AI education
AI glossary
The words of AI, in plain language. Flip through the cards from the easiest essentials to the harder ideas, or read the whole list below. Free for everyone.
Flashcards
Card 1 of 51
All 51 terms
Easiest first. Each term has its own link, so you can share one definition.
Essential · The words you will meet first.
- Artificial intelligence(AI)
- Software that does tasks we usually think need human intelligence, such as understanding language, recognizing pictures, or making recommendations.See also: Machine learning, Generative AI
- Algorithm
- A step-by-step recipe a computer follows to solve a problem or reach a result.See also: Machine learning
- Machine learning(ML)
- A way of building AI where the computer finds patterns in many examples, instead of following rules a person wrote out by hand.See also: Training data, Model, Deep learning
- Training data
- The examples a model learns from. Its quality and variety shape what the model does well and where it falls short.See also: Training, Bias
- Model
- What training produces: a large set of numbers that captures patterns from the data and can then be used to make predictions or create new content.See also: Training, Inference, Parameters
- Training
- The process of adjusting a model's numbers over many examples until it gets good at its task. Training is learning.See also: Inference, Pre-training
- Inference
- Using a trained model to produce an answer. Training is learning; inference is putting what was learned to work, every time you send a prompt.See also: Training, GPU
- Pattern recognition
- Noticing regularities and sorting things into groups, such as telling a cat from a dog or a receipt from an invoice. Much of machine learning is pattern recognition at scale.See also: Machine learning, Neural network
- Prompt
- The question or instructions you give an AI model. Clear context, a goal, and an example of what you want usually give better answers.See also: Large language model, Context window
- Generative AI
- AI that creates new content, such as text, images, audio, or code, rather than only sorting or labeling what already exists.See also: Large language model
- Large language model(LLM)
- A model trained on huge amounts of text to predict the next word again and again, which lets it write, summarize, translate, and answer questions. ChatGPT, Claude, and Gemini are built on LLMs.See also: Token, Transformer, Context window
- Chatbot
- A program you talk with in everyday language. Today's chatbots are usually powered by an LLM and reply in conversation.See also: AI assistant, AI agent
- AI assistant(Copilot)
- AI that works alongside you inside a tool you already use, drafting or suggesting while you stay in charge of the result.See also: Chatbot, Human in the loop, AI agent
- Hallucination
- When an AI states something that sounds confident and fluent but is wrong or invented. Check anything important against a source you trust.See also: Retrieval-augmented generation, Human in the loop
- Bias
- A steady tilt in AI results, often inherited from gaps or imbalances in the training data. Knowing it exists is the first step to catching it.See also: Training data, Alignment
- Human in the loop
- A way of working where a person reviews or approves what the AI produced before it takes effect. It keeps judgment, and responsibility, with people.See also: AI assistant, Guardrails
Intermediate · How today's AI is built and used.
- Neural network
- A model built from many small connected units, called artificial neurons, arranged in layers and loosely inspired by the brain. Each unit does a simple calculation; together they learn complex patterns.See also: Deep learning, Parameters
- Deep learning
- Machine learning with neural networks that have many layers. It powers most of the recent progress in language, images, and speech.See also: Neural network, Transformer
- Parameters
- The adjustable numbers inside a model that training sets. People often describe a model's size by how many it has, in billions.See also: Model, Training
- Token
- A small piece of text, such as a word, part of a word, or punctuation mark, that a language model reads and writes. Limits and prices are usually counted in tokens.See also: Context window, Embedding
- Embedding
- A list of numbers that stands for the meaning of a word, sentence, or image, so that things with similar meanings land close together. It is what makes the famous "king minus man plus woman is about queen" arithmetic possible.See also: Vector database, Token
- Attention
- The mechanism that lets a model weigh which parts of its input matter most for the word it is about to produce, instead of treating every word equally.See also: Transformer, Context window, Sliding window attention
- Transformer
- The neural-network design behind modern LLMs. It repeats blocks that use attention to relate every piece of the input to every other piece. It was introduced in the 2017 paper "Attention Is All You Need".See also: Attention, Large language model, GPU
- Context window
- How much text, counted in tokens, a model can consider at one time, including your prompt, any documents you add, and its own reply. Anything outside it is not seen.See also: Token, Attention, Sliding window attention
- GPU(Graphics processing unit)
- A chip that does huge numbers of simple calculations at the same time. Neural networks are mostly matrix multiplication, which suits GPUs well, so they are the workhorse of AI.See also: Inference, Training
