AI Glossary

grep -ri "{term}" glossary/

76 AI terms in plain English. Search them, filter by topic, or switch to test yourself and flip the page into flashcards.

find

topic

mode

Showing 76 of 76 terms

foundations

Algorithm

A precise, step-by-step procedure for doing something. In AI the word covers both the training procedure that adjusts a model's weights and, more loosely, the trained model itself.

Artificial intelligenceAI

Computer systems doing tasks that normally need human intelligence — understanding language, recognising images, making decisions. Today the term mostly refers to systems that learn from data rather than following hand-written rules.

Dataset

The collection of examples a model is trained or tested on. What is in the dataset decides what the model can and cannot learn — including whose languages and faces it works well for.

Deep learning

Machine learning built on neural networks stacked many layers deep. Almost everything called AI today — chatbots, image generators, speech recognition — is deep learning underneath.

Generative modelgenerative AI

A model that produces new content — text, an image, a DNA sequence — rather than only classifying or scoring existing content.

Inference

Using a trained model to get an answer, as opposed to training it. Cheap compared to training.

Every message you send a chatbot triggers one round of inference.

Labelled data

Examples where somebody has already written down the correct answer. Expensive to produce, especially in medicine and in under-documented languages.

An hour of audio plus its exact transcript is labelled data for speech recognition.

Machine learningML

Building software by showing a computer many examples and letting it work out the pattern, instead of programming the rules by hand. The thing this process produces is a model.

Model

The thing training produces: a program, defined by its weights, that turns inputs into outputs. A language model turns text into more text; a vision model turns an image into a label.

Neural network

A model built from layers of simple units that each pass numbers on to the next layer. Loosely inspired by neurons in the brain, but in practice it is a very large pile of arithmetic whose weights are tuned during training.

Parameters

Another word for weights. When you read that a model has 40 billion parameters, that is how many numbers it contains.

A "7B model" is a model with seven billion parameters.

Training

Showing a model examples and adjusting its weights so its outputs improve. Expensive and slow.

Weights

The numbers inside a model that determine how it turns an input into an output. Also called parameters. Training means adjusting them.

language models

Attention

The mechanism inside a transformer that lets the model decide, for each token, which other tokens matter most right now. It is how the word "it" in a sentence gets connected to the thing it refers to.

Backbone

The large early portion of a model that turns raw input into an embedding. The expensive, general part.

Context windowcontext length, context

Everything you send the model at the moment you ask it something. Think of it as the model's desk: only what is on the desk gets considered.

Embeddingembeddings

A list of numbers — usually a few hundred to a few thousand — that summarises one input. Two similar inputs get similar embeddings.

Fine-tuningfinetuning

A second, much smaller round of training that adjusts an already-trained model to your specific task using your own data.

Foundation model

A big model pretrained once on broad data so it can be adapted to many different tasks, rather than built for one job. Large language models are the best-known examples, but the same idea now runs in biology, vision and speech.

Frozen

A deliberate decision not to change any of a model's weights. You only push data through it and collect the results.

Head

The small final portion of a model that turns an embedding into an actual answer. Cheap to replace and retrain.

Instruction tuning

Fine-tuning a model on examples of instructions paired with good responses — the step that turns a raw next-word predictor into an assistant that follows directions.

Large language modelLLM

A model trained on enormous amounts of text to predict the next token — which turns out to be enough to draft, summarise, translate, answer questions and write code. ChatGPT, Claude and Gemini are all built on large language models.

Open weightsopen source model, open model

The model's numbers are published, so anyone can download, inspect, modify and retrain it. The opposite is a closed model, which you can only send requests to over the internet.

Pretrainingpre-training

The first, biggest round of training, done on a huge amount of general data before anyone has a specific task in mind. This is what makes a model a foundation model.

RLHFreinforcement learning from human feedback

Reinforcement learning from human feedback: people rate pairs of model answers, and the model is trained to prefer the kind of answer people rated higher. A key step in making chat models helpful and polite.

