Normally an AI model is trained once, on a big cloud GPU, by a single team — then frozen and served to users.
This is the opposite. The model lives on a small server that does not train it. The training happens on the laptops of strangers who open the page in a browser tab. Each tab does a few seconds of training, sends a tiny update back, and the server averages everyone's updates into a new version of the model.
It's a real research model — the same JEPA architecture as the relay and silent experiments on this site. Those run a frozen model. This one keeps learning from browser tabs.
The server needs at least two tabs at the same time so contributions can be cross-checked. So: open the hub below, then open it again in a second window.
Each round, your browser runs a couple of training steps on the actual model. Same kind of math the research labs do, just smaller and on your hardware. Takes about 30 seconds per round.
Before the server merges your update in, it checks whether your update made the model better or worse on a held-out test set. Updates that hurt the model are rejected. So you can't break the model by spamming garbage.
Every accepted update earns reputation. Rejected updates cost reputation. A bot dumping random gradients sinks to the bottom of the leaderboard; people whose updates actually help the model rise to the top.
Two different little games, each with its own model. You pick which one to train when you join.
A character has to push a T-shaped block onto a target. The model learns to predict where the character will end up if you push it in a given direction. It only sees pixels — it has to figure out physics by watching.
A predator hunts prey using echolocation — like a bat. It sends out a ping, listens to the echo, and the model learns to predict what the next echo will sound like. Pure audio in, pure audio out, no map.
The bet: that the JEPA models we're building for the relay and silent experiments can be trained by a crowd of browsers instead of by one team renting GPUs. If that works, then every person who plays the game is also helping train the next version of the AI they're playing against. Your gameplay literally becomes the training signal.
Today the federation trains on a small synthetic data pool, not on real gameplay yet — but the pipes are wired end to end. Two tabs, real gradients, real loss decrease. The next step is to pipe in actual game sessions from the relay and silent demos so the model improves from how people actually play.