Education

Glossary Wiki

Key terms in AI, explained simply.

137 terms
A B C D E F G H I J K L M N O P Q R S T U V W X Y Z

A

Agent
An AI system that can take actions on its own, such as using tools, browsing, or running code, to complete a goal rather than just answering in text.
AGI (Artificial general intelligence)
A hypothetical AI that can learn and perform any intellectual task a human can, rather than being good at just one narrow job.
AI safety
The field working to keep AI systems from causing harm, from today's misuse to risks from far stronger future systems.
AI slop
Low-effort machine-generated content flooding the web, recognisable by its confident blandness.
Air-gapped
A machine kept fully off the network; the strictest way to guarantee an AI tool cannot send anything anywhere.
Alignment
The work of making sure an AI system actually does what people intend and holds to human values, especially as it becomes more capable.
API (Application programming interface)
The doorway one program uses to talk to another; AI providers sell model access through APIs priced per token.
Artificial intelligence (AI)
Software that performs tasks we usually associate with human intelligence, like understanding language, recognising images, or making decisions.
Attention
The mechanism inside a transformer that lets a model weigh which earlier words matter most when predicting the next one.

B

1-bit model
A model whose weights are stored as single bits instead of 16, shrinking a 27B model to a few gigabytes with modest quality loss.
Backpropagation
The method that works out how much each weight contributed to an error, so gradient descent knows what to adjust.
Base model
A model fresh from pre-training, before instruction tuning; it continues text rather than answering questions.
Benchmark
A standard test or dataset used to measure and compare how well AI models perform on a task.
Bias
Systematic skew in a model's outputs, usually learned from patterns in its training data, that can make results unfair or inaccurate.

C

Catastrophic forgetting
When fine-tuning on new material erases skills a model already had.
Chain of thought
A prompting style, or model behaviour, where the AI works through a problem step by step instead of jumping straight to an answer.
Checkpoint
A saved snapshot of a model's weights during training, used to resume, compare or roll back.
Chunking
Splitting documents into pieces small enough for retrieval and context windows; the size chosen quietly shapes answer quality.
Closed model
A model reachable only through its maker's API, with weights, training data and inner workings kept private.
Compute
The raw processing power, usually measured in chips and hours, used to train or run an AI model.
Context window
The amount of text a model can consider at once, including your prompt and its own reply. Anything beyond it is forgotten.

D

Data labeling
The human work of tagging training examples, telling the model what counts as correct.
Dataset
The collection of text, images or other examples a model learns from; its quality and coverage shape what the model can do.
Deepfake
AI-generated audio, image or video that convincingly imitates a real person.
Dense model
A model that uses all its parameters for every token, unlike mixture-of-experts designs that activate only a few.
Diffusion model
A model that creates images, audio, or text by starting from random noise and gradually refining it into a clear result.
Distillation
Training a smaller model on the outputs of a stronger one so it inherits capability at a fraction of the cost.

E

Embedding
A list of numbers that represents the meaning of a word, sentence, or image so a computer can compare how similar two things are.
Embodied AI
AI that acts through a physical body, usually a robot, closing the loop between perception and action.
Emergent capabilities
Skills that appear once models pass a certain scale without being trained for directly, like arithmetic or translation.
Epoch
One full pass through the training dataset; models usually train for several.
Evals
Structured tests that measure what a model can actually do, from exam questions to full task simulations.

F

Federated learning
Training one shared model across many devices where the data stays put and only weight updates travel.
Few-shot learning
Getting a model to do a task by showing a handful of examples in the prompt, with no retraining.
Fine-tuning
Taking an already trained model and training it a bit more on a smaller, specific dataset so it is better at a particular task or style.
Foundation model
A large model trained on broad data that can be adapted to many different tasks, serving as a base others build on.
Frontier model
One of the most capable models available at a given moment, usually from a handful of large labs.

G

GAN (Generative adversarial network)
A setup where two networks compete, one creating fake data and one judging it, which pushes the creator to produce realistic results.
GGUF
The file format local runners use for models, packing weights and metadata into one portable, often quantized file.
GPU (Graphics processing unit)
A chip originally built for graphics that turned out to be ideal for the parallel maths behind training and running AI models.
Gradient descent
The core training algorithm: nudge each weight a little in the direction that reduces the model's error, millions of times.
Ground truth
The verified correct answer a model's output is judged against.
Guardrails
Rules and filters added around a model to stop it producing harmful, unsafe, or off-limits responses.

H

Hallucination
When an AI states something false or made up as if it were true, because it predicts plausible text rather than checking facts.
Hyperparameters
The settings chosen before training or generation, like learning rate or temperature, tuned by people rather than learned.

