How modern AI actually works

Strip away the hype and the fear, and modern AI is a specific, learnable stack of ideas: neural networks that learn from data, a transformer architecture that scales, and a training process that turns next-token prediction into general capability. This site explains that stack clearly.

Start here

Foundations

Neural network fundamentals

Neurons, weights, loss functions, gradient descent, and backpropagation: the five ideas everything else builds on.

Architecture

Transformers and LLMs

Attention, pretraining, instruction tuning, RLHF, and reasoning: how a text predictor became the engine of the AI era.

Landscape

The model landscape

The frontier labs and their model families: what distinguishes them, who is open and who is closed, and how to think about capability.

History

The research timeline

From the 1943 artificial neuron to the agent era: the milestones that got us here, in order.

Primary sources

The papers that built the field

Twelve papers that matter, each summarized in a paragraph you can actually read.

Reference

Glossary

The vocabulary of modern AI, from attention to zero-shot, defined plainly.

Why this site exists

Most AI content is either marketing, doom, or math that assumes three years of prerequisites. The reality in the middle is more interesting: a small set of genuinely elegant ideas, scaled to an industrial level, producing systems that surprise even their builders. You can understand those ideas properly in an afternoon, and understanding them changes how you read every AI headline after.

The one-paragraph version

A neural network is a function with millions or billions of adjustable numbers. Training nudges those numbers, example by example, to make the function's outputs match its targets; gradient descent decides the direction of each nudge and backpropagation assigns the blame. The transformer is the particular arrangement of that function that turned out to scale: it processes text by letting every word attend to every other word, and it converts the entire internet's text into general-purpose capability through one humble objective: predict the next token. Everything since: chatbots, coders, reasoners, agents, is refinement and scaling of that core.

Where to go next

If you are new: fundamentals, then transformers, then the timeline. If you already build with these systems: the model landscape and the papers.