The hardest part of getting into AI is the vocabulary. AI, ML, DL, GenAI, LLMs, RAG, agentic AI: they get thrown around as if they were synonyms, but they are not. They are a hierarchy, each one a more specific thing nested inside the last. Once you see the nesting, the whole field clicks into place.
Here is the map.
AI: the big umbrella
Artificial intelligence is the broad goal: machines that do things we would call intelligent, like reasoning, perceiving, or deciding. Everything else on this list is a subset of it. AI is old and wide: it includes rule-based expert systems and game-playing programs (chess, Go) that have nothing to do with modern neural networks.
So "AI" tells you the ambition, not the method.
ML: learning from data
Machine learning is the subset of AI where, instead of hand-coding rules, you let a system learn patterns from data. Show it enough labeled examples and it figures out the rule itself.
Examples: spam detection, recommendation systems, fraud detection. If a system improves by being shown data rather than by being programmed with explicit rules, it is machine learning.
DL: deep neural networks
Deep learning is the subset of ML that uses neural networks with many layers. The "deep" is literally the number of layers stacked up. These layers let the model learn its own features, which is why deep learning took over tasks that defeated earlier ML.
Examples: image recognition, speech recognition, self-driving perception. Whenever you hear "neural network," you are in deep learning territory.
GenAI: models that create
Generative AI is the subset of deep learning whose job is to create new content rather than just classify or predict: text, images, audio, video, code. The shift is from "label this" to "produce this."
Examples: ChatGPT, Midjourney, DALL-E, Stable Diffusion. If the output is new content rather than a label or a number, it is generative.
LLMs: language models
Large language models are a kind of generative AI focused on understanding and producing human language. They are trained on enormous amounts of text and predict the next token over and over, which turns out to be enough to answer questions, write, and reason in language.
Examples: GPT, Claude, Gemini, Llama, Mistral. An LLM is the engine; the chatbot you talk to is that engine wrapped in a product.
RAG: giving models fresh, specific knowledge
A model only knows what it was trained on, which is frozen and general. Retrieval-augmented generation fixes that: before the model answers, you retrieve relevant documents from an external source and feed them in, so the answer is grounded in current, specific knowledge.
This is how you build an assistant over your documents: company policies, product docs, a knowledge base. RAG combines a retriever (find the relevant text) with a generator (write the answer using it), which is why enterprise chatbots and document Q&A are almost always RAG under the hood.
Agentic AI: systems that take action
Agentic AI is the newest layer: AI that does not just answer, but reasons, plans, uses tools, and takes actions toward a goal, often over many steps. Instead of "what is the answer," the question becomes "go achieve this."
Examples: agents that browse, call APIs, write and run code, and chain steps together. An agent is an LLM given goals, tools, and a loop, so it can act in the world rather than just talk about it.
The evolution, in one line
The pieces stack in order of increasing specificity and capability:
AI (think and act) → ML (learn from data) → DL (learn complex patterns) → GenAI (create content) → LLMs (understand language) → RAG (add fresh knowledge) → Agentic AI (act autonomously).
Each is a subset or extension of the one before. That is the whole secret to the jargon.
Why this matters
Knowing the hierarchy is not trivia. It tells you the right tool for the problem: a fraud classifier is plain ML, not an LLM; a chatbot over your docs is RAG, not fine-tuning; a workflow that takes actions is agentic, not a single prompt. Picking the wrong layer is how AI projects waste months.
A few more concepts worth knowing, since they cut across the stack: fine-tuning (adapting a model to your data), prompt engineering (designing inputs for better outputs), multimodal models (text, images, audio together), and AI safety (building systems that are trustworthy and aligned).
Understand it by building it
The fastest way to make all of this concrete is to build the bottom of the stack yourself. When you have hand-coded a neural network, watched gradient descent train it, and built an embedding and an attention layer from scratch, the words stop being buzzwords. You can reason about what a model is actually doing, which is exactly what lets you pick the right layer for a real problem.
That is what the AI and Deep Learning track is for: you build from the foundations up to a tiny GPT, all from scratch in Python, graded in your browser. The first project is free.
Learn the stack from the bottom, and the whole AI landscape stops being a wall of acronyms.