Token

The small chunk a model breaks its input into — roughly a word for text, a single DNA letter for some genome models, a small square for images.

"unbelievable" might be split into the tokens "un", "belie" and "vable".

Tokenizertokenization

The fixed procedure that chops text into tokens before the model sees it. Trained mostly on English, so it often splits other languages into many more pieces — making them slower and more expensive to process.

Transformer

The neural-network design behind almost every modern language model, introduced in 2017. Its key ingredient is attention, which lets the model weigh every part of the input against every other part.

using models

API

The doorway a program uses to talk to a model over the internet: send a request carrying a prompt, get the model's output back. Most AI products are built on a model provider's API rather than on a model the builder trained.

Chain of thoughtreasoning

Having the model write out intermediate reasoning steps before its final answer, which measurably improves accuracy on maths and logic. Newer "reasoning" models do a version of this internally by default.

Few-shotin-context learning

Putting a handful of worked examples in the prompt so the model copies the pattern. No training happens — the examples live only in the context window.

Grounding

Tying a model's answer to specific, checkable source material — retrieved documents, search results, a database — rather than letting it answer from memory alone. Reduces hallucination; does not eliminate it.

Prompt

The text you send a model: the question, the instructions, and any material you paste in. The model's answer depends heavily on how the prompt is worded.

Prompt engineering

Writing and refining prompts so the model reliably does what you want — clear instructions, worked examples, an explicit output format. Less mysterious than it sounds: mostly clear writing plus testing.

RAGretrieval-augmented generation, retrieval augmented generation

Retrieval-augmented generation: before the model answers, the system searches your documents and pastes the most relevant passages into the prompt. The standard way to get a model answering from your own material instead of only its training data.

System prompt

A standing instruction given to the model before the conversation starts, setting its role, tone and rules. Users usually never see it.

"You are a customer-support assistant for a bank. Never give investment advice."

Temperature

A setting that controls how much randomness goes into the model's choice of each next token. Low temperature makes answers more predictable; high makes them more varied. Even at zero, answers are not perfectly repeatable.

Vector databasevector store

A database that stores embeddings and can quickly find the entries most similar to a query. The search engine behind most RAG systems.

Zero-shot

Getting a useful answer from a model without training it on that task at all — just by asking.

agents

AgentAI agent

An AI system that does not just answer, but takes actions toward a goal — calling tools, reading the results and deciding what to do next in a loop, with a model making the decisions.

Agent memorymemory

What an agent retains beyond the current conversation — notes, past decisions, user preferences — usually stored outside the model and pasted back into the context window when relevant.

Agentic AI

The broad label for systems built around agents: AI that plans, uses tools and carries out multi-step work with limited supervision, rather than answering one prompt at a time.

Human-in-the-loopHITL

A design where a person reviews or approves the AI's work at key steps rather than letting it act fully autonomously. The standard safeguard for consequential actions like payments, medical decisions or deletions.

MCPModel Context Protocol

The Model Context Protocol: an open standard for connecting AI applications to tools and data sources, so any compliant model can use any compliant tool without custom wiring for each pair.

Multi-agent systemmulti-agent

Several agents working on one task — splitting the work, checking each other's output, or debating. Useful when one model call cannot hold the whole job.

Orchestration

The code around the model that runs the loop: sending prompts, executing tool calls, handling errors, and deciding when the job is done. In an agent, the model chooses; the orchestration executes.

Tool usefunction calling

Letting a model call functions you define — search the web, query a database, send an email — by writing a structured request that your code then executes. Also called function calling. It is how a model gets hands.

A travel agent bot calling a real flight-search function instead of guessing prices.

evals & safety

Alignment

The problem of making AI systems actually pursue what their builders and users intend — helpful, honest, harmless — rather than whatever their training accidentally rewards.