I

In-context learning
A model's ability to pick up a pattern from examples inside the prompt and apply it immediately, without weight updates.
Inference
The act of running a trained model to get an answer. This is what happens every time you send it a prompt.
Instruct model
A base model after instruction tuning, trained to follow requests and hold conversations; the kind most people actually use.
Instruction tuning
Fine-tuning a base model on example instructions and answers so it learns to follow requests instead of just continuing text.
Interpretability
The study of what is actually happening inside a model's weights and activations, and why it produced what it did.

J

Jailbreak
A prompt crafted to push a model past its safety rules and produce output it was trained to refuse.

K

Knowledge cutoff
The date a model's training data ends; anything after it the model cannot know without being told.
KV cache
Memory that stores what a model has already read in a conversation so it is not recomputed for every new token; it grows with context length.

L

Large language model (LLM)
A model trained on huge amounts of text that predicts and generates language, powering tools like chatbots and writing assistants.
Latency
The wait between sending a request and getting a response; for chat, the wait before the first word appears.
Latent space
The internal map where a model places concepts, where distance means similarity and directions can carry meaning.
Local AI
Running models on hardware you control instead of a cloud API: private, offline-capable and free of per-token bills.
Logits
The raw scores a model assigns every possible next token before they are turned into probabilities.
LoRA (Low-rank adaptation)
A fine-tuning method that trains a small add-on layer instead of the whole model, cutting memory and cost sharply.
Loss function
The score a model tries to minimise during training; it measures how far predictions are from the right answers.

M

Machine learning
A branch of AI where systems learn patterns from data and improve with experience, instead of being programmed with fixed rules.
MCP (Model Context Protocol)
An open standard for connecting an AI assistant to outside tools and data, so a connector written once works in any compliant app.
Memory bandwidth
How fast a chip moves data to its processors; for local text generation it matters more than raw compute.
Mixture of experts (MoE)
A model design that routes each input to a few specialised sub-networks instead of using the whole model, saving compute.
Model card
A model's spec sheet: what it was trained on, what it is for, its limits and its licence.
Model extraction
Rebuilding a rival model's capability by harvesting its outputs at scale, usually against the provider's terms.
Model merging
Combining the weights of two fine-tuned models into one, a common local-scene way to blend abilities without training.
Model server
A program, like LM Studio or Ollama, that loads a model and serves it to other apps through an API.
Multi-agent system
A setup where several AI agents work on one task, splitting roles like researcher, writer and reviewer.
Multimodal
An AI that can work with more than one kind of input or output, such as text, images, audio, and video together.

N

Neural network
A model loosely inspired by the brain, made of layers of connected units that adjust as they learn from data.
NPU (Neural processing unit)
A chip built specifically for AI maths, common in phones and new laptops, using far less power than a GPU.

O

Offloading
Splitting a model between fast GPU memory and slower system RAM when it does not fully fit, trading speed for size.
On-device AI
Running models directly on a phone or laptop, so data stays local and nothing depends on a connection.
Open source
Software whose source code is free to inspect, modify and share; for models the honest term is often open weights instead.
Open weights
A model whose trained parameters are released publicly so anyone can download, run, and adapt it, though the training data may not be shared.
OpenAI-compatible API
The de facto standard interface for chat models; tools that speak it can swap between providers and local servers freely.
Overfitting
When a model memorises its training data too closely and performs well there but poorly on new, unseen examples.

P

Parameters
The internal numbers a model adjusts during training. More parameters can mean more capacity, and counts like 70B refer to these.
Phoning home
When software contacts its maker's servers without the user asking; caught by watching its network traffic.
Post-training
Everything done to a model after pre-training, from instruction tuning to preference training, turning raw prediction into a product.
Pre-training
The first, broad training stage where a model learns general patterns from a very large dataset before any task-specific tuning.
Prompt
The instruction or question you give an AI model to tell it what you want.
Prompt caching
Reusing the processed form of a repeated prompt prefix so long system prompts stop costing full price on every call.
Prompt engineering
The craft of wording and structuring prompts to get more useful, accurate, or reliable results from a model.
Prompt injection
An attack that hides instructions inside content a model reads, tricking it into following the attacker instead of the user.
Prompt steganography
Hiding a signal inside normal-looking prompt text, as with invisible Unicode characters that flag a session without the user seeing.

Q

Quantization
Shrinking a model by storing its numbers at lower precision, so it uses less memory and runs on smaller hardware, with a small quality trade-off.

