How AI actually works, in plain English
Nine mechanical concepts that unlock everything else, plus a visual map of how "AI," "machine learning," "deep learning," and "generative AI" actually relate to one another — and where that tidy nesting breaks into a tree.
Where the ladder becomes a tree
AI → Machine Learning → Deep Learning → Generative AI nest cleanly inside one another. After that, the picture branches — twice.
A diagram showing Artificial Intelligence containing Machine Learning, which contains Deep Learning. Deep Learning then branches into five parallel sibling fields: Computer Vision, Natural Language Processing, Reinforcement Learning, Speech Recognition, and Generative AI. Generative AI itself branches by output type into LLMs (the text branch), Diffusion models (image, video, and audio), and GANs or VAEs. Agentic AI is shown as a dashed wrapper drawn around an LLM, adding tools, planning, and memory, rather than as another nested ring.
Nine concepts, defined plainly
Click a card to expand it. Each definition ends with a concrete example, not just an abstract description.
Core mechanicNeural network
Layers of simple units ("neurons") connected by weighted numbers. Data goes in, each neuron does a small weighted calculation, a prediction comes out — repeated multiplication and addition arranged so the network can approximate very complex patterns given enough layers and the right weights.
Core mechanicTraining
Finding good values for all the network's weights. Show it an example, it guesses, you measure the error ("loss"), backpropagation calculates how each weight contributed, gradient descent nudges every weight to reduce the error — repeat millions of times.
Core mechanicParameter
One learned weight (or bias) inside the network — a single adjustable dial. Modern models have billions of them; more dials generally means more capacity to represent complex relationships, which is why parameter count is used as a rough, imperfect proxy for capability.
ArchitectureTransformer architecture
Older models (RNNs/LSTMs) read text one word at a time, trying to remember everything so far. The 2017 Transformer looks at all words at once using self-attention — every word can consult every other word in parallel, which trains far faster and holds long-range context better.
ArchitectureAttention mechanism
Lets a model decide, for each word, how much to "pay attention" to every other word in the input. In "the animal didn't cross the street because it was too tired," attention helps the model learn that "it" means "the animal," not "the street."
Core mechanicToken / embedding
A token is a chunk of text the model reads at once — often a word or word-piece. An embedding is how a token is represented internally: not as text, but as a long list of numbers positioned so tokens with similar meaning end up near each other in that space.
Working with modelsFine-tuning vs. prompting vs. RAG
Three ways to shape a model's behavior, in increasing cost: prompting (better instructions, no retraining), RAG (fetch relevant documents at question time, without touching the weights), fine-tuning (retrain the weights themselves — most powerful, most expensive).
ArchitectureMultimodality
A model that takes in and/or produces more than one kind of data — text, images, audio, video — through one unified architecture. Images are chopped into patches and audio into clips, converted into the same kind of numerical embeddings used for text.
Working with models"Agent" vs. chatbot
A chatbot answers one message at a time with no ability to act in the world. An agent is the same underlying model wrapped with tool use (search, code, APIs), multi-step planning, and the ability to decide what to do next on its own.
Now that you know how it works — how do you actually use it well?
Prompt structure, image and video prompting, agent design, and the vibe-coding workflow — with real, worked examples.