Baseline

A deliberately simple method you compare a fancy model against. If the simple method does just as well, the fancy model has not earned its place.

Benchmark

A standard test set that different models are scored on so they can be compared.

Bias

Systematic unfairness in a model's outputs, usually inherited from the data it was trained on — for example, speech recognition that is less accurate for some accents than for others.

Data contaminationcontamination

When a model's training data accidentally contained the test questions, so its benchmark score measures memory rather than ability. A constant worry when comparing models on public benchmarks.

Evaluationevals, eval

Systematically measuring how well a model or AI system performs — on benchmarks, on your own test cases, or with human graders. Shortened in practice to evals. The unglamorous work that separates a demo from a product.

Guardrails

Checks wrapped around a model that block unwanted inputs or outputs — filtering unsafe content, keeping an agent inside its permitted actions, stopping leakage of private data.

Hallucination

When a model produces a confident, fluent answer that is simply untrue. Prompting harder does not fix it; grounding the model in real sources helps.

Jailbreakjailbreaking

A prompt crafted to trick a model into ignoring its safety rules. An arms race: providers patch known jailbreaks, and new ones appear.

LLM-as-a-judgeAI judge

Using one model to grade another model's answers, because human grading is slow and expensive. Widely used and genuinely useful — but judges have measurable biases, such as favouring answers by position or length.

Overfitting

When a model memorises the specific examples it was trained on instead of learning the general pattern, so it performs well in testing and badly in reality.

Red-teamingred team

Deliberately attacking your own AI system — adversarial prompts, jailbreaks, strange edge cases — to find the failures before real users or real adversaries do.

Word error rateWER

The standard score for speech recognition: the share of words the system got wrong. Lower is better.

hardware & efficiency

Compute

Shorthand for raw processing power — how many chips, running for how long. The scarcest, most expensive input in modern AI; access to compute increasingly decides who can build and run models.

Data centredata center

A building full of servers where models are trained and run. When you use an AI chatbot, the work happens in a data centre that may be on another continent — with real consequences for latency and data rules.

Distillationknowledge distillation

Training a small model to imitate a large one's outputs, keeping much of the ability at a fraction of the cost. Many fast, cheap models are distilled versions of bigger ones.

GPUgraphics processing unit

A specialised chip originally built for video games, now the standard hardware for training and running AI models.

Latency

How long you wait between sending a request and getting the answer. Depends on model size, server load, and the network distance to the data centre serving you.

On-device AIedge AI

Running a model directly on a phone or laptop instead of on a distant server. No connectivity needed and data stays local — but only smaller models fit.

Quantizationquantisation

Storing a model's weights as less precise numbers so the model needs less memory and runs faster, at a small cost in quality. A big part of what makes large models runnable on small machines.

beyond text

Computer vision

Teaching machines to understand images and video — recognising objects, reading scenes, spotting defects on a production line or diseases in medical scans.

Diffusion modeldiffusion

The model family behind most AI image generators. It learns to turn pure noise into an image step by step, guided by your text description.

Low-resource language

A language with little digitised text or transcribed speech available for training. Most of the world's languages are low-resource, and models serve them noticeably worse.

Machine translationtranslation

Automatically translating text between languages. Quality tracks the amount of training data available per language, so widely spoken languages fare far better than others.

OCRoptical character recognition

Optical character recognition: extracting the text out of an image of a document — a scan, a photo of a form — into characters a computer can search and edit.

Speech recognitionASR, automatic speech recognition, speech-to-text

Turning spoken audio into written text; the technical term is automatic speech recognition, or ASR. Accuracy is scored by word error rate — and varies sharply across languages and accents.

Synthetic data

Training data generated by a model rather than collected from the world, used when real data is scarce, private or expensive. Helpful in moderation; a model trained mostly on model output degrades.

Text-to-speechTTS

Generating spoken audio from written text. Modern systems can clone a voice from a short sample — useful, and an obvious fraud risk.