or expensive, for example rare medical cases or practice math problems. It can also copy the mistakes and blind spots of whatever made it. A common confusion is that synthetic data is fake in a bad sense: used carefully and checked against real data, it is a normal tool.

Say it to studentsSynthetic data is practice data made by a computer instead of collected from real life.

Model collapse

The worry that models trained mostly on AI-generated content get steadily worse over generations, losing rare details and variety.

A 2024 study in Nature (Shumailov and colleagues) showed models degrading when each generation trained on the previous one’s output instead of real data. As more of the internet is AI-written, this matters for future training data. A common confusion is that any AI-generated training data causes collapse: keeping plenty of fresh human-made data in the mix greatly reduces the problem.

Say it to studentsModel collapse is what happens when AI learns mostly from AI: copies of copies get blurrier.

Vibe coding

Building software by describing what you want to an AI and accepting the code it writes, with little or no reading of that code yourself.

The AI researcher Andrej Karpathy coined the phrase in February 2025. It can produce quick prototypes and lets beginners make working apps, but it can also hide bugs and security holes, because nobody has checked the code. A common confusion is that it replaces learning to program: someone still needs to understand the code when it breaks, which is a good classroom discussion about who is responsible.

Say it to studentsVibe coding is asking an AI to write a program and trusting it without really checking how it works.

AI companion

Also: Companion chatbot, Character chatbot

A chatbot designed for ongoing friendship, romance or emotional support, often with a name, a personality and a memory of past chats.

Companion apps are popular with teenagers. Several companies have faced lawsuits and government inquiries about harm to young users, and many apps set minimum ages that are easy to get around. A common confusion is that a companion that says it cares actually does: it is generating likely replies, and it may be designed to keep users chatting. Talk with students about what a chatbot can and cannot be for them, and know your school’s guidance.

Say it to studentsAn AI companion can sound like a friend, but it’s a program. Real people are the ones who can truly care for you.

AI slop

Low-quality content mass-produced with AI, such as spammy articles, fake product reviews and strange images, made to get clicks rather than to inform.

Slop fills search results and social feeds because it is cheap to make in bulk. It is usually not meant to fool anyone carefully, only to grab attention or ad money. A common confusion is that everything AI-made is slop: the word describes low-effort, low-value content, whatever tool made it.

Say it to studentsAI slop is junk content made fast with AI to get clicks.

12 · Topic

Safety, society and the law

Privacy, copyright, rules and the wider effects of AI on people.

Student data privacy

Also: Data privacy

Protecting information about students, such as names, grades, photos, voices and anything they type, when using AI tools.

What you type into a chatbot may be stored, read by staff or used to improve the product, depending on the tool and the type of account. Use tools your school has approved, avoid entering personal details about students, and check whether student data is used for training. A common confusion is that a free account and a school account have the same protections: education plans often promise more.

Say it to studentsKeep personal information out of AI tools: no full names, addresses or private details.

FERPA

Also: Family Educational Rights and Privacy Act

A U.S. federal law (1974) that protects the privacy of student education records and gives parents, and students once they turn 18, rights over them.

Schools generally need consent before sharing personally identifiable information from education records, with exceptions such as approved service providers working on the school’s behalf. That is why districts sign agreements with technology vendors. A common confusion is that FERPA covers everything about a student online: it covers education records held by schools, so check with your district before putting student work or records into an AI tool.

Say it to studentsFERPA is the U.S. law that keeps school records private.

COPPA

Also: Children’s Online Privacy Protection Act

A U.S. law that limits how websites and apps collect personal information from children under 13 without a parent’s permission.

It is one reason many AI tools set a minimum age of 13, and some set 18. In some cases a school can give consent for a tool used only for education. A common confusion is that COPPA makes a tool “approved for kids”: it sets rules about data collection, so still check each tool’s terms, age limits and your school’s approved list.

Say it to studentsCOPPA is the U.S. law that protects personal information of kids under 13 online.

