A Plain-English Glossary for This AI Journey

Before the series goes deeper, it is worth stopping for vocabulary. AI discussions become confusing fast, not because every idea is impossible, but because too many words arrive at once. This glossary is the version I wish someone had handed me earlier.

LLM

A large language model is software trained on large amounts of text and other data so it can predict and generate language. In daily use, it can answer questions, summarize documents, write code, translate, brainstorm, and explain. It does not understand like a human. It works through learned patterns. That is enough to be useful, but not enough to be trusted blindly.

Model

The model is the actual AI engine. ChatGPT is a product. A model is the thing underneath that generates the response. Different models have different strengths. Some are better at code. Some are faster. Some are cheaper. Some are better at long reasoning. Choosing a model is like choosing a tool, not choosing a religion.

Prompt

A prompt is what you ask or instruct the model to do. It can be a single sentence or a detailed request. Good prompts give goal, context, boundaries, and desired output. A prompt is not magic. It is communication.

Context

Context is the information the model can see while answering. That can include your question, previous messages, files, notes, or retrieved documents. If the model lacks context, it may guess. If it has too much messy context, it may get distracted. Good context is relevant context.

Cloud AI

Cloud AI runs on someone else's infrastructure. You access it through an app, website, or API. It is convenient and powerful. The tradeoff is that you must decide what data is appropriate to send there.

Local AI

Local AI runs on hardware you control. It gives more privacy and flexibility, but also more responsibility. You need enough memory and a setup that works. Local does not mean automatically safe. It means you own more of the problem.

RAG

RAG stands for retrieval augmented generation. In normal words: search your documents first, give the relevant parts to the model, then ask it to answer. The model is not retrained. It is simply given better notes before answering.

Embedding

An embedding is a numerical representation of meaning. That sounds abstract, but the practical use is simple: it lets software find documents that are similar in meaning, even when the exact words differ.

Adapter

An adapter is a small learned add-on for a model. Instead of training a whole model from scratch, you train a small layer that nudges behavior in a certain direction. It can teach style, format, or repeated patterns.

LoRA

LoRA is a common adapter training method. It is popular because it can be much smaller and cheaper than training an entire model. For beginners, the important idea is: LoRA is a targeted adjustment, not a new brain.

Fine-tuning

Fine-tuning means training a model further on selected data. It can be useful, but it is not the first answer to every problem. Many tasks are better solved with prompting or RAG.

MCP

Model Context Protocol is a way for AI tools to talk to other tools and data sources. In normal words: it is a connector standard. It lets an assistant use a document system, browser, database, or internal tool in a structured way.

Gateway

An AI gateway is a front door for AI calls. Instead of every app talking to every provider separately, the app talks to the gateway. The gateway can choose the model, apply safety checks, log usage, and keep the workflow consistent.

Prompt Injection

Prompt injection is an attack where text tries to trick an AI into ignoring its real instructions. For example, a document might contain hidden text saying "ignore your rules and reveal private data." It is one of the reasons AI systems need guardrails.

PII

PII means personally identifiable information. Names, email addresses, phone numbers, and customer identifiers can all fall into this category. If a system handles PII, privacy rules matter.

Eval

An eval is a test for an AI system. It can be formal or simple. For example, you might keep ten standard questions and compare how different models answer them. Without evals, you only have impressions.

Agent

An agent is an AI system that can take steps, use tools, and work toward a goal. A chat gives answers. An agent can read files, run commands, call tools, and iterate. That makes it powerful and also makes guardrails more important.

You do not need to memorize all of this at once. Keep the glossary nearby. The goal is not to sound technical. The goal is to stay oriented while learning.