The Concept Lab

Small browser-based tools that take a concept — Fourier decomposition, autoencoders, embeddings — and turn it into something you build, break, and watch work, instead of just reading about it. Open source and free to use.

Tonus

Generate a 2D dataset — or draw your own — and watch eight real machine learning models, from a closed-form linear fit to a live-trained neural network, each draw its own decision boundary.

Inkling

Draw a letter, teach a tiny neural network what it is, and watch real backpropagation update every weight live — then test it on a whole word and export the trained network into a QR code.

Echo State

Type anything and watch a tiny untrained recurrent network's memory trace a unique path through space, character by character — export it as a print, or a QR code that replays the exact model live.

Neural Viaduct

Run a real plain network and a real residual network on your own photo, side by side, and watch the skip connections actually happen — the same mechanism that let vision networks stop breaking as they got deeper.

Vectis

Pick two words as axes and plot anything — text or your own photos — by how similar it is to each one, using a CLIP model running entirely in your browser.

Afterimage

Reconstructs an uploaded photo from a 256-number impression, using a small neural network trained once and shipped frozen — the reconstruction happens entirely in your browser.

Ridgeline

Traces a mountain silhouette from a photo, decomposes it with a discrete Fourier transform, and renders the reconstruction as an image sequence or animation.

Matrix Matrix

What a matrix actually is, what it can represent, and what multiplying two of them means — worked by hand from a 2×2 up to a 4×4, then two matrices of completely different shapes, with a shape checker to test your own.

Hunch

Draw candies from a secret jar and watch your hunch update with each clue — a hands-on walkthrough of real Bayes' theorem.

Attention

Type a sentence and watch real scaled dot-product attention — the mechanism behind every modern language model — run on your own words, from tokens through Query/Key/Value, attention weights, a KV cache, and next-token prediction.