The AI Glossary

An AI system that can autonomously perceive its environment, make decisions, and take actions to achieve a specified goal — often by using tools, browsing the web, or writing and executing code.

The challenge of ensuring that an AI system's goals, values, and behaviour match what its designers and users actually intend — especially as AI becomes more capable.

A set of rules and protocols that allows one software application to communicate with another. AI companies expose APIs so developers can integrate their models into their own products.

A component of neural networks that allows a model to focus on the most relevant parts of an input when generating an output — the core innovation behind the Transformer architecture.

AI systems capable of operating independently to complete complex, multi-step tasks without continuous human oversight or instruction.

A standardised test or dataset used to measure and compare the performance of AI models across specific tasks, such as reasoning, coding, or language understanding.

Systematic errors in AI outputs caused by flawed assumptions, unrepresentative training data, or design choices that result in unfair or skewed results for certain groups.

A prompting technique that encourages a language model to reason step-by-step before giving a final answer, significantly improving performance on complex reasoning tasks.

A software application designed to simulate conversation with humans, typically via text or voice. Modern AI chatbots use large language models to generate contextually relevant responses.

A family of large language models developed by Anthropic, designed with a focus on safety, helpfulness, and harmlessness. Claude is a leading competitor to OpenAI's GPT models.

A technique developed by Anthropic where an AI model is trained to critique and revise its own outputs based on a set of principles (a 'constitution'), reducing harmful outputs without relying solely on human feedback.

The maximum amount of text (measured in tokens) that a language model can process in a single interaction — including both the input prompt and the generated output.

An AI assistant designed to work alongside a human in a specific workflow — most commonly in coding (GitHub Copilot) or productivity software (Microsoft Copilot). It suggests, completes, or generates content in context.

A subset of machine learning that uses neural networks with many layers (hence 'deep') to learn representations of data. It is the foundation of most modern AI breakthroughs in image recognition, language, and speech.

A class of generative AI models that learn to create images (or other data) by learning to reverse a process of adding noise. Used by Stable Diffusion, DALL-E, and Midjourney.

A numerical representation of text, images, or other data as a vector (a list of numbers) that captures semantic meaning, allowing AI models to compare and reason about concepts mathematically.

Capabilities that appear in large AI models that were not explicitly trained for and were not present in smaller versions of the same model — often surprising even the researchers who built them.

A prompting approach where you provide a language model with a small number of examples (typically 2–10) of the desired input-output format before asking it to complete a new task.

The process of further training a pre-trained AI model on a smaller, task-specific dataset to adapt it for a particular use case or domain.

A large AI model trained on broad, diverse data that can be adapted for a wide range of downstream tasks. GPT-4, Claude, and Gemini are all foundation models.

Google DeepMind's family of multimodal large language models, designed to process and generate text, images, audio, and video. Gemini Ultra is Google's most capable model.

AI systems that can create new content — including text, images, audio, video, and code — rather than simply analysing or classifying existing data.

A family of large language models developed by OpenAI, trained using a combination of unsupervised pre-training on large text corpora and reinforcement learning from human feedback. GPT-4 is the current flagship model.

A specialised processor originally designed for rendering graphics, now widely used for training and running AI models due to its ability to perform many parallel calculations simultaneously.

Safety mechanisms built into AI systems to prevent them from generating harmful, offensive, or dangerous content — including content filters, output classifiers, and constitutional constraints.

When an AI language model generates information that sounds plausible and confident but is factually incorrect or entirely fabricated — a significant reliability challenge for current LLMs.

A system design approach where human oversight or approval is required at key decision points in an AI workflow, rather than allowing the AI to operate fully autonomously.

The ability of a large language model to learn from examples provided within the prompt itself, without any weight updates or fine-tuning.

The process of running a trained AI model to generate outputs from new inputs — as opposed to training, which is the process of teaching the model. When you use ChatGPT, you are performing inference.

A prompt or technique designed to bypass an AI model's safety guardrails and get it to produce content it was trained to refuse.