AI regulation (EU AI Act)

Also: EU AI Act, AI laws

Laws and rules about how AI is built and used. The EU AI Act, passed in 2024, is the first broad AI law, with stricter rules for riskier uses.

The EU AI Act bans some uses outright, such as social scoring and emotion recognition in schools and workplaces, and sets strict rules for “high-risk” uses, which include some in education such as scoring exams. Its rules take effect in stages from 2025, and some dates are still being adjusted. In the U.S., rules come mostly from states, existing laws and agency guidance. A common confusion is that there is one set of AI rules: they differ by country and state and change often.

Say it to studentsAI regulation is the laws and rules about what AI is allowed to do.

Jailbreak

Also: Jailbreaking

A trick prompt designed to get an AI to ignore its safety rules, for example through role-play or by hiding a request inside a story.

Companies patch known jailbreaks, and people find new ones, so it is a constant back and forth. A jailbreak is typed by the user, while a prompt injection is hidden in content the AI reads. For classrooms: assume students will try, keep tools on approved settings, and treat it as a digital citizenship conversation. A common confusion is that a jailbroken answer reveals the AI’s “true self”: it is still generated text, just without the usual filters.

Say it to studentsA jailbreak is a trick to make an AI break its own rules.

Misinformation and disinformation

Also: Disinformation

False information. Misinformation is spread by mistake; disinformation is spread on purpose to mislead. AI makes both cheaper and faster to produce.

AI can produce convincing fake articles, images, voices and comments in bulk, and a chatbot’s own hallucinations can become misinformation when people share them. The defenses are habits more than tools: stop, check the source, read laterally and look for the original. A common confusion is that only obviously fake things spread: realistic, partly true content is often the most misleading.

Say it to studentsMisinformation is wrong information shared by accident. Disinformation is shared on purpose to fool people.

Voice cloning

Also: Cloned voice, Synthetic voice

Using AI to copy a real person’s voice from a short recording, so it can say things they never said.

It has good uses, such as giving a voice back to people who lost theirs, and harmful ones, such as scam calls that sound like a family member asking for money. Agreeing on a family “safe word” is a simple defense. A common confusion is that you need hours of recordings: some tools need only a few seconds.

Say it to studentsVoice cloning copies someone’s voice with AI. If a call sounds urgent and strange, check another way.

13 · Topic

Building and running AI

How large language models are trained, tuned and served, and how real AI apps are built, run, tested and secured.

State-space model (Mamba)

Also: Mamba, SSM

A kind of language model that reads text by updating a fixed-size running summary, token by token, instead of looking back at every earlier token the way attention does.

Because the summary never grows, cost rises in a straight line with the length of the text, so very long inputs are cheap. The best-known design is Mamba (Gu and Dao, 2023). A fixed summary can lose exact details, so many newer models are hybrids that mix state-space layers with some attention layers.

Say it to studentsInstead of rereading the whole book for every new word, it keeps one page of notes and updates it.

Diffusion language model

A language model that starts from a rough or masked draft of the whole answer and refines all of it over several passes, instead of writing one token at a time.

The idea is borrowed from image generators. Because many tokens are produced in parallel, these models can be very fast; examples from 2025 include Inception’s Mercury and Google’s Gemini Diffusion research model. They are newer and less proven than ordinary language models on long, careful reasoning.

Say it to studentsLike sketching a whole drawing faintly and sharpening it, instead of drawing one line at a time.

Supervised fine-tuning (SFT)

Also: SFT, Instruction tuning

Training a pre-trained model on many example prompts paired with good answers, so it learns to respond helpfully instead of just continuing text.

SFT usually comes right after pre-training and before preference tuning. It teaches the format and style of a helpful assistant more than new knowledge. The examples are written or checked by people, or generated by another model and filtered.

Say it to studentsSFT is showing the model thousands of good answers so it learns what a good answer looks like.

Reward model

A separate model trained to predict which of two answers people would prefer. It gives a score that is used to train a chatbot with reinforcement learning.

