AI & LLM Glossary

The broad field of building computer systems that can perform tasks usually requiring human intelligence.

A type of AI trained on vast amounts of text that can read, write, summarise, and reason with language.

AI that creates new content — text, images, code, audio, or video — rather than just analysing existing data.

The branch of AI that helps computers understand and generate human language.

A method where computers learn patterns from data rather than being explicitly programmed.

A subset of machine learning that uses layered networks loosely inspired by the human brain.

The underlying computational structure that most modern AI is built on — layers of interconnected mathematical nodes.

A large AI model trained on broad data that can be adapted for many specific tasks.

An AI model that has already been trained on large datasets and is ready to use or customise.

The text, documents, or other content an AI learns from during its development.

The process of a trained AI model actually running and producing outputs — as opposed to being trained.

The finished AI system after training — the 'brain' that processes inputs and produces outputs.

The internal numerical values a model adjusts during training — more parameters generally means more capable.

Another term for parameters — the stored knowledge inside a trained AI model.

The full collection of text used to train an AI model.

The basic unit an LLM reads and writes — roughly equivalent to a word or part of a word.

The information an AI can 'see' at one time — its working memory for a task.

The instruction or question you give to an AI model to produce a response.

What the AI produces in reply to your prompt.

The text an LLM generates to finish or respond to a prompt.

A connection that lets software talk to an AI model programmatically.

AI that can work with multiple types of input — text, images, audio, video — not just text alone.

An AI's ability to work through problems step by step, drawing logical conclusions.

A standardised test used to compare and measure AI model performance.

Capabilities that appear in large models without being explicitly trained — they emerge at scale.

The core architecture that powers most modern LLMs, introduced by Google in 2017.

The part of a transformer that lets the model focus on the most relevant words regardless of where they appear in a sentence.

When a model weighs how much each word in a sentence relates to every other word.

The part of a model that reads and understands input.

The part of a model that generates output.

A model with both encoder and decoder components — suited for translation and summarisation.

A model that generates text one token at a time, each based on what came before.

A numerical representation of a word or concept that captures its meaning and relationships.

A list of numbers used to represent a piece of text mathematically — the language of AI under the hood.

The internal mathematical space where a model stores its 'understanding' of concepts.

A processing stage inside a neural network — deeper models have more layers.

The complete set of tokens a model knows and can use.

The process of breaking text into tokens before the model processes it.

A common tokenisation method that splits text into frequently occurring character sequences.

How a model tracks word order, since transformers don't naturally process text sequentially.

The mathematical function that converts a model's raw scores into probabilities.

The raw numerical scores a model produces before they're converted to probabilities.

The initial, large-scale training phase where a model learns from massive datasets.

Further training a pre-trained model on specific data to specialise it for a task.

Training where the model learns from labelled examples — input paired with the correct answer.

Training where the model finds patterns in data without being given labels.

Training where a model receives rewards or penalties based on the quality of its outputs.

Teaching an AI to improve its responses using human ratings to align it with human preferences.

An AI trained to score outputs, used to guide RLHF training.

Fine-tuning a model specifically to follow instructions better.

An efficient fine-tuning technique that modifies only a small part of a model's weights.

A memory-efficient version of LoRA that uses quantisation to reduce computing requirements.

The algorithm models use to gradually improve during training by reducing errors.

The measure of how wrong a model's outputs are — training aims to minimise this.

The process of passing error signals backwards through a network to update weights.

How quickly a model updates its weights during training — a critical training hyperparameter.

When a model memorises training data rather than learning general patterns.

When a model hasn't learned enough to perform well on its task.

AI-generated training data used when real data is scarce, sensitive, or expensive to label.

Artificially expanding training data by creating variations of existing examples.

When a model loses previously learned knowledge after being fine-tuned on new data.

Techniques to let models learn new information without forgetting old knowledge.

The practice of crafting inputs to get better outputs from AI models.

Instructions given to an AI at the start of a session that shape its behaviour throughout.

Asking an AI to perform a task with no examples provided.

Giving an AI a small number of examples before asking it to perform a similar task.

Giving a single example before asking the AI to perform a task.

Asking an AI to show its reasoning step by step before giving a final answer.

The maximum amount of text an AI can process in a single interaction.

A setting that controls how creative or random an AI's outputs are.

A setting that controls output diversity by limiting which tokens the model can choose from.

A setting that restricts the AI to choosing from only the K most likely next tokens.

A setting that discourages the AI from repeating the same phrases.

The upper limit you set on how long the AI's response can be.

