AI glossary
49 words you’ll meet in AI news, in plain English. For the bigger ideas, read AI Explained.
How models work
- Artificial intelligence (AI) #
- Computer systems that do tasks that usually take human judgment — understanding language, recognizing images, making predictions. In today’s news it mostly means machine learning. Stories
- Machine learning #
- Software that learns patterns from examples instead of following hand-written rules. Stories
- Neural network #
- A machine-learning model built from layers of simple connected units. Training adjusts the strength of the connections until the network gets good at its task. Stories
- Large language model (LLM) #
- A neural network trained on huge amounts of text to predict what comes next. The chatbots and assistants in the news — ChatGPT, Claude, Gemini, Llama and others — are built on them. Stories
- Transformer #
- The neural-network design behind most of today’s language models, introduced by Google researchers in 2017. Its “attention” mechanism lets the model weigh how every word relates to every other word. Stories
- Parameters (weights) #
- The adjustable numbers inside a model, set during training. More parameters usually means more capacity, but not automatically better answers. Stories
- Token #
- The chunk of text a model reads and writes — often a short word or part of a word. Usage limits and API prices are counted in tokens. Stories
- Context window #
- How much a model can take in at once, measured in tokens: your question, any files you attach and the conversation so far. Past the limit, older material drops out. Stories
- Training and pretraining #
- Feeding a model large amounts of data so it adjusts its parameters. Pretraining is the big first pass on broad data; later stages teach it to follow instructions and behave safely. Stories
- Fine-tuning #
- Extra training on a smaller, focused dataset to adapt a model to a task, a style or a company’s own material. Stories
- Inference #
- Running a trained model to get an answer. It happens every time you send a prompt, and it is what most AI products charge for. Stories
- Multimodal #
- A model that works with more than one kind of input or output — text, images, audio or video. Stories
- Reasoning model #
- A model trained to work through a problem in steps — often spending extra tokens “thinking” before it answers — which tends to help on math, code and planning. Stories
- Mixture of experts (MoE) #
- A model design that sends each piece of text to a few specialized sub-networks (“experts”) instead of the whole model, so a very large model can run more cheaply. Stories
- Diffusion model #
- A kind of model that generates images, audio or video by starting from random noise and refining it step by step into a result. Stories
- Embedding #
- A list of numbers that represents the meaning of a piece of text or an image, so similar things sit close together. Used for search, recommendations and retrieval. Stories
- Distillation #
- Training a smaller “student” model to imitate a larger “teacher” model, keeping much of its skill at a fraction of the cost. Stories
Using AI
- System prompt #
- Instructions set by the app’s maker that shape how the model behaves before you type anything. Stories
- Hallucination #
- When a model states something false or made up as if it were true. It’s why AI answers should be checked against sources. Stories
- Retrieval-augmented generation (RAG) #
- Having a model look up relevant documents first and answer from them, so answers can be more current and easier to check. Stories
- Agent #
- An AI system that can take actions over several steps toward a goal — browsing, running code, using apps — rather than answering once. Stories
- Tool use (function calling) #
- A model’s ability to call outside software — a search engine, a calculator, a calendar — and use the result in its answer. Stories
- API #
- A way for other software to use a model directly. Developers usually pay per token instead of a monthly subscription. Stories
Compute
- GPU #
- Graphics processing unit — a chip that does many calculations in parallel, which is exactly what AI needs. Nvidia’s GPUs are the most widely used for AI. Stories
- AI accelerator (TPU and custom chips) #
- Chips designed specifically for AI work, such as Google’s TPUs or Amazon’s Trainium, built as alternatives to GPUs. Stories
- HBM (high-bandwidth memory) #
- Very fast memory stacked right next to AI chips. It feeds data to the chip quickly and is often a bottleneck in supply. Stories
- Compute #
- Shorthand for the processing power used to train and run models — and for the chips, servers and electricity behind it. Stories
- Data center #
- A building full of servers, networking, power and cooling equipment. AI data centers pack in chips that draw far more power per rack than ordinary servers. Stories
- Megawatt and gigawatt #
- Units of power. A gigawatt is 1,000 megawatts — around the output of a large power plant. AI campuses are now planned at hundreds of megawatts and up. Stories
Measuring AI
Safety & policy
- Alignment #
- Work to make AI systems do what their designers and users intend, in line with human values. Stories
- Red teaming #
- Deliberately probing a model for harmful or unsafe behavior before and after release, so problems can be fixed. Stories
- Model card (system card) #
- The document a lab publishes alongside a model describing what it can do, its limits, how it was tested and its safety measures. Stories
- EU AI Act #
- The European Union’s AI law. It sorts uses by risk, bans some, and adds duties for providers of general-purpose models, with rules taking effect in stages. Stories