Makingintelligenceintuitive.
Tensor Thinking turns data science, machine learning and AI into short, cinematic visual lessons, so the intuition arrives before the equations do.
Everything in AI is a tensor.
Images, words, sounds and the models that read them are all organised blocks of numbers. Keep scrolling.
A cube of numbers.
Sixty-four values arranged along three axes. Picture a colour image: height, width and channel.
Reshape it into a matrix.
Same sixty-four numbers, now in rows and columns. Nothing is lost, only rearranged.
Flatten it into a vector.
One long list with a direction and a length. This is where our first series begins.
Collapse it to a scalar.
A single number, like the loss a model tries to shrink. That is tensor thinking.
The ideas behind AI are not out of reach. Most of them were simply never drawn well. We turn vectors, gradients, networks and attention into motion you can see, so understanding comes first and the math feels like the obvious next step.
Four principles behind every lesson.
See it before you solve it.
Every lesson opens with a picture you can feel. Symbols only arrive once the shape of the idea is already in your head.
Built from first principles.
No hand-waving. Each concept is assembled from the ones beneath it, so the whole structure holds when you lean on it.
Precise, never simplified.
Intuitive does not mean approximate. Every curve, arrow and transformation on screen is mathematically exact.
One idea at a time.
Each lesson does one thing well and connects cleanly to what came before, so knowledge compounds instead of piling up.
Vector Pilot.
Steer a comet with vectors. Point at any grid point and click to launch. Chain moves head to tail, slip past obstacles, and land on the star in as few moves as you can.
Four tracks, in the order they build on each other.
Start anywhere. Every concept links back to the ideas it depends on.
Data Science
Vectors, distributions, regression and the stories hiding in a table.Vectors · mean and variance · correlation · PCA
Now airing 02Machine Learning
How models draw lines through data and know when they are wrong.Loss functions · gradient descent · overfitting · trees
Coming soon 03Deep Learning
Layers of simple neurons that learn surprisingly complex things.Perceptrons · backpropagation · CNNs · normalization
Coming soon 04Generative AI
Embeddings, attention and the machinery behind modern language models.Tokens · embeddings · attention · transformers · diffusion
Coming soonWhere the ideas were born.
Twenty-six breakthroughs, from Alexandria to Mountain View. Select a marker to meet the people behind it, or play the story in order.
Data Hunter.
A small break between lessons. Steer the cube, eat data to grow your tensor, dodge the outliers, and use power-ups named after tricks real models rely on.
shape (1,)Best0Data Hunter
The cube chases your cursor or finger. Eat the data points, chain them quickly for momentum, and never touch an outlier.
or press Space. Arrow keys steer too.The engine behind every frame.
Every Tensor Thinking animation is generated from a production-grade Python notebook. The library is available by licence to people who want to teach with the same visuals.
- EducatorsClassroom-ready notebooks and animations for your own lectures and courses.
- TeamsVisual, interactive training to onboard engineers and analysts faster.
- CreatorsLicence the animation toolkit to build explainers in your own style.
The full notebook, render pipeline and assets are available with a licence.
Request accessStay curious.
One email on Sundays with the week's new lessons and a look at what is coming next. No noise.
Delivered through Substack. Unsubscribe any time. See our privacy policy.