AI glossary

42 terms, explained without jargon.

Agent: An AI system that can plan and take actions - such as browsing, running code or calling tools - to complete a goal, rather than only answering a single question.
AGI (artificial general intelligence): A hypothetical AI that can learn and perform most intellectual tasks as well as a human. There is no agreed test or date, and experts disagree on how close current systems are.
Alignment: Making an AI system pursue the goals and values its designers and users actually intend, including avoiding harmful or deceptive behaviour.
API: Application programming interface: a way for software to talk to an AI model over the internet, usually paid per amount of text processed.
Attention: The mechanism in transformer models that lets each part of the input weigh how relevant every other part is. It is why modern models handle context so well.
Benchmark: A standard test used to compare models, such as exams or coding challenges. Scores can be gamed or leaked into training data, so treat them as hints, not proof.
Bias: Systematic unfairness in model outputs, usually inherited from imperfect training data or design choices.
Chain of thought: Having a model write out intermediate reasoning steps before the final answer, which often improves results on maths and logic.
Chatbot: A program that converses in natural language. Modern chatbots are usually built on large language models.
Context window: The amount of text (measured in tokens) a model can consider at once - your prompt, earlier conversation and its own reply together.
Deepfake: Synthetic image, audio or video that convincingly imitates a real person, made with generative AI.
Diffusion model: A kind of generative model that creates images (or audio/video) by gradually removing noise from random static, guided by a text prompt.
Embedding: A list of numbers that represents the meaning of text, an image or other data, so similar things end up close together. Used for search and recommendations.
Fine-tuning: Training an existing model a bit more on specialised data so it performs better at a particular task or style.
Foundation model: A large model trained on broad data that can be adapted to many tasks.
Generative AI: AI that creates new content - text, images, audio, video or code - rather than only classifying or predicting.
GPU: Graphics processing unit: the type of chip whose parallel processing makes training and running large models practical.
Guardrails: Rules, filters and training that keep an AI system within safe and acceptable behaviour.
Hallucination: When a model states something false or invented with confidence, such as a fake citation. Always verify important facts.
Inference: Running a trained model to produce an output (as opposed to training it). Every time you send a prompt you are doing inference.
Jailbreak: A prompt designed to trick a model into ignoring its safety rules.
Large language model (LLM): A neural network trained on huge amounts of text to predict the next token, which gives it the ability to write, summarise, translate and reason in language.
LoRA: Low-rank adaptation: a cheap way to fine-tune a model by training small add-on weights instead of the whole network.
Machine learning: Teaching computers to find patterns in data and improve at a task without being explicitly programmed with rules.
Multimodal: Able to work with more than one type of data, such as text, images, audio and video.
Neural network: A computing system made of layers of simple connected units whose connection strengths (weights) are learned from data.
Open-weight model: A model whose trained weights are published so anyone can run or adapt them. Not always the same as fully open source, because training data and code may be withheld.
Overfitting: When a model memorises its training examples and performs poorly on new data.
Parameter: A learned number inside a model. Larger models have billions or trillions of parameters, but bigger is not always better.
Prompt: The text (and sometimes images) you give a model to tell it what to do.
Prompt injection: An attack where hidden instructions in a web page, document or email trick an AI assistant into doing something its user did not ask for.
RAG (retrieval-augmented generation): Letting a model look up relevant documents first and use them to answer, which reduces made-up facts and lets it use fresh or private information.
Reinforcement learning: Training by trial and reward: the system tries actions and learns which lead to better outcomes.
RLHF: Reinforcement learning from human feedback: people rank model answers and the model is trained to prefer the better ones. A key step in making chatbots helpful.
Red teaming: Deliberately attacking or stress-testing an AI system to find safety and security problems before real attackers do.
Synthetic data: Artificially generated data used to train or test models, often produced by other models.
System prompt: Hidden instructions from the developer that set a chatbot's role, tone and limits before your message arrives.
Temperature: A setting controlling randomness: low values give predictable answers, high values give more varied and creative ones.
Token: A chunk of text (roughly three-quarters of an English word on average) that models read and write. Pricing and limits are counted in tokens.
Training data: The examples a model learns from. Its quality, diversity and legality shape what the model can do and how it fails.
Transformer: The neural network design, introduced in 2017, behind nearly all modern language models and many image and audio models.
Vector database: A database built to store embeddings and quickly find the most similar items.