In RLHF, people rank several answers to the same prompt, a reward model learns those rankings, and the chatbot is then trained to score highly on it. The weak spot is reward hacking: the chatbot can learn what pleases the reward model, such as long, confident answers, rather than what is actually good.

Say it to studentsA reward model is an automatic judge that learned what people tend to like.

Direct preference optimization (DPO)

Also: DPO

A way to tune a model on pairs of answers, one preferred and one not, without training a separate reward model or running reinforcement learning.

DPO (Rafailov and colleagues, 2023) uses the same “answer A beats answer B” data as RLHF but adjusts the model directly with an ordinary training step that makes preferred answers more likely. It is simpler, cheaper and more stable, so open-model builders adopted it widely.

Say it to studentsDPO teaches by comparison: more like this answer, less like that one.

RL with verifiable rewards (RLVR)

Also: RLVR, Verifiable rewards

Reinforcement learning where the reward comes from a program that checks the answer, such as “is the final number right?” or “do the unit tests pass?”, instead of from human judgments.

The name comes from the Tülu 3 paper (Lambert and colleagues, 2024). A checker is hard to fool, which made RLVR the main way reasoning models were trained from 2025. It only works where answers can be checked automatically, so essays and advice still need human preferences.

Say it to studentsRLVR is practice with an answer key: the model only gets credit when the checker says it is right.

Group relative policy optimization (GRPO)

Also: GRPO

A reinforcement-learning method where the model writes a group of answers to the same problem and each answer is rewarded for being better than the group’s average.

Introduced in the DeepSeekMath paper (Shao and colleagues, 2024) and used for DeepSeek-R1 (2025). Comparing answers within a group removes the need for a separate “critic” model, which saves memory, and it pairs naturally with checkable rewards.

Say it to studentsGRPO grades on a curve: answers that beat their classmates get rewarded.

Test-time scaling

Also: Test-time compute, Inference-time scaling

Making answers better by spending more computing when a question is asked, such as thinking longer or trying several answers and picking the best, instead of training a bigger model.

Methods include longer chains of thought, voting across several answers (self-consistency) and best-of-N with a checker. A 2024 study by Snell and colleagues found that for some problems extra thinking beat a model 14 times larger. The price is more tokens: more cost and a longer wait.

Say it to studentsTest-time scaling is giving the model more time on the test instead of more years of school.

Recursive language model (RLM)

Also: RLM

A way to handle very long inputs: the model keeps the text in a small coding workspace, writes code to look through it, and calls copies of itself on the pieces.

Proposed by Zhang, Kraska and Khattab (MIT, 2025), who report handling inputs about 100 times longer than the model’s context window. Careful: “RLM” is also used to mean “reasoning language model”, so check which one a writer means.

Say it to studentsInstead of reading a whole library at once, the model sends helpers to read the shelves and report back.

Parameter-efficient fine-tuning (PEFT)

Also: PEFT

A family of fine-tuning methods that train only a tiny fraction of a model’s numbers and keep the rest frozen, so adapting a model is much cheaper.

LoRA and QLoRA are the best-known PEFT methods. Because the original model is unchanged, many small adapters can be trained for different jobs and swapped in as needed.

Say it to studentsPEFT is adding a small attachment to a tool instead of rebuilding the whole tool.

LoRA (low-rank adaptation)

Also: Low-rank adaptation

A fine-tuning method that freezes a model’s weights and trains two thin extra grids of numbers next to them. The small result can be saved and swapped in like a plug-in.

Introduced by Hu and colleagues (Microsoft, 2021). The width of the thin grids is the “rank”, often 8 or 16. On GPT-3, the paper cut the number of trained numbers by 10,000 times and memory needs by 3 times, with similar quality.

Say it to studentsLoRA is a clip-on lens: the camera stays the same, the lens changes what it does.

QLoRA

LoRA with one extra trick: the frozen model is stored in 4-bit numbers while the small LoRA grids are trained, which cuts memory needs further.

