The AI & Software Glossary

AI applications where the model autonomously plans and executes multi-step tasks using tools.

AI that can perform any intellectual task a human can the long-stated goal of frontier AI labs.

An AI system that pursues a goal by planning, acting, and observing the results of its actions.

AI integrated into a developer's editor or terminal that writes, edits, or reviews code.

An AI agent with a real-time voice interface that can both listen and respond in natural speech.

The transformer operation that lets each token focus on the other tokens most relevant to it.

Asking a model to think step-by-step before answering, which dramatically improves reasoning.

An AI agent specialized for software engineering work reading codebases, running shell commands, editing files.

The maximum number of tokens a model can attend to in a single request, including the response.

An image (or video, audio) generation model that turns noise into content step-by-step.

A numeric vector representation of text (or images, audio) where semantically similar items are close together.

Automated tests that measure how well an AI system performs on representative inputs.

Showing a model a handful of input-output examples so it can generalize to similar inputs.

Continuing to train an existing model on your own data so it specializes for your task.

Two different ways to specialize a model for your use case add knowledge at training time vs at query time.

Letting a language model invoke functions, APIs, or tools and use the results in its answer.

AI that produces new content text, images, audio, video, code rather than just classifying or predicting.

Connecting AI outputs to verified external data sources so answers are based on facts, not training-time guesses.

Programmatic checks around an LLM that prevent unsafe, off-topic, or non-compliant outputs.

When a language model produces fluent text that is factually wrong or unsupported.

A design pattern where AI agents pause and ask a human to approve or correct high-stakes actions.

Running a trained model to produce outputs (as opposed to training the model in the first place).

A neural network trained on huge amounts of text to predict and generate language.

A cheap fine-tuning technique that trains small adapter weights instead of the entire model.

An open protocol from Anthropic for connecting AI models to tools, data, and external systems.

How AI agents retain and retrieve information across turns or sessions.

A system where multiple specialized AI agents collaborate to solve a problem each with its own role and tools.

Models that handle multiple input or output modalities text, images, audio, video, code.

The top-level agent that breaks goals into subtasks and delegates them to specialized subagents.

The discipline of writing inputs to language models that reliably produce useful, structured outputs.

An attack that smuggles malicious instructions into a model's context, hijacking its behavior.

Compressing model weights to lower precision (4-bit, 8-bit) so they're cheaper to run.

Retrieving relevant documents at query time so a language model can answer with grounded, up-to-date information.

Language models trained to think step-by-step before answering, dramatically improving performance on hard problems.

Training a language model to prefer outputs humans rate higher.

Running AI-controlled code in an isolated environment so it can't affect the host system or production.

Search by meaning rather than exact keywords finds 'puppy training' for a query like 'how to teach my dog'.

Forcing a language model's response to fit a specific JSON schema or grammar.

A specialized AI agent that handles one focused task within a larger orchestrated agent system.

A sampling parameter (0-2) that controls how random a model's output is.

The unit a language model processes roughly 4 characters or 0.75 of an English word.

An AI model's ability to invoke external functions (APIs, code runners, browsers) instead of just generating text.

An alternative to temperature: keep only the most-likely tokens that add up to probability p.

The neural network architecture behind nearly every modern AI model, introduced in 2017.

A database optimized for storing and querying high-dimensional vectors (embeddings).

Building software primarily by prompting AI rather than writing code line by line.

The discipline of taking AI-generated prototype code to a production-grade system real users can depend on.

Asking a model to perform a task it was never explicitly trained on, with no examples.