- Pre-training
- The first and most expensive stage, where a model learns general patterns of language from enormous amounts of text. The result is a broad but unpolished base model.See also: Post-training, Fine-tuning
- Fine-tuning
- Training a pre-trained model a little further on a smaller, focused set of examples so it behaves in a particular way or handles a specialty.See also: Pre-training, Post-training
- Post-training
- The steps after pre-training that turn a raw base model into a helpful, safer assistant, such as teaching it to follow instructions and to prefer answers people rate well.See also: Pre-training, Reinforcement learning from human feedback, Alignment
- Reinforcement learning from human feedback(RLHF)
- A post-training method in which people rank a model's answers and the model is adjusted to produce the kinds of answers people rate higher.See also: Post-training, Alignment
- Temperature
- A setting for how adventurous a model's word choices are. Low values give steady, predictable answers; high values give more varied, creative ones.See also: Prompt, Hallucination
- Few-shot prompting
- Putting a few worked examples in the prompt so the model copies the pattern. With no examples it is called zero-shot.See also: Prompt, Context window
- Multimodal
- Able to work with more than one kind of information, such as text, images, audio, and video, in the same model.See also: Generative AI, Embedding
- Retrieval-augmented generation(RAG)
- Look up relevant passages from your own documents first, then give them to the model along with the question, so the answer is grounded in sources you can check.See also: Vector database, Agentic RAG, Hallucination
- Vector database
- A database that stores embeddings and quickly finds the items closest in meaning to a query. It is often the lookup step inside RAG.See also: Embedding, Retrieval-augmented generation
- Evaluation(Evals)
- Tests that measure how well an AI system does its job on realistic examples. Evals are how you know whether a change really helped.See also: Hallucination, Guardrails
Advanced · Agents, safety and the systems around the model.
- Sliding window attention
- A way to save computing on long inputs: each token attends only to nearby tokens inside a window that moves along the text, rather than to every earlier token.See also: Attention, Context window
- AI agent
- A system in which a model plans, takes steps toward a goal using tools, checks what happened, and keeps going. A chatbot answers; an agent acts.See also: Tool use, Agent harness, Chatbot
- Tool use(Function calling)
- The model asks the program around it to run a tool, such as search, a calculator, or a database, and then continues with the result.See also: AI agent, Model Context Protocol, Code execution
- Agent harness
- The software around the model that makes it an agent: the loop, the tools, memory, and permissions. A useful rule of thumb: agent = model + harness.See also: AI agent, Tool use, Guardrails
- Agentic RAG
- RAG where an agent decides what to look up, rewrites its own searches, checks the results, and looks again, instead of doing one fixed lookup.See also: Retrieval-augmented generation, AI agent
- Model Context Protocol(MCP)
- An open standard for connecting AI applications to tools and data sources in one consistent way, so each connection does not have to be built from scratch.See also: Tool use, AI agent
- Code execution
- Letting a model write code and run it, usually in a sandbox, to get exact results or to do real work such as analyzing a file.See also: Sandbox, Tool use
- Sandbox
- A walled-off space where code can run without being able to touch the rest of the system. Agents that run code do it in a sandbox so a mistake stays contained.See also: Container, Virtual machine, Code execution
- Container
- A lightweight package holding an app and everything it needs, run in isolation on a shared operating system. Containers start fast and are the usual unit for deploying software.See also: Virtual machine, Sandbox
- Virtual machine(VM)
- A whole simulated computer, with its own operating system, running inside a real one. It is heavier than a container but separates workloads more strongly.See also: Container, Sandbox
- Guardrails
- Rules and checks around an AI system that keep what it says and does within safe, intended limits.See also: Human in the loop, Prompt injection
- Prompt injection
- An attack where text hidden in a web page, file, or message tricks an AI into following instructions its user never gave. It is a key risk for agents that read untrusted content.See also: Guardrails, AI agent
- Alignment
- The work of making AI systems pursue the goals people actually intend and respect the values people hold.See also: Reinforcement learning from human feedback, Bias, Guardrails
- Mixture of experts
- A model design where only a few specialized sub-networks switch on for each input, so a very large model costs much less to run than its size suggests.See also: Parameters, Inference
- Distillation
- Training a smaller model to imitate a larger one, keeping much of the ability at a fraction of the cost.See also: Quantization, Parameters
- Quantization
- Storing a model's numbers at lower precision to cut memory and cost, usually with only a small loss in quality.See also: Distillation, Inference
Want to go further?
These ideas are taught hands-on in CSE 642, Agentic AI and Application. See the course page, or read more in AI education.