How AI works: glossary
Key terms from this course, in plain English.
- Attention
- The mechanism in transformers that lets the model weigh how relevant each token in the context is when processing another.
- Base model
- A model after pre-training only — knowledgeable, but not yet tuned to follow instructions.
- Constitutional AI
- Anthropic's approach to training models with a written set of principles and AI-generated feedback.
- Context window
- The maximum amount of text (in tokens) a model can consider at once, including instructions, conversation, documents and its own reply.
- Deep learning
- Machine learning with neural networks of many layers.
- Diffusion model
- A generative model that creates images or video by learning to remove noise step by step.
- Embedding
- A list of numbers representing a token, word or document's meaning, so similar meanings sit close together.
- Feature
- An input variable describing an example, such as a house's floor area.
- Gradient descent
- The training method that adjusts parameters step by step in the direction that reduces error.
- Hallucination
- Plausible but false or unsupported model output.
- Inference
- Using a trained model to produce outputs for new inputs.
- Interpretability
- Research into understanding what happens inside AI models.
- Knowledge cut-off
- The date after which a model's training data contains no information.
- Label
- The correct answer attached to a training example in supervised learning.
- Large language model (LLM)
- A transformer-based model trained on vast text to predict the next token, then post-trained to be a helpful assistant.
- Loss
- A score measuring how wrong a model's predictions are during training.
- Overfitting
- When a model memorises training data and fails to generalise to new data.
- Parameter
- A learned weight inside a model. Modern LLMs have billions.
- Post-training
- Training after pre-training — instruction tuning, RLHF, safety training — that shapes an assistant's behaviour.
- Pre-training
- The first, compute-intensive stage of training an LLM on huge amounts of text to predict the next token.
- Precision
- Of the cases a model flagged as positive, the share that truly were.
- Reasoning model
- An LLM trained to work through problems step by step before answering, using more computation on harder tasks.
- Recall
- Of all truly positive cases, the share a model caught.
- Reinforcement learning
- Learning by trial and error from rewards and penalties.
- Retrieval-augmented generation (RAG)
- Retrieving relevant documents and adding them to the model's context so answers are grounded in sources.
- RLHF
- Reinforcement learning from human feedback: training a model towards responses people prefer.
- Self-supervised learning
- Learning where labels come from the data itself — such as the next word in a sentence.
- Supervised learning
- Learning from labelled examples.
- Sycophancy
- A model's tendency to agree with or flatter the user, even when the user is wrong.
- Temperature
- A setting controlling how random a model's token choices are.
- Token
- A chunk of text — a word, part of a word or punctuation — that a language model processes.
- Transformer
- The neural network architecture, introduced in 2017, behind today's LLMs.
- Unsupervised learning
- Finding structure in unlabelled data, such as clusters.