Back to the sky

About the work

Neighbourhoods is a map of two thousand ordinary English words, placed according to how a small language model understands them.

What you are looking at

Each point of light is a word. Bread, winter, grief, ladder, velvet. The model has read every one and given it a place among the others. Words that usually appear in similar contexts end up near each other. Rain sits close to puddle and downpour; mother close to father and grandmother.

The model works in a space with hundreds of directions. What you see is a flattened version of that space, so some of the original closeness remains and some of it disappears.

When you type a word

Your word is given a place in the same way and appears as a brighter star. Its eight closest neighbours light up and are joined to it by thin lines. Those neighbours are chosen in the model’s full space, which is why some of the lines stretch a long way across the sky. Two words can be close in meaning and still far apart in the picture.

You can type any word, or a short phrase. It doesn’t have to be one of the original two thousand.

The trail

Words you’ve already typed stay behind as fainter stars. The longer you stay, the more of the sky fills in.

What you hear

With sound on, a low hum runs under the sky and each neighbourhood becomes a soft chord. Everything is tuned to the Earth. The space between the ground and the upper atmosphere rings at roughly 7.83 beats a second, the Schumann resonance. Raised four octaves, that becomes the low note the piece is built on.

The word you just typed sets the root. Higher in the sky means higher in pitch. Each neighbour adds a note from the same gentle five-note scale, so the chords never clash. Close neighbours land on fuller intervals; more distant ones sound quieter and higher, like stars further away. Every note starts with a small chime when its star appears, then fades into a long echo. The hum underneath slowly breathes in and out, following the same Earth rhythm, slowed down six octaves.

How the model listens

The model is small. It learned from more than a billion pairs of sentences, most of them taken from online conversations. The neighbourhoods it forms are a record of how people actually use words.

It has its own habits. With rarer words it sometimes pays more attention to the sound than the meaning, which is why rhubarb ends up near bar and bear. Those slips stay in the work.

The model runs on your own device.

Two notes

The pink is the colour of the Belt of Venus, the band of light that sits above the Earth’s shadow at dusk, on the line between day and night.

The typeface is Familjen Grotesk. Familjen is Swedish for family. Wittgenstein wrote that the different meanings of a word rarely share one single essence. They are held together by family resemblances, overlapping similarities that cross and connect like the fibres in a thread. The lines in this sky draw one such web.

Thanks

A work by Peter Pechmann, 2025.

Language model: all-MiniLM-L6-v2 by the sentence-transformers project, trained during a Hugging Face community event, converted for the browser by Xenova. Apache License 2.0.

Typeface: Familjen Grotesk by Familjen STHLM. SIL Open Font License 1.1.

Software: three.js (MIT), Transformers.js by Hugging Face (Apache 2.0), ONNX Runtime Web by Microsoft (MIT).

The word list and the sound were made for this work.

Sources

The full licence texts of all software, the model and the typeface are in licences.txt.