AI GLOSSARY
AI Glossary
Short, precise definitions of the terms that show up across our AI and semiconductor coverage — written so you can look one up in thirty seconds and get back to the article.
Fundamentals
LLM (Large Language Model)
An AI system trained on massive amounts of text to predict and generate human-like language.
Transformer
The neural network architecture, introduced in 2017, that made modern large language models possible.
Token
The small chunks of text — often pieces of words — that a language model actually reads and generates, one at a time.
Context Window
The maximum amount of text — measured in tokens — a model can “see” at once, including your prompt and its own reply.
GPU (Graphics Processing Unit)
A chip originally built to render graphics that turned out to be the ideal engine for AI, because it can do thousands of simple calculations at once instead of one at a time.
Parameter
One of the internal numbers a neural network adjusts during training — the '70B' or '400B' quoted in a model's name is a count of these.
Multimodal AI
A model that can take in, or generate, more than one type of data — text, images, audio, video — instead of only text.
Model Lifecycle
Inference
The step where a trained AI model is actually used to generate an answer — as opposed to training, when it learns.
Training (Pretraining)
The process of teaching a model its base knowledge and language ability by having it predict text across a huge dataset.
Fine-tuning
Additional training on top of a pretrained model to specialize its behavior — for a task, a style, or a company’s own data.
RAG (Retrieval-Augmented Generation)
Giving a model outside documents to read at answer time, instead of relying only on what it memorized during training.
Quantization
Shrinking a model's internal numbers down to lower precision so it runs faster and fits in far less memory — often the difference between needing a data center and running on a phone.
Behavior & Risk
Hallucination
When an AI model states something false or fabricated with the same confidence as a correct answer.
Agentic AI (AI Agents)
AI systems that don't just answer a single question, but plan multiple steps and take actions — calling tools, browsing, writing and running code — to complete a goal.
Chain-of-Thought (Reasoning Model)
A model that writes out intermediate reasoning steps before its final answer, which measurably improves accuracy on math, logic, and multi-step problems.
Infrastructure
MoE (Mixture of Experts)
A model architecture that splits its parameters into specialized “expert” sub-networks and activates only a few of them per request, instead of the whole model.
HBM (High Bandwidth Memory)
A type of memory stacked directly next to an AI chip to feed it data fast enough to keep up with its compute — a key bottleneck in the AI buildout.
CUDA
Nvidia's proprietary software platform for programming its GPUs — widely seen as the biggest reason Nvidia's hardware dominance has been so hard for rivals to dislodge.
Foundry (Semiconductor Foundry)
A company that manufactures chips designed by other companies — TSMC and Samsung are the only two that can currently make the most advanced ones at scale.
EUV Lithography
The extreme-ultraviolet light process used to print the smallest, most advanced features onto a chip — made by exactly one company in the world, ASML.