Every week a new AI word shows up in a meeting, and everyone nods like they know what it means. ML, LLM, generative, RAG, agents. Most of it is simpler than it sounds. This is the field guide we give clients before we talk strategy: plain English, real examples, and a single picture that makes the whole vocabulary click. It is the first of a three-part series. Part two covers how you actually put AI to work, and part three is where it all comes together.
Key Takeaways
- AI is an umbrella term, not one technology. Almost everything people call AI today is a branch of it called machine learning.
- An LLM (the thing behind ChatGPT, Claude, and Gemini) is a machine that predicts the next word. That is why it writes so well, and also why it can be confidently wrong.
- Generative AI is the "creates new content" kind. Text, images, video, voice, and code are all flavors of the same idea.
- Coding AI is not a separate species. It is an LLM that has read enormous amounts of software.
- Grounding (RAG) is how you make a general model answer from your own documents instead of guessing.
Start with the family tree
AI (Artificial Intelligence) is the umbrella term for any computer system that does things we normally think require human smarts: understanding language, recognizing patterns, making predictions, or making decisions. It is not one technology. It is a family of them, the same way "vehicles" includes bikes, cars, and semis. Think of AI as the word "sports." Machine learning, LLMs, and agents are basketball, soccer, and hockey. Different games, same umbrella.
The cleanest way to hold it in your head is as a set of nested circles. Each one lives inside the one around it.
An LLM is a kind of deep learning, which is a kind of machine learning, which is a kind of AI. Generative AI is the "creates new content" kind: LLMs are its text flavor, and image and audio generators are its other flavors.
Machine learning: prediction from patterns
Machine learning is software that learns patterns from examples instead of following rules a programmer wrote by hand. You show it many labeled examples, say past intakes marked "returned" or "did not return," and it adjusts itself until it can guess the label on cases it has never seen. Nobody writes the rules by hand. The rules are discovered in the data.
Use it when you have lots of historical data and want a prediction or a score: which clients are likely to miss appointments, which leads are worth a call, how long a case will take to close. Two cautions. It is only as good as the data you feed it, so a biased history produces biased predictions. And it gives you a score, not a reason. Treat the number as a nudge to have a human look closer, not a verdict. It is like a veteran who has seen ten thousand cases and gets a gut feeling about a new one. The gut feeling is just pattern recognition from experience.
LLMs: the next-word machines
A large language model is a specific kind of machine learning model trained on a massive amount of text. It learned the patterns of human language so well that it can read, write, summarize, translate, and answer questions. ChatGPT, Claude, and Gemini are all LLMs with a chat window on top.
Here is how it works without the math: an LLM is a giant next-word predictor. Given the text so far, it guesses the most likely next chunk of words, then the next, and so on. Do that billions of times over a huge body of writing and the result reads fluently. That is also why it can sound completely confident and still be wrong. It is producing plausible text, not looking facts up in a database. Two terms worth knowing. The context window is the model's short-term memory for the current conversation; everything you paste in has to fit inside it. A hallucination is when the model states something false with total confidence, which is why you verify names, numbers, and citations before relying on them.
Grounding (RAG): teaching it your facts
An LLM knows general language, not your organization. Retrieval-Augmented Generation, usually shortened to RAG, is the common fix. Before the model answers, the system fetches the relevant pages from your own material (a policy manual, past cases, contract templates) and hands them to the model along with the question. The model then answers from your documents instead of guessing. It is the difference between asking a smart stranger a question cold, and asking them right after they have read your handbook.
Generative AI, and where images, video, and code live
Generative AI is any AI that creates new content: text, images, audio, video, or code. This is different from older "predictive" machine learning, which only scores or classifies things. Predictive AI says "this client is 80% likely to return." Generative AI writes the outreach email to that client. It comes in one flavor for each kind of content it makes, and they all work the same way underneath, learning patterns from a huge pile of examples, just applied to words, or pixels, or sound.
| If you want to create... | You reach for... | Well-known tools |
|---|---|---|
| Writing and answers | An LLM (the text flavor) | ChatGPT, Claude, Gemini |
| Code | A coding assistant (a specialized LLM) | GitHub Copilot, Cursor |
| Images | An image generator | Midjourney, DALL-E, FLUX |
| Video | A video generator | Sora, Runway |
| Voice and audio | A speech or music generator | ElevenLabs, Suno |
Two things worth knowing. First, coding AI is not a separate species. It is an LLM that has read enormous amounts of software, which is why it can write and fix code in plain conversation. Second, treat all of it as a first-draft engine, not a final-answer machine. It is best used as a fast junior who hands you something to edit, not something to publish unread.
One more distinction: training vs. using
Building a model from scratch (training) is the expensive, one-time, data-center job done by a handful of big labs. Using a finished model to get an answer (called inference) is the cheap, instant part you do every day. When your organization "uses AI," you are almost always running inference on a model someone else already trained. That single fact is what makes private, in-house AI possible, which is exactly where this series is headed.