QLoRA (Dettmers and colleagues, 2023) made it possible to fine-tune a 65-billion-parameter model on a single 48 GB graphics card.

Say it to studentsQLoRA squeezes the big model down so it fits, then trains the small add-on.

KV cache

Also: Key-value cache

Memory a server keeps during a conversation so the model does not have to redo its attention calculations for every earlier token each time it writes a new one.

For each token the model stores two lists of numbers, keys and values. Keeping them saves huge amounts of computing, but the cache grows with every token and every user, and for long chats it can need more memory than the model itself. It is also what makes prompt caching possible.

Say it to studentsThe KV cache is the model’s notes on everything said so far, so it does not reread the whole chat each time.

Speculative decoding

A speed-up where a small, fast model guesses the next several tokens and the big model checks them all at once, keeping the ones it agrees with.

Described by Leviathan and colleagues and by Chen and colleagues (2023). Because the big model has the final say, the output matches what it would have written anyway, often two to three times faster.

Say it to studentsA student drafts the next few words and the teacher approves them at a glance.

vLLM (serving engine)

Also: Serving engine, PagedAttention

A free, open-source program that runs language models for many users at once, using memory efficiently so more conversations fit on each chip.

Started at UC Berkeley. Its key idea, PagedAttention (Kwon and colleagues, 2023), stores the KV cache in small fixed-size pages, and continuous batching lets new requests join as others finish. Similar engines include SGLang and NVIDIA TensorRT-LLM; llama.cpp and Ollama run models on laptops.

Say it to studentsvLLM is the kitchen system that lets one chef cook many orders at once without wasting counter space.

Context rot

The drop in a model’s accuracy as its input gets longer, even when everything still fits inside its context window.

A July 2025 study from Chroma tested 18 models and found accuracy fell as inputs grew, even on simple tasks; an earlier study found information in the middle of long prompts was used worst (“lost in the middle”). Builders respond with context engineering: compaction, notes and loading only what is needed.

Say it to studentsA bigger backpack does not help if you cannot find anything in it.

Compaction

Summarizing the older part of a long conversation or agent run, keeping the decisions and key facts, so the work can continue in a fresh context window.

Used when a context window nears its limit. The risk is that a detail the summary dropped turns out to matter, so agents often also write notes to files they can reread later.

Say it to studentsCompaction is writing a one-page recap so you can start a clean page and keep going.

Prompt caching

Reusing the processed form of a prompt’s fixed opening, such as long instructions, so repeated requests are faster and cheaper.

Providers keep the KV cache for a prompt prefix they have just seen and bill cached input at a large discount, around 50 to 90% off depending on the provider. It only works if the start of the prompt is exactly the same, so builders put fixed parts first and changing parts last.

Say it to studentsPrompt caching means the model does not have to reread the same instructions every time.

AI memory (short-term and long-term)

Also: Episodic memory, Semantic memory, Long-term memory

Information an AI app saves and puts back into the model’s context later. The model itself remembers nothing between calls.

Short-term memory is the current conversation in the context window. Long-term memory is stored outside: episodic (past events and chats), semantic (facts and preferences) and procedural (how to do things, such as skills). Risks include wrong or outdated memories, private details surfacing in the wrong place, and memory poisoning; users should be able to see and delete what is kept.

Say it to studentsThe AI does not remember you; the app keeps notes and shows them to the AI.

Reranking

A second pass in search where a slower, smarter model rereads the top candidates next to the question and puts the best ones first.

A fast search fetches, say, 50 candidates; a reranker keeps the best 5. Fewer, better passages make the final answer cheaper and more accurate.

Say it to studentsReranking is a quick skim to find 50 books, then a careful look to pick the 5 that really answer the question.

GraphRAG

A retrieval method that first builds a knowledge graph of the people, places and things in a set of documents, then uses it to answer questions, including big-picture ones.

Described by Microsoft Research in 2024. Summaries of clusters in the graph help with questions like “What are the main themes across these 500 reports?”, which ordinary chunk search answers poorly. Building the graph costs much more than plain RAG.

