AI Explained.
Finally.

An LLM is the type of AI model that powers most chatbots you use today. It was trained on massive amounts of text — books, websites, code — and learned to predict and generate language. "Large" refers to the billions of numbers (called parameters) it uses to make decisions. Think of it as a very sophisticated autocomplete engine that got really, really good.

Generative AI creates new content from scratch — text, images, music, video, code — rather than just sorting or labeling existing things. Old AI: "Is this email spam? Yes/No." Generative AI: "Write me a professional reply to this email." It's a fundamentally different capability that opened up in the last few years.

A prompt is simply what you type into an AI tool. It's your instruction, question, or request. The quality of what you put in directly shapes what you get back. "Prompting" has become a skill in itself — and the good news is it's not complicated. Be clear, give context, and ask for the format you want.

AI doesn't read word-by-word like you do. It reads in "tokens" — chunks of text that might be a word, part of a word, or a punctuation mark. Roughly 1 token = ¾ of a word. This matters because AI tools have token limits — caps on how much text you can send or receive in one go. Longer conversations or documents use more tokens.

A hallucination is when AI invents facts that sound real but aren't. It might cite a paper that doesn't exist, give you the wrong date, or describe a person who never lived — all with total confidence. This is the #1 risk with AI today. AI isn't lying — it's pattern-matching without fact-checking. Always verify important facts independently.

The context window is like the AI's short-term memory — how much text it can read and remember during a single conversation. Early models could hold a few paragraphs. Modern models can hold hundreds of pages. If you exceed the limit, the AI starts "forgetting" earlier parts of the conversation — like talking to someone with a very short attention span.

An AI agent doesn't just answer questions — it takes actions. It can browse the web, write and run code, send emails, book calendar slots, or fill out forms on your behalf. You give it a goal, and it figures out the steps. This is the fastest-moving area of AI right now, and it's where things get both very exciting and worth watching carefully.

MCP is a technical standard (created by Anthropic) that lets AI tools connect to external apps and data — like Gmail, Slack, Google Drive, or a company's database. Think of it like USB-C for AI: a universal connector so AI can plug into different systems without custom code for each one. It's what makes AI agents actually useful in real workflows.

RAG is a technique that gives AI a reference library to look up before responding. Instead of relying purely on what it memorized during training, the AI first retrieves relevant documents or data, then generates an answer grounded in those sources. This dramatically reduces hallucinations and keeps answers current. It's how AI can answer questions about your company's internal docs.

Fine-tuning is additional training applied to a base model to specialize it for a specific task, tone, or industry. A general AI model fine-tuned on medical records starts responding more like a doctor. Fine-tuned on legal briefs, it writes like a lawyer. Most companies using AI in serious ways are doing some form of fine-tuning to make it fit their needs.

Training data is the massive collection of text, images, and other content that an AI model learned from. For LLMs, this includes huge portions of the internet, books, academic papers, and code — up to a certain date. The AI doesn't memorize this data exactly, but learns patterns from it. What's in the training data shapes what the AI knows and how it thinks. Biases in the data become biases in the model.