An open notebook on modern AI

One idea.
See it work.

Go beyond the explanation. Pull on a thread, change a variable, and watch the mechanism reveal itself.

Try one on this page Browse the drips

On the page / In your hands

What happens
after the first answer?

Follow an agent from its first attempt to a better one.

Goal → Attempt → Check → Revise ↻

Step 01 / Goal

First, make the target clear.

Ask an agent to write a function that sorts a list. Give it a concrete check: the result must be in ascending order.

input: [3, 1, 2] → expected: [1, 2, 3]
Illustrative walkthrough · no model is runningRead more about agent patterns ↗
Amit, your guide through the worlds

New Worlds · free, no account

Walk through a chatbot.

Each world is one idea laid out as a place you scroll through, with Amit as your guide. Season one follows a single sentence from the moment it is chopped into pieces to what changes as models grow. Every number comes from a real run.

Start at stop 1 See all worlds
  1. TokenizationHow a Sentence Gets Chopped
  2. Next-token predictionLearning to Write
  3. AttentionLetting the Translator Look Back
  4. Temperature and samplingHow It Picks the Next Word
  5. The KV cacheWhy It Doesn't Reread Your Chat
  6. Emergent abilitiesEmergence, or a Change of Ruler

Begin anywhere

A good place to start.

All drips
Drip · 14 minTransformers, From First Principles
Drip · 10 minDiffusion Models, Denoised
Drip · 8 minA Natural Introduction to NLP

Everything in the notebook

Explore an idea, learn a subject, or build something.

DripsOne AI concept, a short explanation and a hands-on experiment.53 explainers CoursesLearn a subject step by step, with lessons in a useful order.16 courses BlueprintsA practical guide with working code that ends in something you can run.33 build guides NotesEssays and lessons learned from building with AI.6 essays