AI Glossary

The core innovation in Transformer models that allows AI to weigh the importance of different parts of input text when generating each output token.

An AI system that can autonomously take actions, use tools, and complete multi-step tasks without human intervention at each step.

The challenge of ensuring AI systems behave in accordance with human values, intentions, and societal well-being.

AI systems that operate autonomously over extended periods, taking sequences of actions to complete complex goals.

A software interface that allows developers to access AI model capabilities programmatically — the foundation of all AI-powered products.

The field of research and practice focused on building AI systems that are safe, reliable, controllable, and aligned with human values.

A standardized test used to measure and compare AI model capabilities across specific tasks.

The maximum amount of text (measured in tokens) that an AI model can 'see' and process in a single interaction.

A prompting technique where the AI is guided to reason step-by-step before producing a final answer, significantly improving accuracy on complex tasks.

Anthropic's technique for training Claude to be helpful, harmless, and honest by having AI models critique and revise their own outputs based on a set of principles.

The AI architecture behind most image, audio, and video generation — learns to reverse a noise-adding process to generate content.

Numerical vector representations of text that capture semantic meaning, allowing AI to find conceptually similar content.

Further training a pre-trained AI model on a smaller, task-specific dataset to specialize its behavior.

Providing a small number of input-output examples in the prompt to guide the AI's response format and style.

A large AI model trained on broad data at scale that can be adapted for many different downstream tasks.

Connecting AI model responses to verified, real-world information sources to reduce hallucination and improve accuracy.

When an AI model generates confident-sounding but factually incorrect or fabricated information.

The process of running a trained AI model to generate outputs — what happens when you use an AI tool.

The date after which an AI model has no training data — the 'last updated' date of its knowledge.

A type of AI model trained on vast amounts of text data that can generate, summarize, translate, and reason about language.

The time delay between sending a request to an AI API and receiving the first token of response.

AI systems that can process and generate multiple types of data — text, images, audio, video, and code.

Training a smaller AI model to mimic a larger, more capable model — creating efficient models that approach the larger model's quality.

AI models whose weights and (sometimes) training code are publicly released, allowing anyone to run, modify, and build on them.

The practice of crafting inputs to AI models to reliably get better, more accurate, or more specific outputs.

A technique for reducing AI model size and memory requirements by representing weights with lower precision numbers.

Reinforcement Learning from Human Feedback — a training technique used to align AI models with human preferences and values.

A technique that enhances LLM responses by retrieving relevant documents from an external knowledge base before generating an answer.

AI models specifically optimized for multi-step logical reasoning, math, and complex problem-solving — typically by using chain-of-thought at inference time.

Constraints imposed by AI API providers on how many requests or tokens a user can process per minute, hour, or day.

Instructions given to an AI model before the user conversation begins, shaping its behavior, persona, and constraints.

Constraining an AI model to generate responses in a specific format (JSON, XML, etc.) for reliable programmatic processing.

Search that finds results based on meaning and intent, not just keyword matching.

The basic unit of text that AI language models process — roughly equivalent to 3/4 of a word in English.

The neural network architecture that underpins all modern large language models, introduced by Google in 2017.

A parameter (0 to 2) that controls how random or deterministic an AI model's outputs are.

A sampling parameter that limits token selection to the smallest set of tokens whose cumulative probability exceeds P.

The ability for AI models to call external functions, APIs, or tools to retrieve information or take actions beyond their training data.

The process of splitting text into tokens (the basic units an LLM processes) before feeding it to a model.

A database optimized for storing and searching embedding vectors — the foundation of RAG and semantic search applications.

Asking an AI to perform a task it wasn't explicitly trained on, relying solely on its general capabilities.