📖

AI glossary

Key AI words explained simply.

62 terms

Accuracy: The share of predictions that are correct.; See the lesson: Training and testing
Agent loop: The cycle an agent repeats: think, act with a tool, observe the result, then continue or finish.; See the lesson: What makes an AI agent
Agent memory: Notes or files an AI saves and reloads into its context later, since the model itself forgets between sessions.; See the lesson: Context engineering
Agent skills: Packaged instructions and files an agent loads only when a task needs them, keeping its context lean.; See the lesson: The harness: everything around the model
Agentic AI: AI systems that take actions toward goals over many steps, rather than only answering questions.; See the lesson: What makes an AI agent
AI agent: An AI system that plans and carries out multi-step tasks by using tools.; See the lesson: What makes an AI agent
Algorithmic bias: When an AI system performs worse or unfairly for some groups of people.; See the lesson: Bias and fairness
Artificial intelligence (AI): Computer systems that perform tasks that normally need human intelligence, like understanding language, recognizing images or making decisions.; See the lesson: What is AI?
Attention: The mechanism that lets each token gather information from other tokens in the context.; See the lesson: Attention and transformers
Backpropagation: The method that computes every weight's gradient by passing the error backwards through the network.; See the lesson: Learning = going downhill
Compaction: Summarizing an agent's older history so it can keep working within a limited context window.; See the lesson: Context engineering
Context engineering: Choosing exactly what information goes into a model's context window at each step: instructions, tools, documents, history and memory.; See the lesson: Context engineering
Context rot: The drop in quality when a context window fills with long, loosely relevant text.; See the lesson: Context engineering
Context window: The maximum number of tokens a model can consider at once.; See the lesson: Tokens: how AI reads text
Convolutional neural network (CNN): A network for images that slides small learned filters across the picture.; See the lesson: How computers see
Data: Recorded examples from the world (images, text, numbers, sounds) that a model learns from.; See the lesson: Data: the fuel of AI
Decision tree: A model that makes decisions by asking a series of yes/no questions learned from data.; See the lesson: Decision trees
Deep learning: Machine learning with neural networks that have many layers.; See the lesson: From neurons to networks
Deepfake: A realistic fake image, voice or video created with AI.; See the lesson: Hallucinations and fact-checking
Embedding: A list of numbers that represents the meaning of a token, image or document.; See the lesson: Embeddings: meaning as geometry
Eval (evaluation): A set of test cases used to measure how well a prompt, model or agent performs.; See the lesson: Prompt engineering
Feature: A measurable property of an example that the model uses as a clue.; See the lesson: Data: the fuel of AI
Few-shot prompting: Putting a few example inputs with their ideal outputs in the prompt so the model copies the pattern.; See the lesson: Prompt engineering
Fine-tuning: Further training of a pre-trained model on specific examples to specialize it.; See the lesson: How LLMs are trained
Generalization: A model's ability to perform well on examples it has never seen.; See the lesson: Overfitting and generalization
Generative AI: AI that creates new content such as text, images, audio, video or code.; See the lesson: The AI landscape
Gradient descent: Training by repeatedly nudging parameters in the direction that lowers the loss.; See the lesson: Learning = going downhill
Hallucination: When an AI confidently states something false or made up.; See the lesson: Hallucinations and fact-checking
Harness: The software around a model that turns it into an agent: the loop, tools, context management, permissions, sandbox and interface.; See the lesson: The harness: everything around the model
Human in the loop: A person reviews or approves an AI's key actions before they happen.; See the lesson: Using agents safely
Label: The correct answer attached to a training example.; See the lesson: Data: the fuel of AI
Large language model (LLM): A very large neural network trained on huge amounts of text to predict the next token.; See the lesson: Next-token prediction
Learning rate: How big each gradient descent step is.; See the lesson: Learning = going downhill
Loss function: A number measuring how wrong a model's predictions are. Training tries to minimize it.; See the lesson: Learning = going downhill
Machine learning: A way of building AI where computers learn patterns from examples instead of following rules written by hand.; See the lesson: How do machines learn?
Model Context Protocol (MCP): An open standard that connects AI apps (clients) to tools and data sources (servers) through one common plug.; See the lesson: Tools, function calling and MCP
Neural network: Many artificial neurons connected in layers, learning together.; See the lesson: From neurons to networks
Neuron (artificial): A small unit that multiplies inputs by weights, adds a bias and applies an activation.; See the lesson: Artificial neurons
Open-weight model: A model whose parameters are published, so anyone can download and run it.; See the lesson: Choosing the right model and tool
Overfitting: When a model memorizes its training data, noise included, and does badly on new data.; See the lesson: Overfitting and generalization
Parameter: A learned number in a model (a weight or bias). Large models have billions.; See the lesson: From neurons to networks
Prompt: The instructions, question and material you give to an AI model.; See the lesson: Prompting like a pro
Prompt engineering: Designing, testing and improving the instructions an AI model runs on.; See the lesson: Prompt engineering
Prompt injection: Instructions hidden in content an AI reads (a web page, email or file) that try to hijack what it does.; See the lesson: Using agents safely
Reasoning model: A model trained to think through a problem step by step before giving its final answer.; See the lesson: Prompt engineering
Reinforcement learning: An agent learns by trial and error, guided by rewards and penalties.; See the lesson: Learning by trial and error
Retrieval-augmented generation (RAG): An assistant first finds relevant passages in a set of documents, then answers using them.; See the lesson: Embeddings: meaning as geometry
RLHF: Reinforcement learning from human feedback: tuning a model toward answers people prefer.; See the lesson: How LLMs are trained
Sandbox: An isolated environment where an agent can run code or test changes without risking the real system.; See the lesson: The harness: everything around the model
Structured output: An answer in a fixed, machine-readable format such as JSON with named fields.; See the lesson: Prompt engineering
Subagent: A helper agent with its own fresh context that handles part of a job and reports back to a lead agent.; See the lesson: What makes an AI agent
Supervised learning: Learning from examples that come with labels (the right answers).; See the lesson: Data: the fuel of AI
System prompt: Standing instructions from an app's builder that set the model's role, rules and tone for every conversation.; See the lesson: Prompt engineering
Temperature: A setting that makes a model's output more predictable (low) or more varied (high).; See the lesson: Next-token prediction
Test data: Examples kept hidden during training to fairly measure performance on new data.; See the lesson: Training and testing
Token: A chunk of text (word, part of a word or character) that a language model reads and writes.; See the lesson: Tokens: how AI reads text
Tool use (function calling): A model outputs a structured request to use a tool (search, code, calendar); the app runs it and returns the result.; See the lesson: Tools, function calling and MCP
Training data: The examples a model learns from.; See the lesson: Training and testing
Transformer: The neural network architecture, built on attention, behind modern language models.; See the lesson: Attention and transformers
Unsupervised learning: Finding structure, like groups, in data without labels.; See the lesson: Data: the fuel of AI
Vibe coding: Building software mostly by describing what you want to an AI coding agent and accepting its code, often without reading it closely. Fast for prototypes, risky for serious apps.; See the lesson: The harness: everything around the model
Weight: A number that sets how much an input influences a neuron. Learned during training.; See the lesson: Artificial neurons