AI coverage runs on jargon that didn’t exist five years ago, half of it overloaded with multiple meanings. This glossary defines the terms that actually come up in day-to-day AI news and product decisions, organized by what they describe rather than alphabetically, so related concepts sit next to each other.
Models and how they’re built
- Large language model (LLM): A model trained on huge amounts of text to predict the next piece of language, which turns out to be enough to answer questions, write code, and follow instructions.
- Parameters: The internal numeric values a model adjusts during training. More parameters generally mean more capacity to learn patterns, but not always better real-world performance.
- Mixture of Experts (MoE): An architecture that splits a model into specialized sub-networks (“experts”) and only activates a subset for each input, letting a model have a huge total parameter count while only using a fraction of it per request, saving compute.
- Context window: How much text (measured in tokens) a model can consider at once. A larger context window lets a model work with longer documents, codebases, or conversation history without losing track of earlier details.
- Token: The basic unit a model reads and generates, roughly a word or word-fragment. Pricing, context limits, and generation speed are all measured in tokens.
- Fine-tuning: Further training a general model on a narrower, specific dataset to specialize its behavior for a particular task or domain.
- Reasoning model: A model trained to generate intermediate reasoning steps before its final answer, generally improving performance on math, coding, and multi-step logic at the cost of speed.
- Hallucination: When a model generates confident, plausible-sounding information that’s actually false or fabricated. It’s a known limitation, not a bug specific to any one product.
- Open-weight model: A model whose trained parameters are published for anyone to download and run themselves, as opposed to a closed model only accessible through an API.
Agents and how they act
- AI agent: A system built on top of a model that can plan, take actions, use tools, and adjust based on results, rather than just answering a single prompt. See our full AI Agents coverage.
- Agentic workflow: A multi-step process where an agent breaks a task into stages, executes them, checks its own results, and adjusts course, rather than doing everything in one pass.
- Model Context Protocol (MCP): A standard that lets an AI model connect to external tools, files, and services in a consistent way, rather than requiring a custom integration for every product. See the official MCP specification, or read our deep dive on MCP’s 2026 rewrite.
- Computer-use agent: An agent that interacts with a computer the way a person would, through screenshots, clicks, and keystrokes, rather than through a dedicated API.
- Tool use / function calling: A model’s ability to call an external function, API, or tool mid-response, such as running a calculation, searching the web, or querying a database.
- Orchestrator / subagent: In multi-agent systems, an orchestrator agent breaks work into pieces and assigns them to specialized subagents running in parallel, then combines their output.
Infrastructure and physical AI
- Inference: Running a trained model to generate a response, as opposed to training it. Inference cost and speed are what most product pricing is actually based on.
- Training compute: The processing power spent building a model in the first place, a one-time cost, distinct from the ongoing cost of running it afterward.
- World model: An AI system trained to predict how the physical world behaves, objects, forces, causality, rather than just predicting text. See our explainer on physical AI and world models.
- Physical AI: The broader category of AI systems designed to perceive and act in the real world, spanning robotics, autonomous vehicles, and world models.
- Reality gap (sim-to-real gap): The performance drop that happens when a system trained in simulation is deployed on real hardware, caused by imperfect physics modeling in the simulation.
Safety and governance
- Guardrails: Safety mechanisms built into a model or the system around it that restrict harmful, dangerous, or policy-violating outputs.
- Sandboxing: Running a model or agent in an isolated environment specifically so it can’t affect real systems, used heavily in AI safety testing. See our coverage of what happened when sandboxing failed at two major AI labs in the same month.
- Red teaming: Deliberately trying to make a model misbehave, produce harmful content, or bypass its safety training, to find and fix weaknesses before real-world release.
- RLHF (Reinforcement Learning from Human Feedback): A training technique that uses human ratings of model outputs to steer a model toward more helpful, accurate, or safe behavior.
Key takeaway
This list will keep growing as the field does; bookmark it and check back when a new term shows up in our coverage that you haven’t seen defined plainly elsewhere.