R

RAG (Retrieval-augmented generation)
A technique where a model looks up relevant documents and uses them to answer, so it can cite sources and stay current.
Reasoning model
A model trained to spend extra steps thinking through a problem before answering, which improves results on maths, coding, and logic.
Reasoning tokens
The hidden thinking a reasoning model produces before its visible answer; they cost time and budget like normal output.
Red teaming
Deliberately probing a model to find weaknesses, unsafe outputs, or ways it can be misused, so they can be fixed before release.
Reinforcement learning
Training a model through trial and error, rewarding good outcomes and penalising bad ones, so it learns a useful strategy.
RLHF (Reinforcement learning from human feedback)
A training method that uses human ratings of model answers to teach it to be more helpful, honest, and safe.

S

Safetensors
A file format for model weights that cannot hide executable code, made to be safe to download and load.
Sampling
How a model picks each next token from its probability list; settings like temperature make the choice safer or more varied.
Sandbox
An isolated environment where an agent's code and commands run with limited access, so mistakes cannot touch the wider system.
Scaling laws
The observed pattern that model performance improves predictably as you increase data, parameters, and compute together.
Seed
The starting number for a generator's randomness; fixing it makes the same prompt reproduce the same output.
Self-hosting
Running software, including AI models, on your own machines instead of renting the hosted version.
Self-supervised learning
Training where the data labels itself, as when a model learns by predicting the next word of ordinary text.
Semantic search
Search that matches meaning rather than exact words, so a query about cars finds documents about automobiles.
Small language model (SLM)
A compact model, roughly under 10 billion parameters, built to run fast on phones and laptops.
Speculative decoding
A speed trick where a small model drafts tokens and the big model checks them in bulk, with identical output.
Speech recognition (ASR)
Turning spoken audio into text; local models like Whisper made it free and private.
State of the art (SOTA)
The best published result on a benchmark at a given time; a moving target in AI.
State-space model (SSM)
A neural architecture that handles long sequences with far less memory than attention, used alongside it in hybrid models.
Streaming
Sending a model's answer word by word as it is generated instead of waiting for the whole reply.
Superintelligence
A hypothetical AI that far surpasses the best human minds across essentially all fields.
Supervised learning
Training on examples paired with correct answers, so the model learns to map one to the other.
Synthetic data
Data generated by a model or simulation, rather than collected from the real world, used to train or test other models.
System prompt
The hidden instructions an app gives a model before the conversation starts, setting its role, rules and tone.

T

Telemetry
Data software sends home about how it is used; some AI tools ship it switched on by default with no visible off switch.
Temperature
A setting that controls how random a model's output is. Low values make it focused and predictable, high values make it more varied.
Test-time compute
Extra thinking a model spends at answer time instead of training time; the idea behind reasoning models.
Text-to-speech (TTS)
Generating spoken audio from text; modern local systems produce natural voices on ordinary hardware.
Time to first token
How long a model takes to start answering; long prompts stretch it because the model must read everything first.
Token
The small chunk of text, often a word piece, that a model reads and generates. Models measure input and cost in tokens.
Tokenization
The step of splitting text into tokens so a model can process it.
Tokens per second (tok/s)
The standard speed measure for text generation; around 10 to 15 reads like natural typing pace.
Tool calling
A model's ability to trigger outside functions, like a search or a script, by emitting a structured request instead of plain text.
Top-k sampling
A sampling setting that limits the model's next-token choice to the k most likely options.
Top-p (nucleus sampling)
A sampling setting that limits choices to the smallest set of tokens whose probabilities add up to p, trimming unlikely words.
Training
The process of feeding data to a model so it adjusts its parameters and learns to perform a task.
Transfer learning
Reusing what a model learned on one task as the starting point for another, instead of training from zero.
Transformer
The neural network architecture behind most modern AI, which uses attention to handle sequences like text efficiently.
Turing test
Alan Turing's 1950 test of whether a machine's conversation can pass as human.

U

Unified memory
A design where CPU and GPU share one pool of RAM, letting Macs load models that would need a big graphics card elsewhere.
Unsupervised learning
Training on raw data with no labels, leaving the model to find structure on its own.

V

Vector database
A database built to store embeddings and quickly find the items most similar in meaning to a query. Often used with RAG.
Vibe coding
Building software by describing what you want to an AI and accepting its code with little review, prized for speed and blamed for bugs.
Vision-language model (VLM)
A model that reads images and text together, so it can describe screenshots or answer questions about a photo.
VRAM (Video RAM)
The memory on a graphics card; on most PCs it caps the size of model you can run.

W

Watermarking
Embedding a hidden, detectable pattern in AI output so machines can later tell it was generated.
Weights
Another word for a model's learned parameters, the values that determine how it turns an input into an output.
World model
An AI's internal picture of how an environment behaves, letting it predict outcomes before acting.

Z

Zero-shot learning
Asking a model to do a task with no examples at all, relying on what it absorbed in training.