New AI terms turn up every week, and it is hard to tell which ones matter. We keep this glossary as a reference to the terms our clients ask us about most, grouped so you can scan them in minutes.
The core hierarchy
Start with how the broad terms relate: machine learning is one branch of AI, and deep learning is one kind of machine learning, built on neural networks.
- Artificial intelligence (AI): Systems that perform tasks that would otherwise need human thinking: reasoning, learning, perception and decision-making.
- Machine learning (ML): A branch of AI in which systems learn patterns from data instead of following rules a programmer wrote.
- Deep learning (DL): Machine learning with many-layered neural networks trained on large datasets, each layer learning a more abstract representation than the one before.
- Neural network: A computing model made of connected nodes (neurons), organised in layers, that processes information.
1. AI agents
- AI agent: A system that observes its environment, reasons about it and takes actions on its own toward a goal.
- Agentic AI: AI that can plan several steps ahead, use tools and carry out the work without a person directing each step.
- Tool use: A model's ability to call external functions, APIs or services to complete a task.
- Chain-of-thought (CoT): A prompting technique that asks the model to reason step by step before it answers.
- RAG (retrieval-augmented generation): Retrieving relevant material from external sources and giving it to the model, so its answer rests on that material.
- MCP (Model Context Protocol): An open standard for connecting AI models to external tools and data sources.
- Agent memory: Stored context that lets an agent remember past interactions from one session to the next.
- Agentic workflow: A workflow of several steps in which agents carry out and coordinate the steps themselves.
- Computer use: A model's ability to operate a computer the way a person does, through the screen, mouse and keyboard.
- Spec-driven development: Writing detailed specifications for coding agents to implement, in place of writing the code yourself.
- Human-in-the-loop (HITL): A governance design in which a person reviews and approves the AI's decisions at the steps that matter.
- LLM orchestrator: A framework, such as LangChain or Semantic Kernel, that chains models, tools and APIs into one pipeline.
- Prompt chaining: Feeding the output of one prompt into the next, to solve a problem in several steps.
2. Model architecture
- Transformer: The model architecture behind modern LLMs, built on self-attention (see attention mechanism).
- Large language model (LLM): A neural network trained on very large amounts of text to understand and generate language.
- Foundation model: A large model trained on broad data that can be adapted to many different tasks.
- Parameters: The weights inside a model that training adjusts.
- Hyperparameters: Settings chosen before training starts, such as the learning rate and the batch size.
- Training data: The data a model learns its patterns from.
- Epoch: One full pass through the training data.
- Batch size: The number of examples a model processes before its parameters are updated.
- Loss function: A measure of the gap between the model's output and the correct answer, which training works to reduce.
- Gradient descent: The method training uses to reduce the loss, adjusting the parameters a small step at a time.
- Backpropagation: The algorithm that works out how much each parameter contributed to an error, by passing the error backward through the network.
- Reasoning model: An LLM trained to work through a problem step by step before it answers.
- Small language model (SLM): A compact model built for specific tasks, small enough to run on a device.
3. Generative AI
- Generative AI: AI that creates new content: text, images, audio, video or code.
- Prompt: The text or instruction you give a generative model.
- Token: The unit a model reads and writes text in: a word, part of a word or a character.
- Context window: The maximum number of tokens a model can process in one request.
- Temperature: A setting that controls how random the output is. Higher gives more varied answers, lower gives more predictable ones.
- Inference: Running a trained model on new input to get an output.
- Diffusion model: A generative model that learns to reverse a process of adding noise, and creates new data, such as images, that way.
- Autoregressive model: A model that generates output one token at a time, each based on the tokens before it.
- Prompt engineering: Writing and testing prompts so a model gives the output you want, reliably.
4. Learning paradigms
- Supervised learning: Training on examples where each input is paired with the correct output.
- Unsupervised learning: Training on unlabelled data, so the model finds patterns and structure for itself.
- Reinforcement learning (RL): Learning by trial and error, with rewards for the actions that work.
- Self-supervised learning: Training in which the model makes its own labels from the data, for example by predicting a hidden word.
- Transfer learning: Reusing what a model learned on one task or domain to do better on another.
- Fine-tuning: Further training a pre-trained model on targeted examples so it does one task better.
- Zero-shot learning: Doing a task with no examples of it.
- Few-shot learning: Learning a task from a handful of examples, usually one to ten.
- LoRA (low-rank adaptation): A fine-tuning method that updates only a small number of parameters, so adapting a model is cheaper and faster.
- RLAIF (reinforcement learning from AI feedback): Using a second model to judge which answers are better, so training depends less on costly human labellers.
5. Language
- Natural language processing (NLP): The field of AI that deals with understanding and generating human language.
- Embedding: A list of numbers that represents the meaning of a piece of text, an image or other data, so that similar things sit close together.
- Semantic search: Search that matches on meaning and intent instead of exact keywords.
- Tokenisation: Splitting text into tokens before a model processes it.