A type of attack where malicious instructions hidden in content hijack an AI's behaviour.

Attempts to bypass an AI's safety guidelines through clever prompting.

When a model adapts its behaviour based solely on examples provided in the current prompt.

Linking multiple prompts together so the output of one feeds into the next.

Instructing the AI to produce output in a specific format like JSON, a table, or a list.

The challenge of ensuring AI systems pursue goals that match human values and intentions.

The field focused on preventing AI from causing harm — intentional or unintentional.

When an AI confidently produces false information it has invented.

Another term for hallucination — the AI fills gaps in knowledge with plausible-sounding fiction.

Systematic unfairness in AI outputs, often reflecting imbalances in training data.

The goal of ensuring AI treats all people and groups equitably.

The ability to understand and explain how an AI reaches its outputs.

Making AI decisions understandable to non-technical users.

The ability to examine what is happening inside a model to understand its reasoning.

Deliberately trying to find flaws or harmful behaviours in an AI before deployment.

Anthropic's approach to training AI using a set of principles rather than only human feedback.

Rules or filters built into an AI system to prevent harmful or inappropriate outputs.

Automated or human review of AI outputs to catch problematic content.

A framework for developing and deploying AI in ways that are ethical and accountable.

Policies, frameworks, and regulations that guide how AI is developed and used.

A document that describes a model's capabilities, limitations, and intended uses.

Embedding invisible signals in AI outputs to identify them as machine-generated.

AI-generated synthetic media that depicts people saying or doing things they didn't.

Any content created or significantly altered by AI.

Combining an LLM with a search system so it can look up current or specific information before responding.

A database that stores embeddings and allows searching by meaning rather than keywords.

Searching by meaning and intent rather than exact keyword matching.

The date after which an AI has no training data — it won't know about events after this point.

Connecting AI outputs to verified real-world information to reduce hallucination.

A structured collection of documents or data that an AI can search through.

Using an LLM to answer questions based on specific uploaded documents.

Splitting documents into smaller pieces for more effective retrieval and processing.

Organising content so it can be searched efficiently.

Automatically identifying people, places, organisations, and dates in text.

Combining keyword search and semantic search for more accurate retrieval.

An AI system that can take actions, use tools, and complete multi-step tasks autonomously.

A process where an AI plans and executes a series of steps to complete a complex task.

The ability of an AI to use external tools — calculators, search engines, databases, APIs.

A structured way for AI to trigger specific actions in connected software.

Keeping a human approval step in an AI workflow for oversight and safety.

Multiple AI agents working in parallel or sequence, each handling part of a task.

Coordinating multiple AI models or agents to work together on a task.

Breaking a complex task into smaller sub-tasks an AI agent can tackle one at a time.

A sequence of automated steps that processes information and takes action.

An AI tool that can write and execute code to solve problems.

The ability of an AI agent to control a computer — clicking, typing, and navigating software.

A controlled, isolated environment for testing AI agents safely before deployment.

How often an AI produces correct outputs on a given task.

The proportion of the AI's positive predictions that are actually correct.

The proportion of actual positives that the model correctly identifies.

A combined measure of precision and recall — the harmonic mean of both.

The time between sending a prompt and receiving a response.

How many requests an AI system can handle per unit of time.

Whether an AI's statements are true and verifiable.

Using human raters to judge AI output quality — the gold standard for open-ended tasks.

When a model's performance degrades over time as the real world changes.

The specialised chip used to train and run AI models at speed.

Google's custom chip designed specifically for AI workloads.

Reducing a model's numerical precision to make it smaller and faster, with minimal quality loss.

Techniques to make large models smaller so they run on less powerful hardware.

Training a smaller model to mimic a larger, more capable one.

Running AI models directly on local devices rather than in the cloud.

AI models and services accessed over the internet via a provider's infrastructure.

Running AI models on hardware you own and control, within your own environment.

The infrastructure that makes a trained model available to receive and respond to requests.

The practice of managing AI models in production — deployment, monitoring, updating.

Packaging AI software and its dependencies together for reliable, portable deployment.

The ability of an AI system to handle increasing numbers of users or requests.

An AI assistant embedded in software to help users with tasks in real time.

AI condensing long documents into concise summaries.

AI determining the emotional tone of text — positive, negative, or neutral.

AI assigning text or content to predefined categories.

AI identifying the underlying goal behind a user's message.

A conversational AI interface designed to interact with users in natural language.

A more capable AI companion that handles complex tasks, scheduling, research, and communication.

Connecting AI capabilities into existing software, processes, and workflows.