Key AI Terms Explained
Essential terminology you need to understand AI news. Each term is explained in plain language with context for how it appears in current reporting.
LLM (Large Language Model)
A neural network trained on vast text corpora to predict the next token, enabling it to generate fluent, human-like responses across open-ended tasks.
Foundation Model
A large model pre-trained on broad data that serves as a reusable base — adaptable to many downstream tasks through fine-tuning or prompting rather than training from scratch.
AGI (Artificial General Intelligence)
A hypothetical system that matches or exceeds human-level reasoning across essentially all cognitive tasks, as opposed to today's narrow, task-specific models.
Fine-Tuning
Continuing training of a pre-trained model on a smaller, targeted dataset so it specializes in a domain, style, or task while retaining its general capabilities.
Inference
The process of running a trained model to produce outputs (predictions, text, images). Inference cost and latency — not training — dominate the economics of deployed AI products.
Transformer
The neural-network architecture behind modern LLMs, using self-attention to weigh the relevance of every token to every other, which made large-scale language modeling practical.
Attention Mechanism
The component that lets a model decide which parts of the input matter most for each output token — the core idea that powers transformers.
Context Window
The maximum amount of text (measured in tokens) a model can consider at once. Larger windows let a model reason over longer documents but raise compute cost.
Token
The unit of text a model reads and generates — roughly a word-piece. Pricing, context limits, and throughput are all measured in tokens, not words.
RAG (Retrieval-Augmented Generation)
A pattern that fetches relevant documents from an external store and feeds them into the prompt, grounding answers in current or proprietary data and reducing hallucination.
Hallucination
When a model produces fluent but factually wrong or fabricated output. Mitigations include retrieval grounding, citations, and verification — a central reliability challenge for production AI.
RLHF (Reinforcement Learning from Human Feedback)
A training stage that uses human preference rankings to align a model's outputs with what people find helpful, honest, and safe.
Embedding
A numeric vector that represents the meaning of text (or images) so that semantically similar items sit close together — the basis of search, clustering, and RAG.
Parameters
The learned weights of a model. Counts (e.g. 7B, 70B, 400B+) loosely indicate capacity, though data quality and training method matter as much as raw size.
Mixture of Experts (MoE)
An architecture that routes each token to a few specialized sub-networks, giving a large model's quality at a fraction of the compute per token.
Multimodal Model
A model that processes more than one modality — text, images, audio, or video — in a shared representation, enabling tasks like describing an image or answering questions about a chart.
Quantization
Compressing a model by storing its weights at lower numeric precision (e.g. 4-bit), cutting memory and cost with minimal quality loss — key to running models on cheaper hardware.
Frontier Model
The current most-capable generation of foundation models from leading labs, typically the largest and most expensive to train and the focus of safety and policy debate.
Alignment
The field concerned with making AI systems pursue intended goals and values — covering helpfulness, honesty, harmlessness, and avoiding unintended behavior at scale.
Agent
An LLM-based system that plans and takes multi-step actions using tools (search, code, APIs) toward a goal, rather than returning a single response.
Stay Informed
These terms appear frequently in AI reporting. Bookmark this page as a quick reference when reading our latest intelligence briefs.
Terms are sourced from stories that clear our classification pipeline at a minimum 35 percent relevance threshold. According to that methodology, reviewed July 2026, this follows multi-source corroboration standards recommended by journalism research bodies such as the Reuters Institute for the Study of Journalism.