Privacy and safety

What stays on your device

Everything you type, what it says back and what it remembers. The model runs in your browser, on your own graphics card. I checked the network traffic: while you play, the page sends nothing at all.

What leaves it

Only the download, once. The code comes from jsDelivr, the model from Hugging Face and its graphics code from GitHub. They see your IP address and which brain you picked, not what you type. Gemini Nano is downloaded and kept by Chrome itself. The site counts page visits with Umami; nothing from this game goes there.

What it keeps

One short note about you, in your browser’s storage, and only words you typed yourself. “Forget” wipes it. Unplugging frees the memory. The download stays in your browser until you remove it under Setup, or clear the site’s data.

What the model can do

Pick one move from a fixed list and say a few words. It can’t open links, run code or fetch anything, and what it says is shown as plain text. Ask it to misbehave and the worst you get is a strange line.

Not watertight yet

The code and the model files come from those other servers without an integrity check. If one of them were hacked, it could change what runs on this page. Next on my list: serving that code myself and pinning the files.

The measurements

The gauge shows numbers only. “Copy as JSON” puts them on your clipboard, with your graphics card and browser, never what you typed. Nothing is sent anywhere unless you paste it yourself.

Measurements

Last replies

Nothing measured yet. Plug it in and give it something to react to.

Lab · Beta

Unplug.

A small creature with a language model for a brain, on your own graphics card. It picks how to react. The code checks that choice and plays it. Pull its plug to switch it off.

WebLLM (opens in a new tab)

Brain

Every brain runs in your browser. It downloads once from Hugging Face, and after that nothing you type leaves your device, wherever the model was made.

Temperament

How it’s built

A small language model runs on your own graphics card, in a Web Worker, through WebLLM and WebGPU. The page tells it what just happened, the code checks its answer and plays it with springs on one SVG.

Tech
WebLLM · WebGPU · llama.cpp (wllama) · Prompt API · SVG · Pointer Events · Vibration API
Source
2444 lines · 112.4 kB