A Primer on AI
How modern AI systems work, from a single neuron to the alignment problem
This site explains how modern AI systems work, in enough depth to follow current AI safety research and to start contributing to it. It assumes the basics of linear algebra, calculus, and probability, but no machine learning.
Every page answers one question with one interactive tool. Read the short text, make the prediction each task asks for, then use the tool to check it. Each page ends with a check question and a box for explaining the idea in your own words. Your answers and progress stay in this browser.
The marks the twenty-one pages of the core path, about five hours. They carry the main line of argument, from “a network is a function” to “why a trained system can pursue goals its designers did not intend.” A small circle after a page title fills in once you have answered all of its check questions correctly. A question you miss is asked again the next time you open its page, and questions from earlier sections return at the start of later ones. Sections 1 to 8 each end with a transfer question, which applies the section’s ideas to a case the section did not cover. To follow the core path alone, choose it below; each page then links to the next core page.
Text and figures by Aydin Mohseni, licensed CC BY 4.0. Code licensed MIT, at github.com/amohseni/AI-Safety-Fundamentals.