AI glossary
Plain-English definitions of 36 AI terms, from data and models to transformers, RLHF and alignment.
- AI
- Machines that learn from examples to notice, guess and help.
- Machine learning
- How computers find their own rules from lots of examples.
- Data
- Examples, like pictures, words or sounds, that AI learns from.
- Pattern
- Something that repeats. AI is very good at spotting them.
- Model
- What an AI keeps after learning: lots and lots of numbers.
- Training
- Practising on examples where the answers are known.
- Testing
- Checking the AI on new examples it has never seen.
- Neural network
- Layers of tiny deciders that pass signals on and vote.
- Deep learning
- Neural networks with many layers.
- Pixel
- A tiny square in a picture, stored as numbers.
- Token
- A small piece of text that chatbots read and write.
- LLM
- Large language model: an AI that predicts the next token.
- Prompt
- The words you give an AI to say what you want.
- Hallucination
- When AI says something that sounds true but isn’t.
- Bias
- When AI is unfair because its examples left things out.
- Embedding
- A list of numbers that captures what something means.
- RAG
- Looking things up in trusted notes before answering.
- Agent
- An AI that plans and takes steps using tools.
- API
- A way for apps to send requests to an AI and get replies.
- Deepfake
- A fake photo, voice or video made with AI.
- Vector
- An ordered list of numbers that represents something, like a word or an image.
- Gradient descent
- Training by repeatedly nudging a model’s numbers in the direction that reduces its error.
- Backpropagation
- The method that works out how much each weight contributed to an error, layer by layer.
- Transformer
- The neural network design behind modern language models, built on attention.
- Attention
- A way for a model to focus on the most relevant words when processing each word.
- Context window
- How much text a model can look at in one go.
- Temperature
- A setting that makes answers more predictable (low) or more creative (high).
- Fine-tuning
- Extra training that gives a pre-trained model a speciality.
- Overfitting
- Memorising training examples instead of learning the general pattern.
- Feature
- A clue a model uses to make a prediction, like size or colour.
- Decision tree
- A model that reaches an answer by asking a series of yes-or-no questions.
- Reinforcement learning
- Learning by trial and error, guided by rewards.
- RLHF
- Reinforcement learning from human feedback: tuning a model towards answers people prefer.
- Diffusion model
- An image generator that turns random noise into a picture step by step.
- Interpretability
- Research into what is happening inside a model.
- Alignment
- Making sure AI systems do what people really intend, safely.