Academy · 6 lessons · 7 labs

How AI works. Not read about. Done with your own hands.

Each lesson takes one idea that is genuinely hard, explains it in plain words, then puts the real thing in front of you to operate: a model picking its next word, attention lighting up across a sentence, a reward signal shaping a personality, an agent reading a poisoned email. Where you see numbers, they come from a real model.

Start with lesson 01
LessonsReal model data · runs in your browser · nothing stored
Lesson01
Start here18 min2 labs

How does ChatGPT work? It guesses the next word, one token at a time

Operate a real model's next-word choice: see its odds, turn the temperature, roll the dice.

tokensnextword
Start lesson
Lesson02
Builds on 120 min1 lab

How does a transformer know which words matter? Attention, shown for real

Hover a word and see where a real model looks. Find the head that links 'she' to 'sister'.

attention
Start lesson
Lesson03
Builds on 118 min1 lab

How are ChatGPT and Claude actually trained? Pretraining, human feedback, and why they are so polite

Be the human in human feedback. Rank answers, watch the model learn to please you, catch sycophancy being born.

rlhf
Start lesson
Lesson04
Builds on 116 min1 lab

Why does AI hallucinate? Why ChatGPT is confidently wrong and can't say 'I don't know'

Play the calibration game: 12 questions, your confidence, your curve next to GPT-4's.

calibration
Start lesson
Lesson05
Builds on 214 min1 lab

Why does ChatGPT forget? The context window, explained simply

Fill a context window with your own messages and a pasted PDF, and watch the beginning fall off.

context
Start lesson
Lesson06
Deep end18 min1 lab

What is an AI agent, really? The plan, act, observe loop, tools, and the ways agents go wrong

Run an agent on a fake inbox. One email is poisoned. Choose the permission mode and see what happens.

agent
Start lesson
How we make these

Real numbers, or we say it's a simulation.

Where a lab shows probabilities or attention, they come from a real open model (GPT-2) run once and saved into the page. Where a lab is a simulation, like the reward game or the agent inbox, it says so on the lab, and the text explains how the real thing differs.

Suggest a lesson