A type of AI model trained on vast amounts of text data to understand and generate human language. LLMs are the technology behind ChatGPT, Claude, Gemini, and most modern AI writing tools.

The time delay between sending a request to an AI model and receiving the first token of its response. Low latency is critical for real-time applications.

Meta's family of open-source large language models. Unlike GPT or Claude, Llama models can be downloaded and run locally on your own hardware, making them popular for privacy-conscious users and developers.

A branch of AI in which systems learn from data to improve their performance on tasks without being explicitly programmed for each specific scenario.

An AI image generation tool that creates photorealistic and artistic images from text prompts. Known for its high aesthetic quality and active Discord community.

In AI, a model is a mathematical system trained on data to make predictions or generate outputs. It encodes learned patterns as numerical parameters (weights).

AI systems that can process and generate multiple types of data — such as text, images, audio, and video — within a single model.

The field of AI focused on enabling computers to understand, interpret, and generate human language in a useful way.

A computational model loosely inspired by the human brain, consisting of layers of interconnected nodes (neurons) that process and transform data to learn patterns.

AI models whose weights, training code, or both are publicly released, allowing anyone to download, run, study, and modify them.

The AI research company behind ChatGPT, GPT-4, DALL-E, and Whisper. Founded in 2015, OpenAI is one of the most influential organisations in the AI industry.

When an AI model learns the training data too well — including its noise and quirks — and performs poorly on new, unseen data as a result.

The numerical values (weights and biases) inside a neural network that are adjusted during training. The number of parameters is a rough indicator of a model's capacity — GPT-4 is estimated to have over 1 trillion.

An AI-powered search engine that provides direct answers to questions with cited sources, combining the capabilities of a search engine with a language model.

The input text or instruction given to an AI model to elicit a desired response. The quality and structure of a prompt significantly affects the quality of the output.

The practice of designing, refining, and optimising prompts to reliably get high-quality, accurate, and useful outputs from AI language models.

A technique that combines a language model with a retrieval system, allowing the AI to search an external knowledge base (like your documents or the web) before generating a response — reducing hallucinations and improving accuracy.

The ability of an AI model to work through complex, multi-step problems logically — a capability that has improved dramatically with models like OpenAI's o1 and o3.

A safety testing process where researchers deliberately try to find ways to make an AI model produce harmful, biased, or dangerous outputs — to identify and fix vulnerabilities before public release.

A training technique where human raters evaluate AI outputs and their preferences are used to train a reward model, which then guides the AI to produce more helpful, harmless, and honest responses.

An open-source image generation model developed by Stability AI that can be run locally on consumer hardware. It is the basis for many third-party image generation tools and custom models.

A hidden instruction given to an AI model before the conversation begins, used to set its persona, tone, rules, and constraints. Users typically cannot see the system prompt.

A parameter that controls the randomness of an AI model's outputs. A low temperature (near 0) produces more deterministic, predictable responses; a high temperature produces more varied and creative ones.

AI systems that generate images from natural language descriptions (prompts). Examples include Midjourney, DALL-E, and Stable Diffusion.

The basic unit of text that language models process. A token is roughly 3–4 characters or about 0.75 words in English. Models have limits on how many tokens they can process at once (the context window).

The dataset used to train an AI model. For large language models, this typically includes billions of web pages, books, and other text sources. The quality and composition of training data heavily influences model behaviour.

A technique where a model trained on one task is adapted for a different but related task, leveraging the knowledge it already acquired to learn more efficiently.

The neural network architecture introduced in the 2017 paper 'Attention Is All You Need' that became the foundation for virtually all modern large language models, including GPT, Claude, and Gemini.

A database optimised for storing and searching vector embeddings, enabling fast similarity search — a core component of RAG systems and semantic search applications.

AI systems that can analyse, interpret, and generate visual content — including image classification, object detection, OCR, and image generation.

OpenAI's open-source speech recognition model that can transcribe and translate audio in multiple languages with high accuracy.

The ability of an AI model to perform a task it was not explicitly trained for and has received no examples of, relying solely on its general knowledge and the task description.