Journal of the AI Universe
AI glossary
Short, plain-English explanations of the AI terms that appear in the daily issues.
Terms
- 6G
- The next generation of mobile networks after 5G, still in research and standard-setting, expected to use AI widely in the network itself.
- AGI (artificial general intelligence)
- A hypothetical AI system that could learn and perform almost any intellectual task a person can, rather than excelling only at narrow jobs. No agreed test for it exists, and claims about when it might arrive vary widely.
- AI agent / agentic AI
- Software that uses an AI model to plan and carry out multi-step tasks on its own, such as browsing, calling tools or editing files, with limited human input at each step.
- AI data centre (MW and GW)
- A building full of servers for training and running AI. Size is often given in megawatts (MW) or gigawatts (GW) of power; 1 GW is 1,000 MW.
- AI-RAN
- Using AI inside the radio access network (RAN), the part of a mobile network made of masts and base stations, to manage traffic and capacity, often on shared computing hardware.
- Alignment
- Research and practice aimed at making AI systems pursue the goals their developers and users intend, and behave safely.
- ASI (artificial superintelligence)
- A hypothetical AI that would far exceed the best human minds across most fields. It is discussed mainly in safety, policy and long-range forecasting debates.
- BCI (brain-computer interface)
- A system that reads brain signals and turns them into commands for a computer or device, for example to help people with paralysis communicate.
- Benchmark
- A standard test set used to compare AI systems. High scores do not always mean good real-world performance.
- CBT (cognitive behavioural therapy)
- A structured talking therapy that helps people change unhelpful patterns of thinking and behaviour; increasingly delivered online, sometimes with AI support.
- Chain of thought
- The step-by-step intermediate reasoning a model writes out before giving its final answer.
- CUDA
- NVIDIA’s software platform for programming its GPUs, widely used to build and run AI workloads.
- Deepfake
- Fake audio, images or video made with AI to look or sound like a real person.
- Diffusion model
- A generative model that creates images, video or audio by starting from random noise and removing it step by step until a clear result appears.
- Digital twin
- A live virtual model of a real object, process or place, such as a factory, that is updated with real data to test changes safely.
- Distillation
- Training a smaller, cheaper ‘student’ model to imitate a larger ‘teacher’ model, keeping much of its ability at lower cost.
- Embodied AI
- AI that controls a physical body, such as a robot or vehicle, and learns from interacting with the real world.
- EU AI Act
- The European Union’s law regulating AI by level of risk, with stricter duties for high-risk uses and some uses banned.
- Fine-tuning
- Further training of an existing model on a smaller, focused dataset so it performs better on a particular task or style.
- Foundation model
- A large general-purpose model trained on broad data that can be adapted to many specific tasks, rather than being built for one job.
- GPU (graphics processing unit)
- A chip with thousands of small cores that work in parallel; originally for graphics, now the main workhorse for training and running AI.
- Guardrails
- Rules, filters or checks placed around an AI system to block harmful, unsafe or off-topic outputs.
- Hallucination
- When an AI system states something false or made-up with apparent confidence.
- HBM (high-bandwidth memory)
- Stacked memory chips placed right next to an AI processor to feed it data very quickly; often a bottleneck in AI chip supply.
- Humanoid robot
- A robot with a broadly human body shape, designed to work in spaces and with tools made for people.
- Inference
- Running a trained model to produce answers or predictions, as opposed to training it. Most day-to-day AI computing cost is inference.
- LLM (large language model)
- An AI model trained on very large amounts of text to predict the next word (token). Chatbots and writing assistants are built on LLMs; they can be fluent yet still make factual errors.
- MCP (Model Context Protocol)
- An open standard that lets AI assistants connect to outside tools and data sources in a consistent way.
- Multimodal AI
- AI that works with more than one kind of data at once, such as text, images, audio, video or sensor signals.
- Open-weight model
- A model whose trained parameters (weights) are published so anyone can download and run it; the training data and code may still be private.
- PCCP (predetermined change control plan)
- A plan agreed with the US FDA in advance that sets out how an AI medical device may be updated after approval without a new submission for each change.
- Photonic computing
- Using light instead of electrical signals to move or process data, aimed at faster and more energy-efficient AI hardware.
- Qubit
- The basic unit of a quantum computer. Unlike an ordinary bit it can hold a blend of 0 and 1, but it is fragile and needs error correction.
- RAG (retrieval-augmented generation)
- A technique in which an AI model first looks up relevant documents and then writes its answer from them, so replies can cite current or private information instead of relying only on what it memorised in training.
- Reasoning model
- A language model designed to work through a problem in intermediate steps before answering, usually spending more computing time per question.
- Red teaming
- Deliberately attacking or probing an AI system to find failures and risks before others do.
- Reinforcement learning
- A way of training AI by trial and error, rewarding actions that lead to good outcomes; used for game-playing, robotics and to refine language models.
- Synthetic data
- Artificially generated data used to train or test models when real data is scarce, costly or sensitive.
- Token
- The small unit of text (a word or part of a word) that language models read and write; usage and prices are often counted in tokens.
- Transformer
- The neural-network design behind most modern language and image models; it weighs how every part of the input relates to every other part.
- Vision-language model
- A model that understands images and text together, for example describing a photo or answering questions about a chart.
- VLA (vision-language-action model)
- A model for robots that takes in camera images and written instructions and outputs physical actions, linking seeing, understanding language and moving in one system.
- World model
- An AI model that learns how an environment behaves, so it can predict what happens next and plan, for example in robotics, games or video generation.
Evidence labels used in the journal
- Commercial deployment (evidence label)
- A product or system in real use by customers, patients or the public.
- Company announcement (evidence label)
- A company’s own claim, plan, deal or press release that has not been independently verified.
- Independent validation (evidence label)
- Results checked by someone other than the developer, such as an independent benchmark, regulator test, audit or replication.
- Laboratory demonstration (evidence label)
- A prototype or result shown in a lab or controlled setting by its developer, not yet deployed or published.
- Peer-reviewed (evidence label)
- Published in a journal or conference after review by independent experts.
- Preprint (evidence label)
- A research paper shared publicly, for example on arXiv or medRxiv, before peer review. Findings may change.
Definitions are written for this journal in simple terms; they are not legal or technical standards. Terms are listed when they occur in the issues; the first use of some terms on an issue page links here.