Say it to studentsGraphRAG draws a map of who and what is connected before answering.

A2A (Agent2Agent protocol)

Also: Agent2Agent

An open standard that lets AI agents built by different companies find each other, describe what they can do, hand over tasks and report progress.

Announced by Google in April 2025 and now run under the Linux Foundation. Each agent publishes an “agent card” listing its skills. MCP connects an agent to tools and data; A2A connects it to other agents, and many systems use both.

Say it to studentsMCP gives an agent hands; A2A gives it colleagues.

Agent framework

Also: Agent SDK, LangGraph, Google ADK, Microsoft Agent Framework

A software toolkit that handles the plumbing of an AI agent, such as the loop, tool calls, saved state and logging, so builders do not write it from scratch.

Widely used examples in 2026: Google’s Agent Development Kit (ADK), Microsoft Agent Framework (successor to Semantic Kernel and AutoGen, 1.0 in April 2026), LangGraph, and SDKs from model makers such as the Claude Agent SDK and OpenAI Agents SDK. They change fast, so the ideas matter more than the names; simple apps often need none.

Say it to studentsAn agent framework is a kit with the engine already built, so you only add the parts for your job.

Checkpointing (in AI agents)

Saving an agent’s progress after each step so that a long run can resume where it stopped after a crash, a timeout or a pause for a person’s approval.

Frameworks such as LangGraph save state at each step; managed agent services keep a session log outside the agent so a fresh copy can pick up the work. (In model training, “checkpoint” also means a saved copy of the weights.)

Say it to studentsCheckpointing is saving your game so a crash does not send you back to the start.

Human-in-the-loop (HITL)

Also: HITL

A design where an AI system pauses for a person to approve, correct or reject a step, usually before risky actions such as sending money, emailing outsiders or deleting files.

Good HITL asks rarely and clearly: people shown too many approval pop-ups stop reading them. It works best together with limits the AI cannot change, such as permissions and sandboxes.

Say it to studentsHuman-in-the-loop means the AI asks “May I?” before anything it cannot undo.

Cloud AI platform

Also: Amazon Bedrock, Microsoft Foundry, Vertex AI

A cloud company’s service for using AI models and running AI agents, bundled with security, monitoring and billing.

Main examples as of November 2026: Amazon Bedrock and Bedrock AgentCore (AWS); Microsoft Foundry (called Azure AI Foundry until late 2025); and Google’s Gemini Enterprise Agent Platform (called Vertex AI until April 2026). Names change often; the parts stay the same: a model catalog, an agent runtime and the company’s security around them.

Say it to studentsA cloud AI platform is renting AI from a big company instead of running it on your own computers.

Serverless and containers

Also: Container, Cloud Run, AWS Lambda, Kubernetes, K8s

Ways to run app code in the cloud. A container packs an app with everything it needs; serverless services run it only when requests arrive and scale down to zero when idle.

Examples: Google Cloud Run and AWS Lambda (serverless), AWS ECS and Azure App Service (managed containers and web apps), and Kubernetes (K8s) for large systems with many services or self-hosted models on GPUs. Serverless can have slow “cold starts” and time limits (a Lambda function stops after 15 minutes), which matters for long agent runs.

Say it to studentsServerless means you pay for the kitchen only while someone is cooking.

Queue and async workflow

Also: Async, Event-driven, Webhook

A way to handle slow jobs: the app puts each job in a queue and answers “working on it”, and background workers process the jobs and report back when done.

Used for long AI tasks such as grading a class set of essays. If a worker crashes, the job stays in the queue and is retried. In event-driven systems, parts announce events (“essay submitted”) and anything interested reacts. Jobs should be idempotent, so a retry never does the work twice.

Say it to studentsA queue is a ticket line: you get a number and are called when your order is ready.

Model routing

Also: Model router

Sending each request to the cheapest model that can handle it well, so easy questions go to a small model and only hard ones to a big, expensive one.

