What AI Actually Is (and Isn't)

Big idea

AI, machine learning, and deep learning are nested circles, not synonyms, and today's chat tools are next-word predictors trained at huge scale — forgetting that is how people end up treating them like thinking minds.

The idea

"AI" is a fuzzy umbrella term, and most of the confusion around it comes from people using it to mean three or four different, nested things at once. Artificial Intelligence is the broadest category — any system that does something we'd call "intelligent" if a person did it, including plenty of old-school techniques that have nothing to do with the tools you use today. Machine Learning is a specific way of building AI: instead of hand-writing rules, you show a system lots of examples and let it learn the pattern itself. Deep Learning is a specific way of doing ML, using layered neural networks — it's the approach behind almost everything people mean by "AI" right now.

Think of it like

Think of a grocery store. "Food" covers everything on the shelves. "Baked goods" is one aisle inside it, and "bread" is one shelf in that aisle. A sourdough loaf sits on that shelf: it is food, it is a baked good, and it is bread all at once, but plenty of food is not bread. When people say "AI," they might mean the whole store or just that one loaf. In the same way, AI is the whole store, machine learning is the baked-goods aisle, deep learning is the bread shelf, and an LLM is one particular loaf.

Where LLMs and generative AI fit

A Large Language Model (LLM) is a deep learning model trained on huge amounts of text to predict what comes next. That's it, mechanically — but at large enough scale, "predict the next word well" turns out to produce something that can hold a conversation, write code, and reason through problems. Generative AI is the even broader label for any model that produces new content — text, images, audio, video — rather than just classifying or scoring something that already exists. An LLM is one kind of generative AI; image generators are another.

How a tool like Claude or ChatGPT is actually trained, in plain terms

Three rough stages, in order:

1. Pretraining — the model reads an enormous amount of text from the internet and books, over and over, getting slightly better each time at predicting the next word. This is where almost all of its raw knowledge and language ability comes from.
2. Fine-tuning — the base model gets further trained on curated examples of the kind of behavior you actually want (helpful, well-formatted answers), not just "statistically likely text."
3. Alignment (often RLHF — reinforcement learning from human feedback) — humans rank different possible responses, and the model is nudged toward the ones people preferred, and away from ones that were unhelpful, rude, or wrong.

Why this matters before anything else

Almost every misconception about AI — "it thinks like a person," "it knows things," "it's always right," "it's basically magic" — comes from skipping this distinction. It isn't reasoning the way you do; it's a very sophisticated next-word predictor, trained until that turned out to be enough to do a remarkable range of things. Keeping that straight is what lets you use it well instead of either over-trusting it or dismissing it.

Three generations, one task

The same task — flag spam email — built three ways:

Each step trades hand-written logic for patterns learned from data. The LLM hasn't been trained on your spam specifically — it generalizes from everything it has read.

Key terms

• Machine Learning — A way of building AI where the system learns patterns from many examples instead of following hand-written rules.
• Deep Learning — A kind of machine learning that uses neural networks with many stacked layers.
• Large Language Model (LLM) — A deep learning model trained on huge amounts of text to predict what word comes next.
• Pretraining — The first and largest training stage, where the model reads vast amounts of text and gets better at next-word prediction.
• RLHF — An alignment step where people rank candidate answers and the model is nudged toward the ones they preferred.