Reference
Plain-English definitions of AI terms that actually get used. Not academic, not hype, just enough to follow the conversation.
This is a living reference. The goal is practical literacy: enough precision to ask better questions without pretending every term is settled forever.
What a model is, where it comes from, and what runs when you use it.
The trained system you interact with. Under the hood it is a very large mathematical function, but people usually mean the whole thing they chat with.
Like a recipe distilled from watching many cooks. The original examples are gone, but the pattern remains.
The numbers inside a model. Together they encode the patterns learned during training. More parameters can mean more capability, but also more compute.
Like billions of tiny dial positions inside a machine.
The expensive process of building a model by feeding it huge amounts of data, checking predictions, and adjusting the weights again and again.
Running a trained model to get an answer. Every prompt you send to a hosted AI service triggers inference.
A large language model. It predicts and generates text-like sequences, which is why it can write prose, code, plans, summaries, and structured data.
The vocabulary around prompts, context, and generated responses.
A chunk of text the model reads or writes. Tokens can be whole words, parts of words, punctuation, or spaces.
The amount of text the model can consider at once: your prompt, previous messages, tool output, files, and its own response.
A setting that controls randomness. Lower values are more predictable. Higher values are more varied.
The instruction or request you give the model. Good prompts include goal, context, constraints, and the desired output shape.
Higher-priority instructions that describe how the model should behave before the user request is considered.
A confident answer that is not grounded in reality or supplied context. It is usually a prediction problem, not intentional lying.
Ways teams connect models to specific knowledge or reshape behavior.
Additional training that changes a model's behavior for a domain, style, or task. It is not the same as giving the model documents to read.
Retrieval augmented generation. The system searches a knowledge source, adds relevant chunks to the prompt, and asks the model to answer from that context.
Numeric representations of meaning. Similar pieces of text end up close together, which makes semantic search possible.
A database optimized for storing embeddings and finding nearby matches.
Training a smaller model to imitate a larger one. The goal is often speed, cost, or local deployment.
When models stop only answering and start coordinating work.
A system that uses a model to decide steps, call tools, inspect results, and continue until a goal is handled or blocked.
Letting the model call defined functions or external systems, such as search, shell commands, databases, calendars, or code editors.
Model Context Protocol. A standard way for AI tools to connect to external resources and actions.
Task-specific instructions or workflows loaded when relevant. They keep specialized guidance out of the default context until needed.
Using more than one agent or role to divide work, review output, or explore alternatives in parallel.
More public notes about AI tooling and project context.