CHOOSE A DOOR

Nothing behind these doors is a video. Every room is computed live on your device — organisms evolving, optimizers falling, a planet of real visitors, a radio tuned to the whole world.

TOKENIZE

A model never reads your sentence. It reads pieces — subwords chosen by frequency, not meaning. Type below and watch language get butchered in real time.

ATTEND

Self-attention is the trick that made all of this possible: each token asks every other token "how much do you matter to me?" — and gets a different answer from each head. Hover the words. Switch heads. Same sentence, four different readings.

PROPAGATE

Forty-two neurons, five layers, four hundred weighted edges. Click any neuron on the left and watch activation cascade to the output — bright edges carry the signal, dim edges lose it.

LEARN

A real neural network — 2 inputs, two hidden layers of 8, one output — training live in your browser with genuine backpropagation. Paint two classes of points and watch it carve a decision boundary between them. The spiral is the classic torture test. It will struggle. Then it won't.

DESCEND

Every model you've ever used was trained by rolling downhill on a landscape like this one. Click anywhere to drop three optimizers and race them to the bottom: plain SGD takes every bump personally, MOMENTUM barrels through, ADAM adapts. The pit they find may not be the deepest one there is.

PREDICT

Underneath everything, a language model does one thing: guess the next word. This one starts nearly empty and trains on you — every word you type updates its memory, live. Write a few sentences, repeat yourself a little, and watch it start finishing your thoughts. Press TAB to accept its best guess. Nothing leaves this page.

GENERATE

Sampling made visible. Temperature is chaos, top-p is how much of the field is allowed to speak, entropy is how fast the dream forgets itself. Drag on the canvas to inject your own intent.

LATENT WALK

Generative models don't store images — they store a space, and every point in it is a picture. This field works the same way: a seeded noise landscape you can walk through forever. It never repeats, never stops, and any frame can be exported at 1024px the instant it looks right.

KALEIDO

Draw one line — it appears n times, rotated and mirrored around the center. The hue drifts as you draw, so long strokes become spectra. Every accident becomes ornament. Export the mandala when it stops being yours and starts being good.

HALFTONE

Before machines dreamed images, they printed them — as dots whose size carries the tone. Drop a frame and rebuild it as press-plate halftone: dots, squares, or scan lines, at any screen angle. The classic print-tech look, computed live.

PLOTTER

Drop a frame and the machine redraws it using nothing but marks you define: scanline waves whose amplitude is the pixel, additive R/G/B trichrome lines that fuse into full colour, a single spiral, print-press dashes, stipple fields, or strokes that flow along the image's own gradients. Everything exports at 2048px — plotter-ready, print-ready.

CONVOLVE

Before a vision model understands anything, it drags tiny 3×3 filters across the image. Draw in the first panel — the other three show what the machine actually sees: vertical edges, horizontal edges, and their combined magnitude. Your drawing never arrives whole. It arrives as gradients.

SKETCH ORACLE

Draw, and a point-cloud recognizer guesses what it is — live, while your hand is still moving. It compares your strokes against its learned shapes entirely on this page: no upload, no cloud, no waiting. Play six rounds against the clock, or free-draw and watch it argue with itself.

WORK HERE

NAHITSAI makes high-end motion pictures, TV, and immersive experiences with state-of-the-art AI tools — and we are always hunting for the rare people who treat these machines like instruments, not shortcuts. If any door on this site made you lose track of time, you might be one of us. No CV theatre. Show us what you make.

Every generation stops on a token like this one. Yours stops here — unless you have something worth passing forward.