AI · LANGUAGE · SYSTEMS
AI Glossary
Concise definitions of essential terms covering artificial intelligence, models, agents, and safe product development.
- Agent
- An AI system that pursues a goal, plans steps, uses tools, and processes results or state across multiple actions.
- AI-native
- A product or company whose core workflow is designed around AI capabilities from the start instead of adding AI to existing software later.
- Artificial Intelligence (AI)
- The broad field of computer systems that perform tasks involving perception, language, reasoning, learning, or decision-making.
- Benchmark
- A standardized test or dataset used to compare model capability, quality, speed, or cost.
- Context Window
- The limited number of tokens a model can consider at once for a request.
- Embedding
- A numerical vector representation of text, images, or other data in which similar content can be located close together.
- Evaluation (Evals)
- Systematic tests that measure the quality, reliability, safety, and behavior of an AI system for specific tasks.
- Fine-tuning
- Additional training of a pretrained model on selected examples to adapt its behavior or domain capability.
- Foundation Model
- A broadly pretrained model that can serve as the base for many downstream tasks and products.
- Guardrail
- A technical or organizational safeguard that constrains and checks an AI system's inputs, outputs, or actions.
- Hallucination
- A plausible-sounding output from a generative model that is factually wrong or unsupported by evidence.
- Inference
- Running a trained model to produce a prediction, response, or action from an input.
- Large Language Model (LLM)
- A large model trained on extensive text and often multimodal data to process and generate language.
- Machine Learning (ML)
- A field within AI in which systems learn patterns from data rather than relying only on explicitly programmed rules.
- Multimodal
- The ability to process or generate multiple data types such as text, images, audio, or video together.
- Open Weights
- Publicly available trained model parameters. This does not automatically mean the training data, code, or license is fully open.
- Prompt
- The input or instruction describing context, task, constraints, and the desired output format for a generative model.
- Retrieval-Augmented Generation (RAG)
- A method that retrieves relevant external information and supplies it as context for a better-grounded model response.
- Token
- A unit processed by a language model that may correspond to a word, word fragment, punctuation mark, or other symbol.
- Tool Use
- The ability of a model or agent to call defined functions, search systems, databases, or other external tools.
- Training
- The compute-intensive process of adjusting model parameters across many examples so the model learns patterns and tasks.
- Transformer
- A neural-network architecture that processes relationships within an input sequence through attention mechanisms.
- Vector Database
- A data system that stores embeddings and retrieves content according to numerical similarity.