- Attention mechanism: The part of a model that weighs how much each part of the input matters when it produces each part of the output.
- RLHF (reinforcement learning from human feedback): Training a model on people's judgements of which answers are better, to bring its behaviour in line with what people want.
6. Vision and multimodal
- Computer vision: AI that interprets and analyses images and video.
- Multimodal AI: Models that work across more than one type of data, such as text, images and audio.
- OCR (optical character recognition): Converting images of text, such as scans, into text a computer can read.
- Image segmentation: Dividing an image into distinct regions or objects.
- Object detection: Finding objects in an image and marking where they are.
7. Model behaviour and problems
- Hallucination: When a model produces content that sounds plausible but is wrong or made up.
- Overfitting: When a model memorises its training data and performs poorly on new inputs.
- Underfitting: When a model is too simple to capture the patterns in the data.
- Bias: Systematic errors or unfair results caused by flawed data or design.
- Alignment: Making AI systems behave as people intend, in line with human values.
- Jailbreaking: Prompts written to get a model past its safety limits.
- Prompt injection: Malicious input written to override a model's instructions.
- Drift: A gradual drop in a model's performance as real-world data moves away from the data it was trained on.
- Prompt leaking: When a user's questions get a model to reveal its system prompt or other confidential instructions.
- Automation bias: People's tendency to over-trust AI output, even when it is clearly wrong. It is a governance risk in its own right, separate from hallucination.
8. Evaluation and metrics
- Benchmark: A standard test used to compare models.
- Accuracy: The share of all predictions that are correct.
- Precision: Of the items a model flags as positive, the share that are positive.
- Recall: Of the items that are positive, the share the model flags.
- F1 score: One score that combines precision and recall (their harmonic mean).
- Perplexity: A measure of how well a language model predicts text. Lower is better.
- BLEU score: A score that compares generated text with reference translations.
- Ground truth: The checked, correct answer a model is trained or evaluated against.
9. Safety and ethics
- Explainable AI (XAI): Methods that let people see why an AI system made a decision.
- Red teaming: Attacking a system on purpose to find its weaknesses and the ways it fails.
- Constitutional AI: A training approach that guides a model's behaviour with a written set of principles.
- Guardrails: Checks that stop a model producing harmful or unwanted output.
- Data privacy: Protecting the personal information a model is trained on or processes.
- Model card: A document that sets out what a model can do, its limits and what it is meant for.
- Sandboxing: Running an AI system in isolation, so its actions cannot reach live systems without explicit permission.
- Copyright and IP infringement: The open legal debate over whether training models on publicly available data counts as "fair use".
- AI sovereignty: Nations or organisations hosting their own models and data locally, to depend less on a handful of global providers.
10. Infrastructure and deployment
- API (application programming interface): The interface software uses to send requests to an AI model and get results back.
- Latency: The time between sending a request and getting the model's response.
- Throughput: How many requests a system can handle in a given time.
- GPU (graphics processing unit): A chip that runs many calculations in parallel, used to train and run AI models.
- TPU (tensor processing unit): Google's own chip designed to speed up AI work.
- Quantisation: Storing a model's numbers at lower precision (for example 8-bit in place of 32-bit), so it is smaller and runs faster.
- Distillation: Training a smaller model to copy a larger one, keeping as much of its performance as possible.
- Edge AI: Running AI models on the device itself instead of on cloud servers.
- Vector database: A database built to store embeddings and find the ones closest in meaning to a query. RAG systems use one to find relevant material.
- Inference cost: What it costs in compute to run a model for each request, the ongoing expense of running AI in production.
- Vibe coding: Building software by describing what you want in plain language and letting an AI write the code.
- NPU (neural processing unit): A chip built into many recent laptops and phones to run AI models on the device.
- KV cache: A store of values a model has already computed (keys and values), so it does not compute them again for each new token. It makes text generation much faster.
- Compute governance: How an organisation allocates and budgets its limited GPU and cloud computing capacity.
11. Advanced concepts
- Latent space: The compressed internal space in which a model represents data, with similar concepts close together.
- Emergent abilities: Abilities that appear in large models and are absent from smaller versions.
- Scaling laws: The predictable way a model's performance changes as its size, data and compute grow.
- In-context learning: A model picking up a task from examples in the prompt, with no change to its weights.
- Synthetic data: Data generated by a model or program, used for training or to add to training data.
- Mixture of experts (MoE): An architecture in which separate subnetworks (experts) specialise in different inputs.
- AGI (artificial general intelligence): A hypothetical AI that matches human reasoning in every field.
- ASI (artificial superintelligence): A theoretical AI that exceeds human intelligence in every area.
- World model: A model's internal picture of how the world works, which it uses to predict and plan.
From terms to a working system
We add terms as they come into use. If you have a workflow in mind, we can build it and run it for you, starting with Discovery, a paid audit of how it runs today.
Start with Discovery and tell us about the workflow.