a fun new way to learn Machine Learning
Learn ML the way it should be taught.
Machine Learning done right. Bite-sized lessons. Real intuition. A small glitchy cube named Tensor who is way too excited to teach you backprop. No textbook, no 3-hour MOOCs, no jargon walls.
Also coming toAndroid· spring '26
- no textbooks
- 500 lessons
- made by humans
- machine learning
- gradients welcome
poke me. i dare you.
Meet your gradient buddy
Hi, I'm Tensor.
I'm an undertrained model who gained sentience and is earnestly excited to help you learn machine learning. I'm fluent in math, still figuring out humans. I glitch when surprised or wrong — that's just how the gradients hit. I will never say "great job!"when you didn't.
- short answers
- no scolding
- occasional puns
- never sarcastic
FAQ · Pacing
How many problems should you do a day?
How it works
Three loops, then you ship.
- 01Lesson 02 · descent
Touch the math.
Drag a ball down a loss curve. Tap a point to flip a class. Math becomes muscle memory before it becomes notation.
- 02Lesson 06 · separable
Get one rep wrong.
Tensor glitches the second you pick the wrong answer. No "try again!" cheerleading — just a one-line fix and the next rep.
- 03Lesson 12 · matmul
Build the small thing.
Every five lessons, you make something tiny that runs. A perceptron. A classifier. Then a transformer block. Then your own.
2103124064120
Inside a lesson
5-minute reps. Real intuition.
Each lesson is a tiny interactive moment — drag, drop, click, wrong, right, next. You don't watch ML. You touch it until it sticks.
- Vectors as arrows you can pull
- Loss as a hill the ball rolls down
- Backprop as a chain of nudges
- Attention as a spotlight you can aim
Tap a point to flip its class. Watch the line refit.