A path, not a syllabus
Each stop is one short lesson. Clear a few, crack a chest, then face the world's boss.
A new way to learn AI
Aila is a daily game about what is really going on inside the AI you use. Split words into tokens, catch a model making things up, and keep your streak alive.
hallucinates
3 tokens

Try it
Models see text as tokens: chunks from a fixed vocabulary. Type anything and watch GPT-4o's real tokenizer break it apart.
6 tokens · GPT-4o tokenizer
24/60 characters
“hallucinate” is 2 tokens: halluc + inate. Rare words get built from smaller pieces.
From lesson 1: What is a token?
The game
Lessons are quick rounds of tapping, sorting and spotting mistakes. The game around them rewards you for coming back.
Each stop is one short lesson. Clear a few, crack a chest, then face the world's boss.
Play a little each day to grow your streak. Miss one, and a streak freeze keeps it alive.
Three fresh goals every day. Finish one and open a chest of Sparks.
Every world ends with a boss: 10 questions, 3 hearts. Win and the next world opens.
Stuck on a question? Ask why. Aila explains at the depth you picked.
10 worlds
183 lessons across 10 worlds, each with its own color, island and boss. Start with what AI is. Finish by building with it.
Swipe through all 10
Three ways to learn
Every lesson comes in three depths. Pick the one that fits what you want today, and switch whenever you like in your profile.
Same question, three depths: What is a token?
You want the idea, fast.
AI doesn't read words the way you do. It chops text into small chunks called tokens. Short, common words are usually one chunk each.
You want to know why AI behaves the way it does.
A tokenizer maps text to pieces from a fixed vocabulary. Common words get their own token. Rarer ones are built from parts, so "hallucinates" becomes hall + ucin + ates.
You want the real mechanics, math and code included.
GPT-4o uses byte-pair encoding with the o200k_base vocabulary, about 200,000 tokens. Each token is an integer ID that selects one row of the model's embedding matrix.
You already split words into tokens up there. Lesson one, “What is a token?”, picks it up from there in about 3 minutes.