How AI Works
Start knowing nothing. By the end you'll know what's inside an AI, how it learns, how an agent differs from a chatbot, what happens when AI helps build the next AI, and what could come of it.
Seven chapters, about 40 minutes in all, and each one stands on the one before. No math past fractions. You press buttons and watch things happen, and most of what you run is real.
Short on time? On each page, read the bold "The point" lines, press the buttons, and read the last box, "What you can fairly say".
The path
- 1What's inside an AIA very long list of numbers, and one job: guess the next word. Build a sentence one guess at a time with a real model.
- 2Grown, not writtenNobody writes the rules. Watch a tiny network train, and see why you can't simply find a habit and delete it cleanly.
- 3What scale buysChips, power and money. How far apart the sizes are, what that has bought, and what it hasn't.
- 4Learning from rewardsTrain a small learner in your browser. Then add one badly designed reward and watch what it does.
- 5Chatbots and agentsThe same model can be either. Step through one task both ways, and see why long jobs are hard.
- 6The LoopAI helping build the next AI. Run a simulator of two clocks: how fast AI improves, and how fast people can check it.
- 7What could happenA map of the serious outcomes, not a forecast. Pick your assumptions and see where they lead.
How to read the numbers
Every number on these pages says what kind of number it is.
- measured
- Somebody counted it.
- third-party
- Someone else reports it, and these pages can't check it themselves.
- assessed
- A judgment call.
- simulated
- It follows from assumptions.
Want to track this as it moves? The Absorption Index is this site's running reading of AI's pace, one editor's judgment with its rules shown.
Want more on Chapter 1? The Parrot Test is the deep version, about 20 minutes, with GPT-2 opened up floor by floor.