To a language model, the word "cat" is just a long list of numbers. So is "dog". So is "democracy".
What surprised me is that those numbers aren't random. Words with similar meanings end up close to each other, and some relationships even turn into directions you can do math with. I wanted to actually see that, not just read about it. So I built WordCanvas3D.
π Try it: https://wordcanvas3d.vercel.app/
π» Code: https://github.com/Akage1234/WordCanvas3D
It's free, open source, and runs in your browser. No signup.
Here's what you can explore.
Fly through a map of meaning
Every point is a word. Grab the scene, spin it around, zoom in, and you start noticing neighborhoods: animals near animals, countries near countries, emotions near emotions.
Search for a word and the camera flies straight to it, so you can see who its neighbors are.
You can also switch between two ways of squashing hundreds of dimensions into three:
- PCA keeps the big picture, the overall spread of the data.
- UMAP keeps local neighborhoods tight, so clusters pop out more clearly.
Switching between them on the same words is one of the most interesting things to try. The same data tells a slightly different story each way.
See how a model actually reads your sentence
Models don't read words, they read tokens: chunks that can be whole words, parts of words, or single characters.
Type a sentence and watch GPT and LLaMA split it differently, side by side. Some fun things to try:
- A long or rare word
- Numbers and dates
- A sentence in another language
It explains a lot about why "token limits" and "token pricing" work the way they do.
Do math with meaning
This is the part that hooked me in the first place:
paris β france + italy β rome
Take "paris", remove the "france" part, add "italy", and you land near "rome". The model was never taught capitals. That relationship just emerged from reading lots of text.
Try your own equations and see where they land. Not all of them work, and the failures are just as interesting.
Learn the ideas behind it
While building this I had to learn a lot myself, so I wrote short visual guides on embeddings, tokenizers and vectors. They're meant to be read in a few minutes alongside the tool.
Small things that make it nicer
Full screen mode hides the panels so it's just you and the words:
It works on phones and tablets, with touch controls for rotating, panning and pinch-zooming:
And it speaks four languages: English, δΈζ, ζ₯ζ¬θͺ and EspaΓ±ol.
How it's built
- Next.js for the app
- Three.js for the 3D view
- Tailwind CSS for styling
- next-intl for translations
- Python to prepare the data ahead of time
The heavy lifting (loading embeddings and reducing them to 3D) happens beforehand in Python. The browser only loads the finished coordinates, which keeps things smooth even on a phone.
One honest caveat: squeezing hundreds of dimensions into three always loses some information. Nearby words are usually related, but distances are approximate. Think of it as a map, not a measurement.
Go explore
π https://wordcanvas3d.vercel.app/
If you find a weird cluster or an equation that surprisingly works, I'd love to see it in the comments.
And if you want to help: the code is open source, and translating the Learn articles is a great first contribution.
Top comments (0)