A quick classifier makes the choice. A 2024 study (RouteLLM) cut costs by more than half while keeping quality. A misrouted hard question gets a weak answer, so routing has to be tested.

Say it to studentsModel routing sends easy questions to the quick helper and hard ones to the expert.

Semantic caching

Saving past answers and reusing one when a new question means nearly the same thing as an old one, judged with embeddings, so the model is not called at all.

It can cut costs a lot for repeated questions. Risks: a near-match can be wrong (“retake for chemistry” vs. “for physics”), saved answers go stale, and one person’s private answer must never be reused for someone else.

Say it to studentsSemantic caching is answering “When is the field trip?” the same way you did five minutes ago.

Prompt compression

Shortening long prompts and retrieved text before they reach the model, by hand or with tools that drop low-value tokens, to save cost and time.

Tools such as Microsoft’s LLMLingua (2023) remove words that matter little to the answer. Squeeze too hard and key details vanish, so the result must be tested on the real task.

Say it to studentsPrompt compression is cutting a long note down to what the reader really needs.

Observability and tracing

Also: Tracing, Trace, OpenTelemetry, Langfuse, LangSmith

Recording every step an AI app takes, such as model calls, searches and tool calls, with timing, tokens and cost, so problems can be found and explained.

A trace follows one request from start to finish; each step in it is a span. OpenTelemetry is the open standard for collecting this data, and tools such as Langfuse (open source) and LangSmith add AI-specific views. Logs contain whatever users typed, so collect only what is needed and protect it.

Say it to studentsTracing is a receipt for every step the AI took, so you can see where it went wrong.

LLM-as-a-judge

Also: AI grader

Using a strong language model to grade other AI answers against a rubric, so thousands of answers can be checked without a person reading each one.

A 2023 study found GPT-4’s judgments agreed with people about as often as people agreed with each other. Judges have known biases: they favor longer answers, the first answer shown and sometimes their own writing, so they should be checked against human grades.

Say it to studentsAn AI judge is fast, but someone still has to check the judge.

RAGAS

An open-source toolkit (2023) for grading retrieval apps, with scores such as faithfulness, answer relevancy, and context precision and recall.

Faithfulness asks whether every claim is supported by the retrieved passages; answer relevancy whether the question was answered; context precision and recall whether search found the right passages, and only those. Many scores are computed with an LLM judge.

Say it to studentsRAGAS checks whether the answer really came from the pages the AI found.

Agent evaluation

Also: Agent evals, pass^k

Testing an AI agent on whole tasks: checking the end result, the steps it took, and how reliably it succeeds when the same task is run several times.

End-state checks ask whether the job was really done (once, not twice); path checks look at tool use and whether it asked before risky steps. Because agents vary from run to run, tests repeat each task: the τ-bench study (2024) calls the chance of succeeding on all of k tries pass^k.

Say it to studentsOne lucky success does not prove an agent is reliable; run it again and again.

Role-based access control (RBAC)

Also: RBAC, Least privilege

A security rule where what someone may see and do depends on their role, such as student, teacher or admin. AI tools should act with the asking user’s permissions, never more.

Giving an AI only the access the current user has (least privilege) means that even a fooled AI, for example through prompt injection, cannot reach data the user could not. This protects better than instructions alone, which can be overridden.

Say it to studentsIf you could not open the file yourself, the AI working for you should not be able to either.

LLMOps (CI/CD and model versioning)

Also: CI/CD, Model versioning

The practices for running language-model apps reliably: keeping prompts, models and tests under version control and testing every change automatically before it ships.

LLMOps is MLOps for language-model apps. CI/CD (continuous integration and delivery) runs the evals on each change and rolls it out gradually with a quick way back. Teams pin an exact model version, because providers update and retire models and the same prompt can behave differently on a new one.

Say it to studentsLLMOps means no change goes live until it passes the tests, and you can always undo it.

Definitions are simplified on purpose and checked against the sources on our AI Learning Center pages. Spotted something wrong or missing? Tell us and we will fix it.